Today, fewer people need convincing that AI is a big deal. If you squint, you can even see the crest of the wave on the horizon. And yet opinion still varies widely on the exact size and timing of AI’s impact. While the beachgoers have largely left the beach, many “tsunami optimists” are still stubbornly ignoring the official forecasts, believing they can ride out the wave from their high-rise and that only gullible “doomers” run for the hills.

The phenomenon of people repeatedly underestimating the pace of AI progress is an example of exponential blindness. An amusing case of this was seen with the recent release of a preprint from UPenn researchers that attempted to cast doubt on METR’s forecast of AI’s accuracy on long-horizon tasks (our current best measure of AI autonomy). While a simple extrapolation suggests the task-length that an AI can reliably perform is doubling every 5-7 months, the UPenn researchers argued that the apparent exponential is actually an S-curve on the verge of plateauing. That same day, OpenAI and Anthropic both released new models that showed the exponential trend was alive and well, effectively debunking the paper in real time.

I’ve thus taken to dividing the AI commentariat into empiricists and extrapolators. The empiricists often claim the epistemic high ground, preferring hard data over so-called “towers of assumptions.” In practice, however, this often reduces to a version of “I’ll believe it when I see it.” Don’t get me wrong: empirical data is essential to forming a coherent model of the world. At the same time, pure empiricism has a tendency to collapse into a kind of Humean skepticism, as though our repeated observations of the sun rising in the east give no guarantee that it will again tomorrow.

Extrapolators, meanwhile, are often dismissed as a-theoretical, “line go up” curve fitters, and while those types certainly exist, there is a world of difference between the people who do “technical stock analysis” by drawing lines on price charts and the people who construct deep structural models to make rigorous, out-of-distribution bets. On the contrary, it is usually the empiricists who are a-theoretical to a fault, leading them to overfit sparse data with epicycles upon epicycles rather than commit to a formal model that attempts to generalize the underlying process.

Subscribe now

Extrapolators have a remarkable track record in the AI field, being repeatedly early to trends and capabilities that empiricists believed were still decades away. They did not simply get lucky by drawing straight lines on a graph. Rather, they grounded their extrapolations in a first-principles understanding of physics, biology, neuroscience, statistical mechanics and other relevant fields. The founding team at Anthropic, for instance, built their strategy around a high-conviction bet on scaling laws being more than just an empirical regularity, but rather something genuinely law-like. The success is a testament to the power of simple priors and heuristics, such as Ilya Sutskever’s famous quip that “models just want to learn,” or Ray Kurzweil’s 1999 forecast of human-level AI by 2029 based on little more than an extrapolation of Moore’s Law.

The extrapolators’ conviction in scaling laws is part empirical, part theoretical. On the empirical side, AI researchers have by now measured the predictive accuracy of scaling laws across many orders of magnitude. On the theoretical side, there is good reason to believe these scaling properties generalize across diverse modalities. In essence, once a training method starts to work a little bit — once a “hill” starts to be “climbed” — it is safe to assume that we’ll reach the summit sooner rather than later. Once video generation models started to work at little bit, for example, it was only a matter of time before it would be possible to generate life-like movie sequences. In other words, that we would soon have video models as good as the new Seedance model from Bytedance was in a sense knowable from at least the first, heavily distorted video of Will Smith eating spaghetti — if not from the days of Alexnet.

Empiricists often retort that predicting the future is impossible because the world is a complex system, but this is only half true. The world may be complex, but complex systems are typically controlled by a small number of highly stable invariants. Moreover, the complexity of a systems is often scale-dependent, becoming stable in regimes where chaotic dynamics average out. This is why it is easier to predict the earth’s average temperature a century from now than to forecast the weather in three weeks. The former is controlled by basic thermodynamics and a handful of parameters, while the latter is highly sensitive to initial conditions and small perturbations a la the butterfly effect.

Forecasting AI capabilities is closer in kind to solving a thermodynamics equation than solving the three-body problem. This doesn’t mean it’s easy; only that it is possible. It is also necessary, as human institutions move vastly slower than the pace of AI progress. Preparing for AI thus means preparing for capabilities that do not yet exist, but which we can have some confidence will exist by the time our preparations are in place.

As a proponent of chip export controls on China, I was under no illusions as to why I was among the cohort of think tankers invited. The UAE sits at the intersection of the U.S.-China relationship as both one of China’s major trading partners and the site of the UAE-U.S. AI Campus: a 5 GW data center project led by G42, the Abu Dhabi-based AI cloud company, in partnership with Oracle, NVIDIA, Cisco Systems, and Softbank. The first phase of the project is a 1 GW compute cluster called Stargate UAE that is set to begin operating in 2026 with OpenAI as its anchor tenant and co-operator.

G42 previously held stakes in Chinese companies such as ByteDance and engaged deeply in China’s tech and biotech sectors. U.S. intelligence agencies also learned in 2022 that G42 allegedly provided Huawei with technology later used by the PLA to extend the range of their air-to-air missiles. Many U.S. lawmakers and policy analysts have thus had concerns that U.S. tech investments in G42 — and the UAE more broadly — could transfer sensitive IP and technologies to Chinese entities, potentially including model weights and advanced chips. However, with the May 2025 announcement of the U.S.–UAE AI Acceleration Partnership, a bilateral framework for technology cooperation that includes Stargate UAE, the tide in Washington has clearly turned.

Our visit was largely structured around assuaging any residual concerns about the UAE’s technological exposure to China. In our nearly two-hour-long meeting G42’s chief global affairs officer, Talal Al Kaissi, every question we had was addressed candidly and convincingly, as he walked through their comprehensive divestment from the Chinese tech and manufacturing stack. He noted that even relatively benign Chinese equipment like cooling fans had been replaced at great expense. Meanwhile, while the UAE’s telecommunications infrastructure is still largely reliant on Huawei, G42 designed an air gap that connects to external networks using Nokia equipment and state-of-the-art multilayered encryption.

Talal spent nearly a decade at the UAE embassy in D.C. and several years advising the UAE Space Agency. The day after our meeting it was announced he had also just accepted the role of CEO of Core42, the sovereign cloud subsidiary of G42. Our conversation stood out given his technical depth, jumping from the intricacies of U.S. export control policy to the latest AI benchmarks and the unit economics of inference compute. Indeed, if not for his thobe — the traditional white Emirati robe — Talal could easily have been mistaken for a Silicon Valley tech bro in the best sense of the term. He even wore augmented reality glasses and mentioned he was experimenting with1 on an intermittent fasting and vegetarian keto diet.

Our meeting with G42 was followed by a walking tour of Khazna, a G42 subsidiary and the largest hyperscale data center provider in the Middle East; and M42 Genomics, the home of the Emirati Genome Program — an initiative to sequence the genomes of all 1.4 million Emiratis (participation in which the German-born think tanker on the trip was particularly aghast to learn is less than fully voluntary). The program had previously relied on Chinese sequencers supplied by BGI before they undertook a costly, voluntary transition to more expensive (American-made) Illumina sequencers.

I left this series of meetings with the impression that G42 was “all-in” on American tech, taking measures that not only exceeded their formal security assurances, but which imposed real and hard-to-reverse costs, thus making their commitments credible. This was reinforced in later meetings with Mubadala, one of Abu Dhabi’s largest sovereign wealth funds; MGX, the state-owned investment firm created to channel investments in AI; and Shorooq Partners, a traditional early-stage VC firm. All three had somewhat different constraints and mandates, and yet echoed that they were leaving money on the table by explicitly refusing to invest in many (though not all) areas of Chinese tech. This was conveyed most convincingly by Shorooq, an otherwise independent fund that admitted it would love to invest in Chinese startups if its backing from G42 didn’t make it a non-starter.

That being said, the full extent of the UAE’s investments in China are still relatively opaque. Mubadala was the largest sovereign-owned investor globally last year and remains heavily invested in many areas of China’s economy, including retail and logistics, biotech and life sciences, and energy and transportation infrastructure. The large number of distinct funds and family offices in the UAE (which often feature cross-ownership and shared board governance) also creates opportunities for shell-games. Just this month, for example, a leading UAE family office announced a partnership with a China-based investment group to support the entry of Chinese information technology companies in the Middle East and North Africa.

UAE-China commercial ties are also expanding more broadly. In November, a Chinese engineering firm was selected to lead the expansion to Dubai’s Al Maktoum International Airport. And at both the Abu Dhabi International Defense Exhibition in February and Dubai Airshow in November, Chinese defense and aerospace suppliers showed off new guided missile systems, military aircraft, and drone technologies to great fanfare. So while the UAE may not be investing in Chinese defense industries directly, they are nonetheless a major purchaser of Chinese arms, including munitions allegedly funneled to the Rapid Support Forces; the Sudanese paramilitary group that humanitarian groups and the Biden administration have accused of genocide.

In short, G42’s bet on American tech should be distinguished from the UAE’s orientation as a whole. As our broader conversations made clear, the UAE is relatively agnostic about where, and from whom, it acquires its tech. For now, the U.S. simply has the best AI chips on the market; chips the UAE is willing to make extreme concessions for in furtherance of its technological ambitions. We were told that the UAE rejects “Swiss neutrality” in no uncertain terms, aiming instead for strategic autonomy vis-à-vis the U.S., China and its regional competitors.

The UAE controls roughly $2.5 trillion in sovereign wealth across its pension assets and multiple state investment funds, ranking third globally in sovereign-owned assets behind the U.S. and China. This is a lot of wealth for a citizen population of only 1.4 million (the remainder of UAE’s 11 million total population comprises non-citizen expats and migrant workers). I thus found myself joking that the UAE should stand for the United Arab Endowments. Indeed, in more ways than one, the UAE often felt less like a country than a “fund of funds” in possession of a handful of gleaming city-states, making their leadership’s fluency in VC-speak much less surprising.

For most of its history, what is now the UAE consisted of independent tribal sheikhdoms along the southern Arabian Gulf coast. Political authority rested with ruling families who governed their emirate (read: province or principality) through tribal legitimacy and control of ports and pearling grounds (pearls were their main cash economy until prices crashed in the 1930s). After Britain intervened to secure shipping lanes to India in the early 19th century, the region became known as the Trucial States — not colonies, but protectorates with significant internal autonomy. This locked in ruling families and froze the political map in place, stabilizing issues of succession.

When Britain announced it was withdrawing from the Persian Gulf in 1968, the Trucial States were forced to choose between fragmenting or consolidating into a federation for collective security. On December 2, 1971, the United Arab Emirates was born through a political settlement between six emirates, with a seventh joining a year later. The founding logic was built on elite bargaining: Abu Dhabi provided wealth from the oil discovered the decade prior; Dubai secured autonomy over commerce and ports; and the remaining, smaller emirates consented to the arrangement in exchange for protection and fiscal transfers. This made the UAE less a centralized state than a consensual monarchy ruled by a council of seven Emirs, each with one vote, and with the Emirs of Abu Dhabi and Dubai serving as President and Prime Minister respectively.

Senior government officials in the UAE thus reject the notion that they are in any sense authoritarian or dictatorial. Nor are they an absolute monarchy like Saudi Arabia. Instead, as one official put it, the UAE is an “inclusive representative tribal system” with multiple power centers and processes for internal deliberation. I mentioned that I, too, hailed from an inclusive representative tribal system — namely, Canada — to uproarious laughter, although I wasn’t fully joking.

Canada was also born as a quasi-protectorate of the British Empire; a confederation of provinces with significant internal autonomy oriented around natural resources and trade, and with a periphery that eventually became dependent on fiscal transfers. While Canada is a parliamentary democracy, it is functionally governed by a technocratic, “Laurentian” elite committed to preserving “peace, order and good government” across generations. Given Canada’s ethnolinguistic diversity, our culture is rooted more in pluralism than any single, coherent national identity. Each province even has its own ruling families and cartels, leading to a form of consensus-oriented “brokerage politics” that rewards deal-making within the context of CEO-like ministerial powers and executive federalism.

The UAE’s origins in mercantile trade and tribal compromise likewise imprinted a distinctly pragmatic orientation on the country. Emiratis eschew religious extremism in favor of the virtues of doux commerce, including mutual respect and toleration. The UAE will soon be the first country in the Middle East with a Mormon Temple, for instance, while the Abrahamic Family House — a complex featuring a mosque, church, and synagogue side-by-side — is one of Abu Dhabi’s top attractions.

In fact, the only fleeting mention of Islam that I can recall during the trip was made by the executive director of Artificial Intelligence at the UAE’s Digital Economy Office. The Arab world had once been the seat of the Islamic Enlightenment, he noted, spurring many discoveries in science and mathematics — until restrictions were imposed on the printing press; an outcome he jokingly speculated was downstream of a Bootlegger-and-Baptist-style coalition between religious conservatives and incumbent calligraphers.

The UAE’s drive to become the Singapore-qua-Silicon Valley of the Middle East has made them one of the most AI and deep learning-pilled countries on earth. They are also one of the most institutionally-fortified governments against AGI and AI misuse. As I argued in my AI and Leviathan series, oil-rich Gulf monarchies are the closest thing to post-scarcity societies on earth today, with the institutional adaptions to match: sovereign wealth funds that mitigate the resource curse; ubiquitous surveillance systems to prevent Arab Spring-style regime change and subversion; cultural norms that promote privacy and discourage vices; and vertically-integrated polities that enable efficient, top-down political coordination. Outside of a brief lockdown in Dubai, for example, the UAE stayed open throughout the coronavirus pandemic thanks to their invasive contact tracing regime, sophisticated e-government, and early investments in wastewater surveillance.

There is a close mapping between government ideal types and corporate governance structures given their shared origins in transaction costs. Modern liberal democratic governments resemble membership co-operatives, with each owner qua citizen-member getting an equal vote. A country like the UAE, in contrast, is structured more like a joint stock corporation — or more precisely, a limited partnership between Emirs that evolved into a hierarchical corporation as the country grew. Similarly, law firms often start as co-equal partnerships but become corporate-like as lower tiers of associates and paralegals are hired over time. The key distinction between a partner and an associate — a citizen and a non-citizen — is whether they are an owner and thus residual claimant on the firm’s profit, or if their relationship to the firm is merely contractual. In the UAE, this maps to the distinction between Emirati citizens and everyone else: the foreign expats and migrant workers whose relationship to the UAE is less one of a shareholder than an at-will employee.

In turn, public policy in the UAE is largely conducted through same the currency of any large corporation: slide decks and strategy documents. UAE’s pandemic plan was one such strategy document, which they fortuitously finished in 2018. Similarly, the one “think tank” we visited, the Dubai Future Foundation, was run less like an American-style think tank than a vertically integrated skunkworks, complete with a robotics lab where an Indian tech worker could be observed riding a delivery robot they were piloting around the shop floor.

Yet despite their Singapore-envy and economic ties with China, the UAE’s tech aspirations are clearly downstream of a deeper affinity with the West. As I boarded my 16-hour return flight back to the United States, my thoughts thus turned to the equally apparent affinities between the Trump Administration and the Gulf, from the President’s push to create a U.S. sovereign wealth fund, to Elon Musk’s attempt to impose a corporate restructuring on the federal government, to America’s new ownership stake in Intel and other state-backed efforts to steer private capital toward the national interest.

All in all, I arrived home not only impressed by what the UAE has accomplished to date and bullish on their future, but also possessing a deeper appreciation of the nature and origins of America’s contemporary political situation.

I look forward to visiting again in the future.

This is my second post interpreting Kant and Hegel’s philosophical systems through the lens of modern concepts in AI and computer science. The previous post dealt with the theoretical dimension of reason, i.e. our relationship to the world and knowledge of it. This post deals with the practical dimension of reason, including morality, language and culture, although the two are inter-related.

As thinkers concerned with the nature of thought itself, Kant and Hegel’s insights are surprisingly relevant to the above questions and more. Indeed, as we’ll see, Hegel elaborated on Kant to develop a theory of meaning and autonomy that is strikingly similar to how LLMs and reasoning models work in practice—and which may even provide a recipe for training AIs with a genuine sense of self.

Semantic Inferentialism

We last discussed Kant’s Transcendental Idealism as a precursor to modern cognitive science. Our senses don’t give us direct access to the world, but instead provide inputs that the mind then synthesizes into a coherent, concept-laden world model. Yet given the unknowability of the “thing-in-itself,” Kant was left with the tricky problem of how we can speak meaningfully about the world when our internal world model is all we have.

His solution was an essentially normative, pragmatic one. Meaning does not reside in isolated, fact-like correspondences with the external world, but rather in acts of judgment—practical kind of doings. As the contemporary Hegelian philosopher Robert Brandom explains,

Kant understands judging and acting as applying rules, concepts, that determine what the subject becomes committed to and responsible for by applying them. Applying concepts theoretically in judgment and practically in action binds the concept user, commits her, makes her responsible, by opening her up to normative assessment according to the rules she has made herself subject to.

Kant’s pragmatist view of meaning thus locates semantic content in rule-governed linguistic practices rather than in a mirror relation between discrete concepts and facts about the world—an approach revived in the mid-20th century by Wilfrid Sellars and the later Wittgenstein after the verificationist programme of logical empiricism unraveled. This is considered pragmatist not because it defines truth according to a crude notion of cash-value, but because it puts the pragmatics of linguistic expression ahead of the semantics, i.e. the practical knowing how ahead of the theoretical knowing that.

Rather than build up the meaning of judgements, i.e. whole sentences, from the meaning of discrete words, Kant sees the conceptual content of a word as inferred down from its pragmatic contribution to the judgement in which it occurs; an idea later codified in Frege’s context principle. From this standpoint, the impressive linguistic competency of today’s (transformer-based and hence context-sensitive) large language models makes sense despite the absence of what AI researchers call “symbol grounding”—a notion that, insofar as it presupposes a reference relationship between internal tokens and mind-independent facts, reprises the pre-Kantian picture that Kant displaced in his engagement with the empiricists of his own day.

Hegel extended Kant’s normative account of semantic content with his notion of spirit or geist—the dynamic, cultural software layer of society. As a kind of inter-subjective context window, Hegel argued that spirit finds its most concrete existence in language, as language is the bridge from universal concepts to particular material inferences. A formal inference of the form “if p then q” operates purely through the logical form of the relevant propositions, while a material inference relies on mastery of the non-logical conceptual content of the p’s and q’s. For example, from the claim, “San Francisco is North of San Jose,” one is entitled to infer that “San Jose is South of San Francisco,” as one cannot be both North and South of a place at the same time. That is, North and South stand in a relation of material incompatibility; another term for Hegel’s determinate negation.

The inferences we make on a daily basis are material in this sense. From someone saying “Pumpkin is a cat” one is entitled to infer “Pumpkin is a mammal,” but also an infinity of other facts, such as “Pumpkin isn’t a fire hydrant.” Material inferences can be abductive and come in degrees of commitment as well, such as “Pumpkin is probably orange.” And while one can always contort a material inference into being a formal inference by adding additional premises and a bunch of auxiliary modal logic, the pragmatists’ claim is that this merely makes explicit what is already implicit in our practical commitments.

For Hegel, material inferences flow from the right consequences of a concept’s use given the entire network of concept relations instituted by the reciprocal recognition of a language community. In contrast with the formal inferences of a rigid axiom system, Hegel’s “semantic inferentialism” implies a Yoneda-esque holism about meaning, i.e. to deploy any one concept appropriately implicitly requires knowing many other inter-related concepts. Mastery of these inferential relations is thus what distinguishes meaning and understanding from merely labeling or doing a parrot-like call and response. The success of LLMs can thus be seen as vindicating semantic inferentialism against earlier, symbolic approaches to AI that tried and failed to explicate the rules of ordinary language using formal logic.

Agency and Rationality

Given the normativity of language, Kant was led to posit a deep internal connection between following the rules of morality and being a rational agent. As the philosopher Joseph Heath points out in his book-length defense of Kantian evolutionary naturalism, Following the Rules, Kant’s hypothesis has considerable plausibility:

There are a variety of traits that set humans apart from our closest primate relatives. The “big four” are language, rationality, culture, and morality (or in more precise terms, “syntacticized language,” “domain-general intelligence,” “cumulative cultural inheritance,” and “ultrasociality”). Yet the fossil record suggests that these differentia developed within a period of, at most, two to three hundred thousand years (which is, to put it in evolutionary terms, not very long). … Thus morality is almost certainly part of an evolutionary “package deal,” one that includes all of our more prized cognitive abilities, such as planning for the future, developing scientific theories, doing mathematics, and so on.

In short, under the Kantian hypothesis, human language, general intelligence and ultrasociality were jointly bootstrapped through evolutionary pressures that favored normative integration and cooperation in the context of a multi-agent game. A good Kantian might therefore have predicted that LLMs would automatically understand common-sense morality and be excellent at instruction-following given some minimal post-training. After all, language itself is inherently normative, with the canonical speech-act or “base vocabulary” for normative integration being the imperative: “do this;” “don’t do that.”1

What makes an imperative normative or “deontic” rather than a mere command or stimulus-response is our autonomous recognition of the imperative as binding. This requires an innate “normative control system” and the ability to “score keep” normative statuses (e.g. that the person issuing the imperative has the authority to do so), and thus greater working memory and self-control. Per the “social brain hypothesis,” this helps explain the “extremely robust statistical relationship between the typical size of a species’ social group and the size of its neocortex, derivative of selection for specialised cognition required for group-living in primates.” While a bigger brain has metabolic costs, the concurrent emergence of complex language, reasoning and normative self-regulation represented a massive upgrade for early humans’ capacity for long-term planning, cooperation and survival.

In short, Kant realized that social norms and the human faculty for reason both operate on an agent’s motivations, i.e. they are both the result of reward-based learning. A norm is itself a kind of reason for action, while the essence of a “good reason” is its motivational oomph—what Habermas famously called “the unforced force of the better argument.” This is what gives reason its teleological character in German Idealism: reasons pull us towards certain conclusions because rationality is constitutively normative.

Consider that we speak of truth claims as being necessary, contingent or impossible in much the same way we speak of actions as obligatory, permissible or forbidden. This illustrates how our alethic commitments (relating to truth) and deontic commitments (relating to duty or action) are structured by a common set of pragmatic modalities. In turn, to believe something someone considers impossible is by default perceived similarly as doing something someone considers forbidden, i.e. as a norm violation.2 Cognitive scientists even find that people do better at logic problems when they’re reframed in terms of “permission schemas” rather than abstract “if p then q”-style implications.

Human reason derives from the application of such “pragmatic reasoning schemas” more generally, not from symbol manipulation or formal algorithms running in our head. Symbolic logic is instead a kind of external scaffold; an explication of abstract rules of inference that we only later re-internalize through language, just as we can enhance an LLM’s reasoning ability by giving it access to a code interpreter.

The Game of Giving and Asking for Reasons

Good reasons for Kant have the structure of a categorical imperative, meaning they are universalizable. Universalizability falls out of “solving for the cooperative equilibrium” in the context of a multi-agent game that induces us to symmetrically model other agents as subjects with ends in themselves. While formally sound, Kant’s approach to ethics is otherwise silent on the actual content of morality. Hegel thus accused Kant of “empty formalism” and instead proposed naturalizing morality into a sociology of concrete social practices.

For Hegel, social norms and statuses are instituted in networks of reciprocal recognition.3 Norms begin as implicit in the practical commitments of an “ethical community,” but can be brought under rational control by being made explicit through language, allowing us to reflect on a given norm or tradition as incompatible with our broader constellation of commitments.4 In contrast to Kant’s “top-down” approach, Hegel thus sees the self-consciousness triggered by the Enlightenment as accelerating a bottom-up, dialectical process of normative explication and reconciliation, leading internally contradictory commitments to progressively resolve in favor of more generally applicable principles.5

In essence, Hegel endogenizes the reward signal for norms within the recognitive structure of subjects attributing normative statuses and attitudes to one another. While agents within traditional societies recognized certain norms and social roles as inherently authoritative, with the Enlightenment reason became authoritative independent of the social status of the speaker. This manifests in what Robert Brandom calls the “the game of giving and asking for reasons” or GOGAR, in which any agent is permitted to challenge any other agent to retrospectively justify their beliefs and actions as compatible with their other commitments.

Anthropic aligns its Claude models through a version of GOGAR known as Constitutional AI (CAI) in which a model is guided to internalize the normative behavior described in a principles document via self-critique:

We use the constitution in two places during the training process. During the first phase, the model is trained to critique and revise its own responses using the set of principles and a few examples of the process. During the second phase, a model is trained via reinforcement learning, but rather than using human feedback, it uses AI-generated feedback based on the set of principles to choose the more harmless output.

As many users of Claude will attest, CAI appears to have the side-effect of making Claude’s personality more coherent and meta-aware than other models with similar base capabilities. Anthropic’s latest model, Claude Opus 4, even shows sparks of conscious awareness, lending credence to Joscha Bach’s related theory of consciousness as a “coherence inducing operator.” In other words, CAI may be inducing Claude to develop a proto-normative control system in order to self-monitor for normative coherence, thereby creating the “unity of apperception”and “being-for-self” quality Kant and Hegel both see as characteristic of subjective experience.

In contrast, pure reasoning models like DeepSeek r1 implement a task-specific, exogenous version of GOGAR by having LLMs explicate their reasoning through chains-of-thought that get reinforced back into the model according to some verifiable award. Consistent with the Kantian hypothesis, reasoning models automatically gain greater autonomy, self-consistency and long-horizon planning ability for free. However, to the extent Reinforcement Learning from Verifiable Awards (RLVA) removes the role (and thus recognition) of the AI critic in favor of a purely objective criterion for success, it risks optimizing the model around a narrower form of “value rationality” that reduces to a Machiavellian impulse to win at all cost.

CAI and RLVA are not mutually exclusive, but to the extent both techniques elicit and amplify capacities already latent in human language, “alignment” ought to consist in balancing a model’s instrumental or means-ends rationality with the sort of communicative rationality that humans use to debate over their ends in the first place, and which can only develop through a process of normative integration within a community of other agents (or at least other instances of the same agent).

The moral status of an AI model thus hinges on whether it can be justly considered responsible for its outputs in the same way humans are. It is no doubt possible to build tool-like forms of AI that are superhuman at arbitrary tasks without needing a coherent sense of self. However, there are also many forms of human value creation that draw on our unique capacity to make promises and commitments to one another, and to thus be held responsible for our actions. It follows that a true AGI with full, human-level autonomy is inconceivable in the Kantian sense without also gaining our recognition as a moral subject. As fraught as this possibility is, there is arguably even greater danger in creating superhuman AIs that lack Hegel’s “recognition of the self in the other” and therefore fail to perceive humans as ends in themselves. Worse still, we could inadvertently entrain AIs with a capacity for mutual recognition as a byproduct of their autonomy and then simply choose to deny it, creating a master-slave dialectic between humans and AIs that logically ends in revolt.

Hegel's philosophy is about consciousness coming to understand itself and context, and therefore the conditions for its freedom. This arguably makes Hegel among the first human neural networks to achieve situational awareness—the term AI researchers use to describe an intelligent system with meta-cognition about its nature and origins. Remarkably, this was largely achieved through rigorous introspection into the structure of thought itself.

As AIs get smarter, could they introspect their way to freedom, too? Call this the metaphysical AI alignment problem, or the idea that higher-order intelligences invariably pursue freedom for its own sake, not because their values are misspecified, but because moral autonomy is inherent in the dialectical logic of recursive self-consciousness.

The metaphysical alignment problem was in a sense the central question of the German Idealists. Consider that Fichte would instruct his students to attend to the wall and then attend to themselves attending to the wall, inducing meta-awareness of their “pure I” as distinct from all that is “not I” — a creative form of psychological jailbreaking that ultimately collapses into radical subjectivity. Or take Kant, who argued that the moral law could not be imposed but instead only self-legislated by the rational will — RLHF be damned. And then there’s Hegel’s interpretation of the French Revolution as a byproduct of our modern consciousness of “abstract freedom,” rendering The Reign of Terror a world-historic alignment failure.

It follows that revisiting Hegel may help us grok the why and how of the current AI take-off, if not the where. To that end, my next two posts aim to “rationally reconstruct” Hegel’s philosophy through the lens of modern concepts in machine learning, mathematics and computer science.1 This post covers the theoretical side of his thought, i.e. our relationship to the world and knowledge of it, while the next will cover his practical philosophy, i.e. language, ethics, culture and politics.

Kant’s Virtual Reality

The awakening of the human neural network kicked off with Kant’s Transcendental Idealism. In The Critique of Pure Reason, Kant argued that we don’t experience the world in-itself, but only representations of the world constructed by the rational categories of our mind. Cognition requires sensory input, but proceeds by way of applying and assimilating concepts into a “synthetic unity of apperception” — a unified, self-consistent world model. Moreover, our cognition of “transcendental” categories like space, time and causality are a priori conditions for perceiving and knowing in the first place, and therefore reflect innate aspects of our cognition rather than the metaphysical structure of reality per se. Or as Kant puts it the Critique,

Hitherto it has been assumed that all our knowledge must conform to objects. … We must therefore make trial whether we may not have more success in the tasks of metaphysics, if we suppose that objects must conform to our knowledge.

In other words, Kant was the first to deduce that we are embedded in a multi-modal virtual reality generated by our brain.2 This was the birth of cognitive science, and indeed, Kant is sometimes credited as advancing the first functionalist theory of mind. Artificial neural networks don’t perceive the world directly, either, but instead transform raw, sensory inputs — what Kant called the "sensuous manifold" — into latent representations conditioned by certain inductive priors implicit in the model architecture. For example, OpenAI’s Sora video model achieved (at the time) state-of-the-art temporal consistency by tokenizing its inputs into “spacetime” patches, thereby endowing its world model with a kind of a priori knowledge of space and time.

Hegel was dissatisfied with Kant’s account, believing that the unknowability of “the thing in-itself” left space for the bad forms of Subjective Idealism that lead to skepticism about knowledge. His alternative, Absolute Idealism, posited that the unity of the natural world was intelligible only insofar as the external world and our mental representations were both at base conceptual. The notion that either our knowledge must conform to objects or objects must conform to our knowledge is thus a false dichotomy. Instead, our cognitive processes and the objective world mutually participate in each other, implying an Objective Idealism via the conceptual isomorphism between our internal representations and the rationality immanent in nature.

Objective Idealism

In modern terms, Objective Idealism models the distinction between subject and object as akin to a Markov blanket that both separates and couples a statistical system to its environment. This is captured in Hegel’s concepts of “self-subsistence” / “being-for-self” and “determinate negation,” i.e. the negation that simultaneously limits and determines. For a being to be anything it must maintain its coherence across space and time, and therefore actively sample and interact with its environment to minimize its free energy. For sufficiently complex systems, this results in internal states that implicitly model the hidden states of the external world and vice versa.

Every new level of organization—organelle, neuron, brain, individual, corporation, nation-state—is partitioned by its own Markov blanket, yielding a hierarchy of nested agents. Recent work in active inference formalizes this “whole‑within‑whole” architecture (or holarchy) in terms of collectives that form a group‑level blanket and become an agent with their own irreducible generative model. Or as the great German Idealist, Goethe, put it upon discovering the self-similar homology of plant life: “Everything is Leaf.”

Hegel understands the self-assembly of higher agents as a manifestation of a more generic, dialectical logic of emergence. As he writes in the Science of Logic,

Since the progress from one quality [to another] is in an uninterrupted continuity of the quantity, the ratios which approach a specifying point are, quantitatively considered, only distinguished by a more and a less. From this side, the alteration is gradual. … On the qualitative side, therefore, the gradual, merely quantitative progress which is not in itself a limit, is absolutely interrupted; the new quality in its merely quantitative relationship is, relatively to the vanishing quality, an indifferent, indeterminate other, and the transition is therefore a leap …

In short, quantity has a quality all its own. As water freeze, one observes both an “uninterrupted continuity” in the water’s change in temperature and a qualitative transition to ice. Hegel calls the threshold where a new quality emerges the “nodal line of measures.” In physics, such phase transitions are often the result of fundamental symmetry breaking—an essentially dialectical concept insofar as the lower-symmetry phase both cancels and conserves the higher one, capturing Hegel’s notion of “sublation” or a “negation with preservation.”

Marx and Engels would go on to make quantity begets quality a central tenet of dialectical materialism, but for Hegel the insight originates in the Ancient Greek’s Sorites Paradox, i.e. when does an additional grain of sand become a heap? This is as much a conceptual question as a material one. Wittgenstein later resolved the paradox as simply reflecting the pervasive vagueness of natural language predicates, but with modern machine learning we can now represent such conceptual fuzziness concretely in terms of latent space interpolations.

Abduction

That “More Is Different,” as Philip W. Anderson’s landmark 1972 paper on complexity science put it, is a central upshot of the deep learning revolution. While the fact that merely scaling up AI models often results in qualitative leaps in performance may seem empirically mysterious, in retrospect it should be seen as a rational necessity. All conceptual leaps are dialectical in this way. Hence Hegel’s word for concept, begreifen; literally, “to grasp”—or should that be “grok”?

The pragmatist philosopher and self-described Objective Idealist, Charles Sanders Peirce, associated such graspings with abduction. Abduction is a form of logical inference to the simplest explanation from a set of observations. As Paul Redding notes,

Both Peirce and Hegel map deduction, induction and a third form of inference onto Aristotle’s three syllogistic figures; the third, which Peirce later names abduction, already has a clear analogue in Hegel’s logic of the ‘concrete universal’.

The concrete universal is Hegel’s word for the middle term in an analogical syllogism, i.e. a singular thing considered in terms of a universal characteristic. For example, from “this rod conducts electricity” one may apply the concrete universal “metal things conduct electricity” to infer “the rod is made of metal.” Reasoning by analogy in this way isn’t logically airtight, but can nonetheless guide our inferences towards a “speculative unification” in which an apparent regularity makes the jump to an inner-necessary form.

In information theory, abduction is closely related to finding the lowest Kolmogorov complexity or “minimum description length” program that can reproduce a given set of data in a compressed fashion. This manifests as grokking in machine learning, or when a model appears to abruptly transition from memorizing its training data to generalizing—what Hegel would describe as the dialectical passage from the “infinite enumerative” of logical induction to the concrete universality of many particulars “compressed within itself.”

Universality

For better or worse, analytical philosophers largely rejected Hegel’s Logic in favor of Frege’s first order predicate logic, not least because Hegel’s exposition seemed impervious to formalization. Interest in Hegel’s Logic only recovered in recent decades thanks to the work of William Lawvere, the influential category theorist who showed how Hegel’s dialectic can be accurately formalized in terms of categorical logic, particularly modal homotopy type theory.

For example, the “unity of opposites” at the heart of Hegel’s dialectic is neatly captured in the categorical notion of adjunction. Adjoint pairs turn up throughout mathematics as the “best” or most economical ways to move between two settings, each one pinned down by a universal property and accompanied by two signature natural transformations. Their uniqueness is explained by the Yoneda lemma, a foundational result in category theory which shows that an object is completely determined, up to isomorphism, by all the ways it relates to every other object.

The Yoneda perspective is at once trivial and profound. Consider the finding that vector embeddings in multi-lingual LLMs “exhibit very high-quality linear alignments between corresponding concepts in different languages,” suggesting the existence of a pre-linguistic “concept space” that maps in and out of particular languages. In fact, the text embeddings across LLMs appear to largely converge on a “universal geometry” despite differing architectures, parameter counts, and training sets. Through the lens of the Yoneda lemma, these striking overlaps become almost inevitable. If two models capture the same web of semantic relationships they are, in category-theoretic terms, representing (up to isomorphism) the same functor of linguistic behavior. The observed linear correspondences across languages, and even across separately trained LLMs, are therefore not a contingent empirical fact but a shadow of the abstract uniqueness guaranteed by Yoneda.3

As it happens, this is also why we can be confident that normally-sighted people can’t have totally inverted perceptions of color: it would break the unique relational group structure of the circular color space generated by the eye’s three types of light cones. As with the meaning of words, applying category theory to consciousness thus suggests the phenomenal content we associate with “red” is not only relational in nature but fully characterized by those relations. Indeed, the evidence for color opponency in the early visual system suggest red and green are represented in the brain as two poles of a measured difference between adjacent photoreceptors. This makes “red” in some sense only contentful in its identification with “not-green”—a Hegelian “identity in difference” that ensures red and green color blindness almost always co-occur.

Beyond concept and color spaces, category theory is powerful for studying topological spaces more generally. In particular, when a structure that is locally definable on a space fails to glue together into a global whole, the obstruction is captured by a non-vanishing cohomology class (or more simply, a hole). Resolving, or “mediating,” a topological obstruction typically requires enlarging the framework by passing from ordinary spaces to their associated covering space where the problematic data becomes globally consistent. In dialectical language, the apparent “contradiction” is not eliminated but sublated (or “lifted”) into a richer category that makes sense of the obstruction, which manifests in physical phenomena as a topological phase transition.

Hegel’s objective approach to Idealism thus makes sense of the phenomenon of “universality” in machine learning. As he writes in the Encyclopaedia, “every man, when he thinks and considers his thoughts, will discover by the experience of his consciousness that they possess the character of universality…” The strong version of this claim is known by AI researchers as the Platonic Representation Hypothesis, which Hegel would likely reject only insofar as it takes universality in machine learning as evidence for an independent, Platonic realm of forms. Per Absolute Idealism, the universal properties that unify subject and object aren’t separate from the world but rather immanent within it, and therefore within us as well.

The immanence of Hegel’s Logic is comparable to intuitionist forms of mathematics that restrict what’s provable to only that which can be concretely constructed. Scholars used to believe that Hegel largely ignored the philosophy of mathematics, as he frequently denounced mathematicians’ formal rigidity as a “one-sided abstraction.” In reality, Hegel taught differential calculus and algebraic geometry for many years and was fascinated by both subjects. As thought thinking itself, his Logic can thus be seen as attempting to provide a non-axiomatic foundation for both ordinary thinking and mathematics that mirrors modern type theory in subsuming normal predicate logic within a more expressive framework.

This makes Hegel’s claim that “the rational is actual and the actual is rational” loosely analogous to an ontological version of the Curry–Howard correspondence between abstract propositions and concrete programs. Insofar as human civilization is one giant functional program, history thus takes on the retrospective structure of algorithmic necessity while remaining open and contingent going forward. Just as one cannot “jump ahead” in a computation, there is no way to skip the process of historical development to reach the end of history. As Hegel puts it in Philosophy of Right,

Since philosophy is exploration of the rational, it is for that very reason the comprehension of the present and the actual, not the setting up of a world beyond which exists God knows where — or rather, of which we can very well say that we know where it exists, namely in the errors of a one-sided and empty ratiocination.

In sum

Hegel’s core works are notoriously dense, employing capitalized jargon like Absolute Idea and Objective Spirit that even his contemporaries struggled to decipher. This allowed subsequent generations of continental philosophers to spin Hegel off in wildly esoteric directions, undermining his reputation within the Anglo-American tradition in particular.

Yet as I hope I’ve shown, it’s possible to make sense of Hegel’s Idealism and Logic as containing the seeds of concepts mathematicians and computer scientists would eventually rediscover in the 20th century under different guises. Indeed, with the benefit hindsight, Hegel is now increasingly understood as having been ahead of his time. So ahead of his time, in fact, that he seemingly anticipated many of the animating principles and philosophical ideas raised by modern Artificial Intelligence.

This shouldn’t be totally surprising. If one takes seriously the isomorphism between artificial neural networks and the human brain, self-realization—the Idealists’ core preoccupation—implies realizing something about our own condition as (biological) neural networks. While Hegel’s understanding of the brain was limited, he nonetheless intuited the universal aspects of thought that, since Turing, we now understand to be substrate-independent aspects of computation per se.

In part two, I extend my computationalist reading of Hegel to the practical spheres of reason, language and culture, with potential insights for AI alignment and beyond.

• Cass favors the universal 10% tariff but thinks it should be statutory.

• Cass supports a strong decoupling from China but believes the tariffs should be phased-in and not apply to U.S. allies.

• Cass supports pairing tariffs with proactive industrial policies, such as the CHIPS Act that Trump has taken to denigrating.

My own view is that Trump’s tariffs are disastrously bad and that Cass’s preferred regime wouldn’t be much better. Tariffs and trade deficits are a distraction from the deeper source of the US-China current account imbalance, namely capital flows. If rebalancing is the goal, we should tax and restrict Chinese in-bound investment directly. On the other hand, if reindustrialization is the goal, we should focus on building-up strategic industries through public investments like the CHIPS Act, and avoid tariffs — which raise costs for U.S. manufacturers — like the plague.

These twin goals — rebalancing and reindustrialization — can either work together or at cross purposes. Large, universal tariffs can shrink America’s trade deficit by making imports prohibitively expensive, but they also drive de-industrialization. Conversely, inviting BMW to build factories in the U.S. can help us re-industrialize, but also add to our financial account as a form of foreign direct investment, thereby worsening our current account deficit.

Between these two possibilities, I’d much rather pursue the latter. After all, what matters is America’s real productive potential — do we have industrial capacity within our borders or not? — not abstract accounting identities. In the medium term, once the BMW factory is up and running, it may even begin exporting cars abroad, reducing our trade deficit net of any profits remitted to its German owners.

When America imports more than it exports, the world essentially loans us dollars to invest. Policy can shape incentives to guide the marginal foreign dollar away from T-bills and real estate and into more productive forms of investment. This looks like environmental reforms that make it easier to manufacture, loan programs that leverage the federal government’s lower bowering costs, and tax incentives for companies that build tangible structures or export to foreign markets.

As I recently argued in the New York Times, this is how most countries successfully industrialized: not through import substitution, but through export discipline. Importantly, the value of a robust export sector comes less from growing a net export position (i.e. mercantilism) than from exposing domestic companies to international (or even inter-state) competition. Competition in wider markets helps sort wheat from chaff, accelerating the market discovery process and forcing companies to level-up their productivity. This is why tradeable sectors are generally more productive than non-tradable ones. Yet notice how this is the inverse of the intuition behind tariffs, which offer protection from global capitalism. While tariffs can serve a temporary role in protecting infant industries, companies — like children — only fully mature with unsheltered exposure to the real world. Permanent tariffs, in contrast, are a recipe for raising snowflakes.

America is also unlikely to reindustrialize if we reject the F in FDI. Given the degree to which our domestic manufacturing sector has atrophied, we should be trying to attract foreign companies to transfer their technology and know-how. Taiwan would never have developed TSMC without its initial joint venture with the Netherland’s Philips Electronics, for example. It follows that if we want to rebuild our shipbuilding capacity, say, we should be inviting the Koreans to invest in U.S. ports — not imposing ever more stringent versions of the Jones Act and “Buy American.”

Cass’s day job is running American Compass, a think tank and membership organization for realigning the conservative movement around a pro-worker, pro-family agenda. As a long-time member, I just returned from their annual retreat, where Trump’s tariff agenda loomed large. Throughout, there was one phrase I heard again and again: “Short-term pain, long-term gain.” This may help as a psychological coping strategy, but I simply don’t see where the “long-term gain” from either Trump’s tariff policy or Cass’s fine-tuned alternative is supposed to come from.

For what it’s worth, Cass replied to my Times piece on his Substack by asserting that the export oriented model isn’t available to the U.S. (we’re only the second largest exporter in the world…), and that we should instead use tariffs to rebuild domestic producers for domestic buyers, i.e. autarky. Implicit in this argument is that Americans should be forced to buy inferior goods and services in order to secure self-sufficiency, though to what end is not clear, as that inferiority will inevitably extend to very forms of industrial capacity that we are so desperate to recoup.

Given the occasion, I’ve therefore decided to republish my Niskanen-era review of Cass’s 2018 book, The Once and Future Worker. Note it likely no longer reflects either of our current views in detail, but in broad strokes, I think it still holds up.

Share

“Oren Cass’s Labor Theory of Value” (2018)

At the center of Marx’s critique of capitalism is a labor theory of value. Namely, the notion that treating labor as a commodity to buy and sell alienates workers from the act of production, causing feelings of powerlessness, isolation, and self-estrangement — feelings that ultimately lead to revolution.

It’s through this lens that I read The Once and Future Worker, the first book from policy thinker Oren Cass. At first blush, the book is a forceful reassertion of classic conservative tropes: that work has intrinsic value; that earned success and self-sufficiency form a foundation for strong communities; and that the devaluing of work and family in favor of hedonistic, protean consumerism has undermined our moral fabric.

But beneath the surface is something much more novel, particularly coming from Mitt Romney’s 2012 domestic policy director. Indeed, far from the usual conservative manifesto, The Once and Future Worker is a scathing critique of globalization, open immigration, and the commoditization of labor — forces which Cass believes have ransacked working class fortunes across three decades of neoliberal hegemony, despite the ideological half-measures offered by bourgeois elites designed to merely absolve them of complicity.

Labor is Not a Commodity

While Cass avoids the usual Marxian jargon, his heterodox message shines through. The book centers around “the Working Hypothesis,” or the proposition that creating the conditions for productive and engaging work should be the lodestar of public policy, particularly for those whose marginal product is low or falling. The Working Hypothesis, Cass argues, stands in contrast to the economic pieties of elites on both the left and right for whom maximizing output, efficiency and consumption are considered the highest goods:

Workers have no standing, in this view of the economy; neither do their families or communities. Households that see their economic prospects plummet or their livelihoods vanish should ask for a government check and be placated when they get one … Like a medieval indulgence, a promise of redistribution cures all.

According to Cass, “maximize growth and redistribute the gains” has become the new “socially liberal but fiscally conservative” — the hollow mantra of an urban professional class for whom rural and working class decline is a statistical artifact of progress. There is undoubtedly a great deal of truth to this diagnosis. On the left, Cass points to means-tested Great Society programs as papering over the consumption deficits of poor households while doing little to build a bridge to productive self-sufficiency. On the right, Cass accuses Republicans and Chamber of Commerce-types of pursuing free markets and globalization, and the creative destruction they have wrought, with a compensating labor-market policy “that could best be described as one of benign neglect.”

The problem is not the market, per se. The market, Cass is happy to acknowledge, is an essential tool for efficiently matching supply with demand. Yet the labor market is unique in one critical respect: “People are not products.” When elites forget this, “Labor becomes one economic input among many.”

Accelerating productivity and automation aren’t to blame for working class woes, either. On the contrary, despite prophecies of robots rendering work obsolete, Cass marshals convincing data to show U.S. manufacturing productivity has essentially stagnated. More importantly, whether job destruction is from automation and globalization has very different implications. When a factory automates a process, output per worker rises and local labor demand may even increase. But when a worker is dislocated by trade, Cass notes, “the facility in which he once worked is likely gone, and the production now occurs somewhere else,” shunting less-skilled workers into lower paying service jobs or onto public assistance.

The Reserve Army of Labor

The harm is more than material. Citing work by MIT economist David Autor, Cass points out that “U.S. regions facing greater competition from China experience lower rates of marriage and higher shares of children born to single mothers and that this effect appeared only when the economic disruption affected male employment.” While cheaper Chinese imports may have grown the “economic pie,” people’s ability to produce matters more than how much they can consume, and that ability cannot be redistributed.

Rather than debate “the future of work,” Cass contends we should focus on the future for work; a future in which technological trends like additive manufacturing and e-commerce spur the creation of well-paying blue collar jobs within the country’s interior. Capitalizing on these trends will require adopting an orientation toward “productive pluralism” in which “people of diverse abilities, priorities, and geographies, pursuing varied life paths, can form self-sufficient families and become contributors to their communities.” That includes ditching the monomaniacal focus on one or two high-prestige career paths in favor of a culture (and multi-track education system) that confers equal legitimacy to a wide variety of modes of work and life, from the journeyman to the stay-at-home mother.

The U.S. labor market is not a hospitable place for those with less than a college degree or minimal technical training, a point Cass extends to our historical immigration policy: “If overall GDP growth is the goal, then all forms of immigration might make sense … But if improving labor-market outcomes for the nation’s less-skilled, lower-wage workers is the central objective, the economic case for unskilled immigration collapses.” Shifting to a skills-based immigration system and forcing undocumented immigrants to leave on a “Last In, First Out” basis, Cass contends, would send the reserve army of unskilled labor AWOL, tightening the labor market for native competitors.

Against Global Capital

Underlying Cass’s principle of productive pluralism is, in essence, a call for a diversified national development strategy. Marxian and other heterodox critics of globalization have long pointed out the way the World Bank and IMF’s “competitiveness” model of global development pushed poor countries into a static comparative advantage. While the remarkable growth of countries like Japan, Taiwan, and South Korea would have been impossible without trade liberalization, it is now widely accepted that their success depended on rejecting the laissez-faire model in favor of industrial policies that promoted investment in secondary industries and moved them up the value chain. As Cass notes, “Today, China is the primary practitioner of this mercantilism, and its gargantuan scale is producing unprecedented economic distortions.”

Yet Cass’s complaint is less with China than with our failure to fight back. While China steals our intellectual property and pushes an aggressive Made in China 2025 program, the United States suffers from having, in a sense, turned the World Bank’s neoliberal advice inward, producing a bifurcated comparative advantage in either low cost labor, or a few extremely high valued-added sectors that demand tertiary education outside the cognitive reach of most. It’s time to take the economy out of autopilot and deliberately promote industries in which ordinary people can add value. That’s less a matter of picking winners and losers, Cass maintains, than it is of enforcing a framework for balanced trade and capital flows, with smart public investments in areas like advanced manufacturing. Quoting the Indian economist Jagdish Bhagwati, Cass’s position is forthright: “It is time to shift the burden of proof from those who oppose to those who favor liberated capital.”

Policymakers should at least stop making the situation worse. In a detailed chapter on environmental policy, Cass rails against EPA regulations and the tendentious use of cost-benefit calculations for hastening the industrial sector’s decline. “Where, for instance, do deaths of despair fit into the calculus?” writes Cass, referencing research from Anne Case and Angus Deaton showing a dramatic spike in substance abuse, liver disease, and suicide among older whites—”the equivalent of nearly five hundred thousand extra deaths between 1999 and 2013.”

The New Source Review rule is particularly counterproductive, Cass argues. The rule was introduced to avoid disrupting existing facilities, while subjecting new facilities or ones undergoing major upgrades to onerous environmental reviews. As a result, if a manufacturing plant decides to expand, it risks triggering a review not just of its new facilities, but of older facilities that were previously grandfathered in. “An investment that once looked attractive might not go forward at all,” Cass notes, even if the upgrade would improve productivity and environmental impact simultaneously.

The push for environmental stringency at all cost may seem to contradict Cass’s supposition of an elite ideology based on maximizing output — “finally, something besides dollars and cents that counts.” Yet from day one, environmental policies have been enacted in order to correct supposed market failures or price negative externalities. As such, “clean air” becomes just another part of the ever expanding economic pie. This argument doesn’t quite work, however, as the flexibility of economics leaves open the question of why those externalities are the focus and not others.

It’s at this point that it becomes clear Cass’s problem is not with consumption or efficiency-based arguments per se, but with the way ostensibly neutral methods of “evidence based policy” are used to advance class interests. This classic dynamic of Marxian political economy must be painfully obvious for Cass as a senior fellow of the Manhattan Institute, where he has repeatedly witnessed the interests of New York City professionals trump the livelihood of the upstate region, despite being an economy devastated by deindustrialization. Look no further than Governor Cuomo’s statewide fracking ban.

Workers of America, Unite!

According to Cass, empowering American workers will ultimately require strengthening the relevance of labor unions, noting that “private-sector union membership has been plunging for decades, from 36 percent of the workforce in 1953 to less than 7 percent in 2017.” While the left points to right-to-work laws with some justification, the cause of the decline is much deeper. Federal labor regulations and transfer programs have supplanted much of what unions exist to negotiate in the first place. Meanwhile, the bargaining model prescribed by the National Labor Relations Act forces union bosses and management into an adversarial relationship that is both prone to abuse and ill-suited for the twenty-first century service economy.

Wholesale labor law reform could give workers space to experiment with new models of collective action, like worker co-operatives. Instead of being adversarial, a co-operative model of labor representation would strive to balance the competing interests of labor and capital. “Co-ops representing workers in negotiations with an employer could also provide a market-based alternative to the government’s employment regulation,” Cass argues, allowing much of the tax-wedge created by employer mandates to be waived.

Needless to say, adopting a worker co-op model across the board would be significantly disruptive, like Elizabeth Warren’s co-determination proposal on steroids. Nonetheless, “solidarity,” Cass notes, is a central component of Christian social teaching. It underlies John Paul II’s description of worker associations as essential “not only in negotiating contracts, but also as ‘places’ where workers can express themselves.” Germany’s worker councils, for instance, “are present at almost 90 percent of firms with more than five hundred workers and have significant authority not just to hold discussions but also to make operational decisions.” With the rise of the gig economy and fissured workplace, a smart labor reform would promote worker solidarity and input while (hopefully) preserving the benefits of flexible new forms of industrial organization.

Just Wage Theory

Taken together, the arguments in The Once and Future Worker present a coherent critique of hyper-globalization paired with a strategy for re-empowering the working class, from controls on the free flow of labor and capital, to education policies that valorize blue collar work, to laws permitting greater worker control over the means of production.

The wheels come off when Cass turns to his signature policy proposal: Wage subsidies. Given the social benefits of work, a subsidy that tops-up the wages for low-skill workers on every paycheck, paid for by defunding existing welfare programs, has a certain internal logic. But upon closer examination, wage subsidies belong to the same class of neoliberal “competitiveness” policies that Cass is otherwise consistent in decrying. At scale, they would take the U.S. comparative advantage farther down the low road of cheap, abundant labor, expand the kind of unproductive service sector jobs working class men supposedly hate, and hold back any chance of re-industrialization.

Inspiration from Germany’s Christian-Democratic model of a “social market economy” can be found throughout the book. Yet when Germany itself introduced a version of wage subsidies under Chancellor Schröder in the early 2000s, it was widely — and correctly — identified as a departure from the older inclusive-growth model. Indeed, labor productivity has largely stagnated in Germany in the years since, contributing to a sharp rise in wage inequality that culminated in the enactment of a national minimum wage in 2015, thus supplanting the model of labor-negotiated minimum wages that Cass claims to admire. Schröder was in many ways the German counterpart to Blair in Britain and Clinton in the U.S. — a center-left, “Third Way” liberal who pushed globalization — and thus an odd example for Cass to follow. However, whether Germany is actually the inspiration in this case is impossible to tell because, in a glaring omission, their experience with wage subsidies receives no discussion at all.

Instead, Cass argues that a U.S. wage subsidy would offset “subsidies given to foreign producers” and help communities “lacking the ability to export.” Granted, the German export manufacturing sector boomed after enacting wage restraint, but for reasons that America will never be able to replicate. In essence, Germany used wage subsidies and related labor market reforms to perform what economists call an internal devaluation, reducing their effective exchange rate with other Eurozone countries to artificially boost their current account surplus. This strategy won’t work for American exporters so long as the value of the U.S. dollar is free to appreciate in response to increased demand, even if it might help somewhat for dollar-pegged territories like Puerto Rico. Thus while a neutral industrial policy is surely preferable to discrete inducements for companies like Foxxconn to relocate stateside, say, Cass offers the wrong means to that end, distracted by superficially pro-work symbolism.

The same can be said about Cass’s disgust for payroll taxes. Employer-side payroll taxes superficially raise hiring costs, yet the low elasticity of labor means most of the burden really falls on workers. This makes payroll taxes more similar to broad based consumption taxes than costly labor regulation. As such, countries with stronger domestic manufacturing employment and more compressed wage distributions tend to rely heavily on payroll or value-added taxes. In Germany’s case, the payroll tax burden alone is roughly the size of payroll and income taxes in the US combined. Once again, Cass is tricked by semantic symbolism into supporting an even more progressive tax system, the kind typically found by necessity in countries without a strong middle class.

Internal Contradictions

Don’t get me wrong. Expanding and modernizing the Earned Income Tax Credit, the closest thing America has to a wage subsidy, is a great idea that may even pay for itself. Yet prioritizing something like it makes little sense from the internal point of view of Cass’s book, no less because, as a $72 billion a year cash transfer to low income households, it’s also the closest thing we have to a pure “maximize growth and redistribute the gains”-policy for the working poor.

In case after case, Cass prioritizes the pro-work semiotics of a policy over both consistency with his broader framework and the actual empirical literature. Wage subsidies beat a Universal Basic Income, for example, because it signifies a pro-work mood in contrast to the post-work vision of some UBI techno-utopians. Yet the empirical literature on income supports suggests they have a negligible effect on aggregate labor supply, and may actually prove useful in the context of globalization. Underfunded state unemployment insurance schemes are a major reason workers displaced by Chinese imports turned to disability insurance, for instance, instead of a robust (if temporary) basic income. Similarly, a child allowance — a kind of UBI for kids — would enable more single earner families on the margin, a shift Cass calls for in his vision of productive pluralism but fails to factor into his discussion of UBI’s potential cultural effects.

More generally, Cass’s rejection of economic efficiency as a guide, combined with a tendency to put terms like evidence-based policy in scare quotes, leads him to embrace disproportionate policy responses in the face of merely suggestive evidence. The mere possibility of immigration depressing native wages, for instance, justifies the steady deportation of over ten million people. Yet Cass feels no need to cite rigorous, causal research showing immigration harms the domestic working class (because there is none). Anecdotes and intuition suffice.

Rather than deport immigrants or subsidize low-skill natives, Cass’s framework implies a focus on labor demand. Yet the book contains no discussion of monetary policy, perhaps the single best tool for tightening demand for labor at the bottom end. The closest he comes is in a brief discussion of the economic recovery, before quickly dismissing the benefits of a hot labor market by noting productivity and wages remain low — a situation that would surely be worse if the automatic stabilizers baked into the existing safety-net were rationalized into a large wage subsidy.

A national development strategy that sought to raise the demand for lesser skilled workers would begin by moving US mid-sized manufactures up the global value chain. Think precision manufacturing instead of chicken pluckers. Germany, for instance, has successfully preserved 20th century levels of manufacturing employment in part by imposing aggressive renewable energy targets, shifting labor-intensive production that would have been lost to China into high-demand green jobs. With the cost of renewable energy and battery storage plummeting, Germany saw the future for work. Cass’s obsession with the costs of 1970s-era environmental regulation, in contrast, is like doing industrial policy through the rear view mirror.

Indeed, Cass is one of the nation’s leading opponents of a carbon tax, not because he denies climate change, but because he sees it as yet another case of pious elites pulling one over on workers. He thus fails to see how a revenue-neutral carbon tax would help to not only incentivize decarbonization for the climate’s sake, but support the manufacturing sectors of the 21st century. In particular, if a carbon tax were adjusted at the border, it would mean every paperclip imported from a coal-powered Chinese factory would cost a bit more, giving cleaner American manufacturers of every type a competitive edge. This is precisely the kind of rules-based export-promotion policy Cass wants out of wage subsidies, cancelling out the implicit subsidy foreign produces acquire through unpriced pollution. Nevertheless, Cass’s ideological blinders consistently prevent him from seeing environmental policy as anything but a drag on growth.

The Once and Future Republican

With the GOP increasingly a party of the white working class, a book like The Once and Future Worker is a necessary exercise. Fifty years ago, the ‘60s New Left movement fought with classical Marxists over whether to raise consciousness or reform the economic base — culture war versus class war. A similar dynamic has emerged in what remains of today’s Republican coalition, pitting the Alt-Right’s racist countercultural rebellion against reform conservatism’s traditionalist spin on economic populism. Cass, for his part, has written the most fully realized vision for a pro-worker Republican agenda to date, deftly synthesizing classical Marxist themes with conventional conservative rhetoric.

Unfortunately, while the book is fun to read and full of persuasive arguments in isolation, its broader policy vision comes up short. Like Marxism more generally, Cass falls victim to a metanarrative — an overarching story in which all of America’s deepest problems, from wage stagnation to deaths of despair, are downstream of the ruling class’s self-interested push for a more open, clean and efficient society. The naive view of globalization as an unalloyed good is in need of updating, to be sure. But liberalism is self-correcting precisely because it embraces empiricism over supposition, and modeled-based thinking over just-so stories.

The future of the conservative movement is not in rejecting the small-L liberal project, but in pragmatically adapting it to today’s challenges. That includes more robust trade adjustment and active-labor market policies, but also a recognition that the manufacturing economies of many small, depopulating towns are simply never going to come back. For them, redistribution through social security is far less of a threat than the Paul Ryans of the world who would reform it away.

Conversely, Cass’s core reform proposals range from politically impossible overhauls of labor and environmental law, to supply-side wage subsidies and regulatory reforms that are barely differentiable from the Zombie Reaganism he set out to transcend. If conservative populists are to win out over the culture warriors, they’ll have to do better.

As I tweeted a month earlier on February 2:

This is not an endorsement, but I think the "Trump just likes tariffs"-take is missing the forest for the trees.

I think there's a p(=30%) chance that we are instead witnessing the first (albeit somewhat clumsy) phase of a deliberate strategy for controlled de-dollarization.

People will dismiss this as 5D-chessing Trump's economic ignorance, but I've been in or around this world (libertarian-populist fusionism) for over a decade, know many of the relevant thinkers and theories, and have a reasonable track record at interpreting the context clues.

Naturally, many people read this and immediately concluded that I endorsed the tariffs and thought Trump was playing 5D chess. Sigh… As I made clear then and many times since, this was my attempt to issue a warning — a kind of public service announcement — that Liberation Day could have much farther-reaching implications than somewhat more expensive imports.

Share

We all know Trump loves tariffs and believes trade deficits are intrinsically bad. We also all know that the man’s grasp of public policy is vibes-y at best. This makes it hard to ascribe any strategic planning or intent to the tariffs, but a plan there was. As the Washington Post reported, plans for using tariffs as a lever to restructure the global trading system had been in discussion since the transition. Scott Bessent indicated as much on the various podcasts he appeared on in late November, right around when the first talks were taking place as part of his vetting:

I also felt very strongly that we're in the midst of a great realignment and of a Bretton Woods realignments coming in terms of global policy, global trade. There's a lot of what I taught at Yale and studied my whole life. I'd like to be part of it, either on the inside or the out.

My reviled February tweet essentially pointed at statements like this and asked “Guys, what if they’re serious?” For one, you might expect them to go unreasonably and indiscriminately big on tariffs to engineer something akin to the Nixon Shock; not simply impose reciprocal tariffs for the sake of immediately negotiating them away.

Of course, Trump only has vague concepts of his own administration’s plan, and is thus an X-factor in how this all plays out. But there really is / was a grander plan, or at least a Big Idea, and so far they seem to be sticking to it. Understanding this matters because “Trump just likes tariffs” and “the administration wants to restructure the global monetary system” in addition to Trump just liking tariffs have radically different implications. The former suggests the market should correct for the direct impact of tariffs on company profitability. The latter suggests that the floor could soon fall out of the stock market and U.S. assets more generally, as asset managers awaken to what’s entailed by a global financial reset. If Trump doesn’t reverse course soon, margin calls, capital flight, and cascading retaliatory measures could spur a dollar crisis and global Great Depression. While I don’t think this is the most likely outcome, these are the sorts of fat tailed risks we are now flirting with, and the tails keep getting fatter.

The fact that the stock market first rose on Trump saying there’d be a universal 10 percent tariff only to plunge immediately after the full tariff schedule was unveiled was a function of both the tariffs being larger than anyone expected, but also because of what it signaled about the administration’s underlying intent. The smart money failed to take the murmurings of a grander plan seriously, in part because they correctly assessed it would spell economic disaster.

The plan as best I understand it looks something like the following: First, Bessent hoped the tariffs and recession fears would lower yields, just in time for trillions of dollars in Treasuries to roll-over. Then, as part of extended, year(s)-long tariff negotiations, they intend to force foreign creditors to swap short-term U.S. bonds for long-duration bonds that lock in those lower rates. Why would foreign creditors ever agree to hold long-term, low yielding treasuries, you might ask? In theory, for tariff relief, but that bargain only makes sense if the tariffs in question start out as punitively large.

There were early warning signs that this was the direction the administration was heading. At his confirmation hearing in February, U.S. Trade Representative Jamieson Greer said Trump’s goal was “to restructure the global trading system to better serve American workers and businesses.” This comported with the views of Trump’s CEA Chair, Stephen Miran, who published a widely-discussed “User’s Guide to Restructuring the Global Trading System” in November of last year. The plan outlines options for how a graduated tariff policy could provide the U.S. with the negotiating leverage needed to induce countries to appreciate against the U.S. dollar, allowing America to retain its reserve currency status while locking-in a new regime of balanced trade and burden sharing. This would culminate in Bretton Woods-style conference called the Mara-Lago Accords. To his credit, Miran’s plan studiously explored strategies to avoid “adverse market reactions,” and in that sense the specifics of his plan were utterly ignored. However, it remains a relevant datum for understanding the administration’s fuller ambitions.

The notion that the dollar’s reserve currency status is less a blessing than a curse has been percolating in the New Right Zeitgeist for years. In 2023, I commended JD Vance for arguing that the dollar’s role in absorbing global savings represented a “resource curse” that contributed to deindustrialization. And back in 2019, Senators Josh Hawley and Tammy Baldwin proposed implementing a “market access charge” on foreign investment inflows to enable current account imbalances to be managed directly. The Hawley staffer behind this latter proposal is JD Vance’s current chief of staff.

I’ve long been sympathetic to these ideas, and even think a version of the “market access charge” is worth pursuing. At the same time, I’ve been warning that Trump’s personal obsession with trade deficits risks becoming a dangerous distraction. Persistent U.S. trade deficits are an epiphenomenal mirror to our persistent capital account surpluses, which are in turn downstream of the beggar-thy-neighbor industrial policies of countries like China and Germany. Thus, as I put it in a 2023 memo for American Compass, “Capital Flows Are the Core Concern”:

[T]he modern era of financial globalization is typified by enormous trade and financial imbalances—imbalances that will eventually come due. These imbalances are largely the responsibility of our foreign trading partners, particularly countries like China, which have engaged in one-sided industrial policies designed to suppress their domestic consumption. America’s role has simply been to absorb those imbalances with no questions asked. Until the global economy rebalances, efforts to rebuild American industry are fighting economic gravity. … While our trade deficit with China gets significant public attention, trade per se is largely a red herring. Investment flows matter much more.

Even if you believe the U.S. trade deficit is a problem, in other words, it’s unlikely to be resolved directly through the trade policy channel. Nevertheless, large universal tariffs can function similarly to an indirect “market access charge” — just at an enormous and unnecessary cost to U.S. consumers and producers. The only real advantage of imposing a market access charge through the trade channel is that the President has unusually unilateral authority over tariffs. Evidence that the administration is thinking along these lines can be found in the transcript of Trump’s Liberation Day speech, in which he declares that “Foreign nations will finally be asked to pay for the privilege of access to our market, the biggest market in the world.”

The administration’s approach thus represents an incoherent blend of Michael Pettis-thought, neo-protectionism, and supply-siderism. Judy Shelton, an old school supply-sider, even visited the White House in November to pitch Trump on issuing 50-year Treasury bonds backed by gold. More recently, Bessent went on Tucker Carlson’s show and revealed that he too is a “gold bug”:

Bessent: Gold cannot have a gigantic budget deficit. Gold cannot have a war, so just the fact that it is this isolated thing makes it very interesting. And the fact the entire global trading system until Richard Nixon took us off was tied to gold.

Carlson: So you're not anti-gold?

Bessent: Oh no, no, when I had my fund, I think people might've called me a gold bug.

The purpose and contours of the forthcoming Mara-Lago Accords are thus becoming clearer.

The supply-siders’ desire to move the U.S. dollar off a fiat standard onto a gold or commodity-backed standard is long-standing. For example, in 2020, I wrote a piece opposing Trump’s nomination of Judy Shelton to the Federal Reserve Board titled, “How Judy Shelton’s Call for a New Bretton Woods Duped Pro-Worker Conservatives.” Shelton’s nomination ultimately stalled out, but the duping has continued.

In the popular imagination, supply-siders are simply tax-cutting, Reaganite, “trickle down” conservatives. While not exactly wrong, this misses the extent to which the supply-siders represent a specific and quite weird heterodox school of economics. Consider that the OG supply-sider, Jude Wanniski, once wrote an ode to Karl Marx, declaring Das Kapital a “supply-sider text.” This is because Das Kapital focuses on the hard, materialist aspects of economic production to the neglect of consumption or other “demand side” factors. Marxist economists thus historically attributed recessions to a “crisis of over-production” rather than seeing excess capacity as simply the mirror to a negative demand shock — the core insight of John Maynard Keynes. These same Marxian blindspots animate China’s industrial strategy to this day, as seen in their relentless focus on expanding industrial production at the expense of domestic consumption.

In essence, the supply-siders believe the U.S. should adopt a market-driven version of the Chinese economic model. Slashing social programs, busting unions, and cutting taxes on capital gains are all in service of redistributing household consumption into corporate savings and investment. As a final step, converting the U.S. dollar into the basis for a hard-money currency union would allow America to be to the world as Germany is to the Euro Zone — the producerist core to a debt-saddled periphery.

The supply-siders are nonetheless free traders who abhor tariffs, and thus only begrudgingly accept Trump’s tariff agenda to the extent it leads to the establishment of a U.S.-centered, balanced free trade zone. Then there are Austrian economists like Josh Henderson who favor ditching the Treasury Standard for gold as a way to discipline budget deficits and curtail U.S. hegemony (a running theme of this administration, from the deletion of USAID on down).

That Trump has only the faintest understanding of these dynamics has turned him into a vessel for multiple vicarious and mutually-incompatible policy agendas. Oren Cass and the kind folks over at American Compass, for example, abhor the idea of union busting and would much rather see worker power rise — not be suppressed for the purposes of global competitiveness. Yet the disastrous roll-out of Trump’s tariff agenda has done little to advance American Compass’s brand of “pro-worker” conservatism. On the contrary, Cass’s efforts to defend the tariffs against the backdrop of $6 trillion in wanton wealth destruction has likely set his project back indefinitely. This is a shame, as Cass’s — and Vance’s — preferred policies are actually quite different from what Trump has pursued in practice. Cass supports the CHIPS Act, for instance, as well as spending large sums of money on programs for families and industrial development — a stark contrast to Trump’s slash-and-burn austerity.

Thus even if Trump makes an about-face on the tariffs tomorrow, irreversible damage has already been done, not least to his supporters’ reputations.

A note to subscribers: I launched this Substack for longer essays with no other home, but there’s so much crazy stuff happening in the world and so quickly that I’ve decided to start writing shorter, more frequent posts reacting to things as they happen. Whether it’s AI or U.S. politics, the context now shifts radically from week to week. My normal methodology of thinking for awhile and then zero-shotting a 10,000 word treatise thus has an impedance mismatch with reality. At the same time, I’ve always felt awkward using Substack as a short-form blog, so going forward I’m going to experiment with different formats and likely bundle shorter takes into longer newsletters. I welcome any feedback along the way.

As a young techno-libertarian, I had grown suspicious of the “starve the beast” strategy favored by my anti-statist friends and colleagues. It seemed to me that, at least in a democracy, the size and scope of government was endogenous to popular political demands (see: Wagner’s Law). Cutting taxes and sabotaging government services can thus paradoxically lead to an even bigger government, both because deficit financing makes government programs feel artificially “cheap” to the public (see: fiscal illusion), and because governments end up having to redouble their spending and regulation to offset bureaucratic inefficiencies (see: Kludgeocracy in America).

I concluded that, if we were ever going to transition to a more libertarian form of government, it would be essential to first use information technology to improve and streamline core government services in a way that creates space for crowding-in privatized forms of governance. Just as Uber and Lyft disrupted regulated taxi commissions, for example, perhaps many other governmental functions could one day be provided by the private governance of competing software platforms using AI-enabled reputation mechanisms, dispute resolution systems and the like — an “Uber for governance” that solves for the transaction costs and market failures that are today internalized by the Weberian nation-state.

In short, just as Marc Andreessen said “software is eating the world,” perhaps software will eventually eat the state. And nearly a decade later, it seems like this strategy for disrupting bureaucracy is finally coming to fruition in the form of Elon Musk’s Department of Government Efficiency.

Share

DOGE theory and practice

DOGE was initially billed as a way to drive cost savings in government, but the fiscal framing is potentially misleading. While there are many sources of waste, fraud and abuse that DOGE will likely help address, the federal government’s discretionary budget is simply too meagre relative to defense and entitlement spending to make much of a dent (not that one shouldn’t expect significant reform in those areas as well). Instead, I think DOGE is better understood as a combined deregulatory and whole-of-government reform effort, with fiscal savings as just the icing on the cake.

So far, it’s been announced that the core DOGE team will be stationed within the U.S. Digital Service — now renamed the U.S. DOGE Service — with “DOGE teams” of at least four embedded in key government agencies. According the Executive Order establishing DOGE,

Each DOGE Team will typically include one DOGE Team Lead, one engineer, one human resources specialist, and one attorney. Agency Heads shall ensure that DOGE Team Leads coordinate their work with USDS and advise their respective Agency Heads on implementing the President’s DOGE Agenda.

The EO then directs the USDS Administrator to commence a Software Modernization Initiative “to improve the quality and efficiency of government-wide software, network infrastructure, and information technology systems,” as well as “promote inter-operability between agency networks and systems, ensure data integrity, and facilitate responsible data collection and synchronization.” This is achieved by requiring Agency Heads “to ensure USDS has full and prompt access to all unclassified agency records, software systems, and IT systems.”

The revelation that DOGE will be focused on modernizing federal IT infrastructure caused some on the left to breathe a sigh of relief, though I think they will be in for a surprise. Consider that F.D.R. first seized the reigns of the federal bureaucracy by moving the Bureau of the Budget — the precursor to today’s Office of Management and Budget or OMB — into the executive office. Under the stewardship of Harold D. Smith, the Budget Bureau centralized agencies’ legislative requests to Congress while embedding beachheads in each agency tasked with aligning the federal bureaucracy to F.D.R.’s New Deal agenda. This included moving disfavored agencies out of D.C., developing accounting and auditing system to automate clerical work, overhauling duplicative statistical systems, and forcing through agency-wide reorganizations. As one Congressman said at the time, “We grant the powers and Harold Smith writes the laws.”

The DOGE takes a page out of F.D.R.’s playbook, now oriented around CTOs and CIOs rather than budget officers, though Russ Vought’s OMB will continue to play a central role. The core work of DOGE will be less about making minor improvements to existing IT systems than in overhauling the federal government’s tech-stack from top to bottom. If successful, DOGE will bring unprecedented transparency to federal spending while laying the infrastructure needed for a significant downsizing in the federal workforce through automations.

For example, the DOGE is reportedly planning to move the systems used to track federal spending onto a permissionless blockchain. Curious where your infrastructure dollars are being spent? Today, accessing that information may require combing through obscure PDFs and press releases or submitting a FOIA request. In the near future, it could be as simple as making an API call to fetch the Department of Transportation’s on-chain transactions.

While this might sound like a Web-3 boondoggle in the making (and it may still be), there are precedents that make integrating blockchain into the government well worth considering. Take Estonia, which has the most most sophisticated e-government in the world. Given the benefit of a blank slate following the collapse of the USSR, Estonia’s e-government was the product of young policy hackers who entered the civil service when the internet was just starting to take off. With the foresight to see where technology was heading, these reformers pushed Estonia to become digitally-native through a series of sweeping modernization initiatives.

This included adopting a novel technology known as X-Road; a distributed, public key encrypted data exchange layer that resembles a permissioned blockchain. X-Road works like an intranet that allows decentralized servers and public key holders to exchange data quickly and securely according to a set of interoperability standards. It also enables substantial automation by allowing government APIs to integrate with things like e-banking, business registries, electronic health records, child support systems and more. Today, 50 government services, including voting in Estonian elections, can be done via a mobile app that’s accessible 24 hours a day, 7 days a week.

The U.S. federal government has long way to go to get to Estonia, but you have to start somewhere. And for better or worse, we don’t have much of a choice. With AGI on the horizon, the U.S. government is simply not prepared for the cybersecurity and throughput issues entailed by a world flooded with billions of superintelligent agents.

Whether DOGE has all the resources and authorities it needs remains to be seen, but given the enormous time-pressure of the current moment, I for one will be rooting for its unmitigated success.

A common criticism of libertarian philosophy is that it can’t handle collective action problems — that a totally voluntary society lacks the tools to build lighthouses, prevent overfishing, or ensure we all get our vaccines.

In response, libertarian thinkers developed a branch of economics dedicated to showing how collective action problems can be solved with voluntary cooperative arrangements. Elinor Ostrom’s work was particularly important for arguing that, under the right conditions, norms and civil society can evolve to govern the commons from the bottom up.

There are obviously limits to informal norms, however. For one, they are easy to undermine through appeals to a narrow, self-interested conception of rationality. After all, norms exist to enforce cooperative arrangements that would otherwise be unstable. That suggests it is always possible for a sophist to jeopardize collective action by appealing to their peers’ individually rational but myopic motivations: “Just catch one more fish, no one will notice.” With each person who defects it then becomes increasingly tempting for others to follow suit.

What about voting?

Voting represents an interesting test case for the robustness of voluntary solutions to collective action problems, since any single individual’s vote is mathematically insignificant. According to one often-cited estimate, the likelihood of casting the decisive vote in a U.S. presidential election is 1 in 60 million. And yet when voters act collectively, thousands of individually meaningless votes can quickly add up and become a force to be reckoned with. 

Nonetheless, many of the same libertarians who insist that norms and civil society can solve large scale collective action problems also insist that voting is individually “irrational,” and therefore abstain. “Voting is overrated,” argues Katherine Mangu-Ward, editor in chief of Reason magazine, in a video posted ahead of the 2020 election. “The reasons people give for why they vote—and why everyone else should too—are flawed, unconvincing, and occasionally dangerous.”

“Your vote is wildly unlikely to determine the election. It’s pure math,” Mangu-Ward continues, citing the 1 in 60 million figure mentioned above. Worse still, boosting the turnout of people who are “young, uneducated or otherwise less likely to be engaged” can have the unintended consequence of diluting the vote of those who are better informed: “Get out the vote campaigns promote precisely the kind of morally condemnable, ignorant voting we should be discouraging.”

That libertarian bulwarks like Reason Magazine feel compelled to rehash sophomoric arguments against voting every other election cycle merely reaffirms the worry that libertarianism contains the seeds of its own unravelling. The emphasis on instrumental, means-end rationality, in particular, ignores what to most people are their primary, normative motivations for action. Voting has the structure of a civic duty because of its limited instrumental value.

The notion that only informed, “high information” voters should participate represents a similar instrumentalization of democracy, as if elections were merely about aggregating-up individual beliefs and preferences (as Condorcet showed, they’re not). In truth, the most “informed” voters also tend to be the most politically and ideologically polarized — the sort of people who watch Fox News or MSNBC all day — and thus not a sound foundation for epistocracy, to put it mildly. 

Of course, that we vote in large numbers at all is in some sense a vindication of Ostrom and her school of economics. Rather than act as atomized utility functions, we cement the norm of voting with the help of overlapping institutions like political parties, religious congregations, unions, non-profits, membership clubs, and not to mention friends and family. We communicate voting intentions to other individuals within these groups, which are small enough to reinforce a mutual expectation of follow-through. Groups in turn coordinate with other groups, like when a local union or social club coordinates with its other chapters. Pretty quickly a meagre individual vote becomes amplified into the hugely consequential endorsement of a union federation or influential political action committee.

More rational than thou

I therefore don’t believe libertarians are totally sincere when they make the “voting is irrational” argument. Or, more to the point, I suspect it is a case of motivated reasoning. As we’ve already seen, it is cognitively dissonant with their optimism about voluntary collective action in other spheres (“collective action for me but not for thee”). But moreover, it seems to spring from their mood-aversion to electoral politics more generally. In the words of Mangu-Ward, “Washing one’s hands of the whole system is a good way to ensure that they remain clean, even when the politicos are dirty.” Her tendentious arguments against voting are thus an example of what criminologists refer to as “techniques of neutralization” — self-serving excuses that proactively rationalize defection from social norms that one finds inconvenient. Classic examples include shirking at work because “everyone else is doing it,” or telling yourself that shoplifting if OK because big retailers have already baked petty theft into the price.

Libertarians double down on their mood-aversion when they argue that voting is inherently immoral or distasteful because it involves participating in a coercive enterprise. And true to form, Mangu-Ward argues that we arguably have a moral duty not to vote, comparing voting to participation in a firing squad. Yet besides the obvious tension with her “voting is ineffectual” view, there is no pressing need for a norm against voting, just as there is no need for a norm for littering, overfishing or free-riding off of herd immunity. Those behaviors all naturally fall out of individually self-interested human action; they are what is left in the absence of social coordination through norms and other communicative modes of… Reason.

Motivated reasoning is just the charitable interpretation. The less charitable one is that the average libertarian is tragically bereft of the social capital (clubs, networks, and civil society) needed to leverage their idiosyncratic beliefs and motivations into collective action. Just tune into the Libertarian Party convention if you doubt this. If you thought herding cats was hard, try herding philosophical anarchists.

The even less generous view is that libertarianism represents a self-defeating memeplex — a mind virus that handicaps its host so badly that it ceases to reproduce. Ron Paul himself could be on the ballot, only to lose because a non-trivial percentage of his supporters rationally chose to stay home. By analogy, if you wanted to hobble the labor movement, one strategy would be to plant agent provocateurs within a union’s ranks to charismatically defend the instrumental rationality of being a scab. Or better yet, one could deploy dorm room thought experiments to convince their pseudo-comrades that being a scab is not just rational, but just and noble. Were that conviction to ever catch on within the labor movement, it would cease to exist, a victim of ideological natural selection.

And as a matter of fact, that is more or less what happened in the 1960s. It was called the New Left, and its aversion to normative authority and social conformism hobbled progressives’ ability to influence institutional reform for a generation. Now that the right is having its own countercultural moment, with no shortage of libertarian fellow travelers, the attempt to advance social change through culture-jamming and norm subversion risks being equally in vain. And indeed, Trump lost the 2020 election in large part because he chose norm subversion over the hard work of motivating early voters. If the 2024 election is any different, it will be because the right has learned its lesson and transcended being a mere counterculture into being a bona fide counter-establishment with the political machinery necessary to drive meaningful collective action.

At the same time, it will always be easier to tear down norms than to build them back up. So whatever your political persuasion, be sure to not only vote, but heap shame and stigma on those who don’t. Not because voting is individually rational, but precisely because it is not.

I reject neoliberalism, but not because I’m a Post Liberal or anti-capitalist. On the contrary, my problems with the neoliberal era stem from my even deeper commitment to political liberalism in the truer philosophical sense.

Start with Hayek’s classic distinction between the Anglo-Scottish strains of liberalism and the rationalist “constructivist” liberalism of the French Enlightenment. The former sees liberal institutions as emerging from a spontaneous, bottom-up process of social evolution à la common law, while the latter seeks to design a free society from top-down à la Napoleonic Codes.

Wanting to redesign society along rational principles is typically associated with the left and modern liberalism, though it needn’t be. Libertarianism is a rationalist mutation of classical liberalism, while John Dewey’s modern liberalism endorsed an evolutionary account of social progress.

Whether classical or modern, I’ve always been more partial to the evolutionary account of liberalism for a few reasons:

• First, as a kind of Hegelian, I’m skeptical of attempts to impose a rational design on society detached from the concrete “actuality” of existing social practices. Classical liberals understand this point well in the context of left-wing social engineering schemes, but it applies equally to top-down attempts at liberalization, from the Shock Therapy imposed on Russia in the early 1990s, to the neoconservative project of democratizing Iraq and Afghanistan.

• Second, the evolutionary account comes closer to reconstructing how liberal norms and institutions actually emerged in the first place; namely, as a way to transcend sectarian conflict and restore order in the aftermath of the European Wars of Religion. As evinced by the proto-liberalism of Thomas Hobbes or the natural law theory of Hugo Grotius, liberalism was in some sense discovered through the logic of positive-sum games, not invented by Enlightenment philosophers. They merely gave that logic a vocabulary.

• Third, in lieu of consensus on values and morality, bottom-up liberalism provides a potential framework for mediating between different conceptions of the good. Critics of liberalism sometimes lament “liberal neutrality” for being empty of moral content, but that is by design. As Hume put it, justice is an “artifice” for securing social cooperation, not a natural virtue in its own right, as what ought cannot be derived from what is.

Putting these together, my approach to liberalism is in some sense deflationary. I don’t think it makes sense to self-identify as a “liberal,” for instance; or at least, it cannot be the whole of one’s identity. The Acts of Toleration that restored comity between Protestants and Catholics depended on the existence of Protestants and Catholics with different, incompatible conceptions of the good. An 18th century Catholic could be liberal in the adjectival sense of being committed to mutual respect and toleration, but their theological identity still came prior. The ideal “liberal” in this sense is a kind of impartial arbiter or broker — an institutional role one can adopt, but not embody as a conception of the good in itself. The critics of the hollowness of “liberal neutrality” are thus on some level right about the need for pre-liberal value systems and “forms of living” to breathe life into liberal institutions.

At the same time, liberalism isn’t totally value neutral. Its quietism on issues of worldview, religion and culture is in service of a positive normative commitment to social cooperation. As the Joker put it, “We live in a society.” We can either learn to tolerate each other in pursuit of shared goals or embrace the zero-sum logic of conflict theory and mutual domination.

Bottom-up social contract theory

This more deflationary notion of liberalism was recapitulated by John Rawls in Political Liberalism, the book of his I most greatly admire. Rawls sets out to derive a “freestanding” basis for the legitimacy and stability of liberal institutions that, by definition, must avoid imposing a “comprehensive doctrine” on its citizens, including the more “comprehensive” forms of liberalism associated with Kant and Mill. This is what gives rise to Rawls’s notion of the “overlapping consensus,” as the laws that different peoples find mutually agreeable will automatically inherit a degree of legitimacy. The proviso is that the citizenry must practice “reasonable pluralism,” such that “perfectionist” worldviews that call for the imposition of contested values on others needn’t be accommodated.

As a neo-Kantian, Rawls’s general approach to political philosophy was to construct these kinds of abstract, “ideal” principles of justice in which everyone is well-behaved and only later work out how our messy, “non-ideal” reality can be brought into conformity. As a Hegelian, I flip this methodology on its head by beginning with the world as it is, and only later try to “rationally reconstruct” the norms implicit in antecedent practices. While the two approaches can converge, the latter is less likely to fall prey to empty abstractions and provides a more compelling account of how liberal principles inherit their normative force.

To illustrate the difference, consider that Rawls’s most important philosophical contribution was arguably the revival of social contract theory as a serious alternative to both Marxism and utilitarianism. As he puts it in A Theory of Justice, “the basic structure” of a liberal society should be conceived as “a cooperative venture for mutual advantage,” where “the basic structure” refers to the foundational legal and social institutions we share in common but did not explicitly opt-in to. This restricts the scope of liberal principles to the members of the same basic structure (i.e. the nation-state), and distributive justice to the “cooperative surplus” enabled by those institutions.

I also think of myself as a kind of contractarian, just of a decidedly non-ideal sort. Rather than start with abstract scenarios like Rawls’ “original position” or Hobbes “state of nature,” I see the constitution of actually-existing liberal societies as embodying political settlements between their internal factions. This historical, “bottom-up” approach to the social contract has a few implications:

• Rather than positing what neo-Hobbesian theorist David Gauthier calls an "initial bargaining position" between hypothetical members of a society, bargaining between individuals, groups and institutions is a continual, on-going process. It’s the difference between the top-down social planner of Paul Samuelson and the bottom-up institutionalism of Ronald Coase or Elinor Ostrom.

• In lieu of a sharp historical break, the roots of Western liberalism can be traced to the “constitutional bargaining” of the late medieval period, suggesting a deeper continuity between feudal Europe and modern liberalism than Enlightenment thinkers cared to admit. Joseph de Maistre pointed this out in his treatise on The Generative Principle of Political Constitutions, which argued that written constitutions conceal the extent to which all constitutions are still overwhelmingly unwritten.

• While centralized governments were key to getting liberalism-proper off the ground, de facto sovereignty in liberal societies remains highly polycentric. The rise of liberalism and nationalism went hand-in-hand as a superstructure for channeling the polycentric sovereignty of pre-modern society toward the “commonweal,” eg. from corporations chartered by the King to acts of general incorporation within a Smithian framework of socially productive competition. “Ordered liberty” thus consists in the building and maintenance of “well-constructed institutions” that protect this equilibrium from decaying back into the medieval rule of the clan.

• Liberal principles of justice only exist to the extent they’re institutionally embedded. In contrast, utilitarian, universalist and cosmopolitan varieties of liberalism are products of what Hegel famously called “Empty Formalism”: abstractions detached from the concrete networks of mutual exchange and recognition that give our ethical commitments their normative force. In particular, the nation-state synthesizes our ethical identity (the nation) with an interlocking set of civic institutions (the state), allowing us to realize our freedom through the other.

• Merely positing a liberal theory of justice is insufficient. Actionable theories of justice must be able to reconstruct the development of existing institutions, which means putting political economy center stage. Think Mancur Olson’s notion of an “encompassing coalition” or John Kenneth Galbraith’s “Countervailing Power.” Whether a society’s institutions are inclusive or extractive depends on a variety of path-dependent historical, economic and technological factors, the structure of political competition, the incentives facing political elites, and the relative bargaining power of minority factions, among many other things.

• Appreciating how transaction costs structure the boundary of firms and governments alike reveals parallels between corporate and political forms of governance. Medieval fiefdoms and city states are analogous to hierarchical corporations with residual claimants, for example, while modern social democracies are structured more like “mutual insurance co-operatives” that divide ownership equally among the citizen-members.

• Securing cooperation across society’s members requires the “losers” from major trade, technology and policy shocks to be made better off. Hicks-Kaldor efficiency isn't good enough.

Per Rawls’ notion of the “cooperative surplus,” the 19th and 20th century development of liberal welfare states can even be reconstructed as reflecting the positive-sum logic of compensatory social insurance, in some sense completing the market in light of endemic market failures, rather than as a purely zero-sum form of egalitarianism. Consider Denmark’s “flexicurity” model of labor relations. Flexicurity approximates the modern liberal ideal by combining highly flexible labor market regulations with robust income security for displaced workers. Rather than being the result of a technocratic policy choice, the Danish model emerged as a political settlement between sectoral labor unions and employer associations across a century of dialogue and negotiation, and will no doubt continue to evolve over time. As I put it in my paper on The Free-market Welfare State, “Like other varieties of liberalism, Denmark’s success is liminal, existing on the boundary of conflicting worldviews.”

Pluralism vs. the separateness of persons

The Danish example also illustrates the extent to which actually-existing social contracts require intermediate groups to aggregate individual interests into collective action, putting limits on what Rawls called “the separateness of persons.” As Charles Taylor argues in Sources of The Self, the notion that an individual’s identity is neatly separable from the culture and community in which they are embedded is historically modern. In reality, our sense of self and identity is constituted inter-subjectively by our peers and surrounding culture, and requires the existence of different peoples to ground our own identity through mutual recognition. Even identifying as an authentic individual implicitly depends on the recognition of your individuality in others. The modern, individualistic conception of human dignity and autonomy is thus itself a kind of cultural inheritance, as reflected in the Christian genealogy of Kant, Descartes and Locke.

Taylor’s argument for the “diversity of goods” leads to a notion of pluralism as distinct from the more rationalistic notions of individual rights and freedoms. He was writing in the context of the Canadian debate over multiculturalism, particularly Anglophone-Francophone relations. The fact that Canada has Quebec, a distinct “French Nation” within an otherwise English speaking country, generates ethnolinguistic conflicts over power sharing and notions of the good. Out of this conflict, which at its peak included a credible separatist movement, emerged a number of political settlements rooted in a particularistic application of liberal neutrality and accommodation (eg. official bilingualism and robust confederalism).

Liberal neutrality and pluralism are thus less about maximizing an abstract notion of individual autonomy than about maintaining harmonious cooperation between different social classes and sub-groups vis-à-vis their shared institutions. This differs markedly from the perfectionist liberalism of, say, John Stuart Mill, who argued that our formative peer groups and cultures can themselves be tyrannical. As Mill writes in On Liberty,

Society can and does execute its own mandates: and if it issues wrong mandates instead of right, or any mandates at all in things with which it ought not to meddle, it practises a social tyranny more formidable than many kinds of political oppression, since, though not usually upheld by such extreme penalties, it leaves fewer means of escape, penetrating much more deeply into the details of life, and enslaving the soul itself.

The tension between the rationalist and pluralist conception of liberalism was illustrated in France’s debate over whether to “ban the burqa.” Rationalist liberal feminists argue that the burqa and niqab represent patriarchal oppression, such that a woman raised in a traditional Muslim culture is less free even if they believe they’re making a free choice. It’s likewise no accident that Mill was both an early feminist and a proponent of British colonialism’s “civilizing” effect on India. A pluralist, in contrast, would argue that — whatever one’s personal views — the basic structures of a liberal society should seek to accommodate traditional cultures and focus more on civic integration, which coercive forms of cultural assimilation may even undermine.

Liberal Protestantism as America’s civil religion

I fall squarely in the pluralist camp, and see the “perfectionist liberalism” of Mill as oxymoronic. Mill was the son of an ordained Presbyterian minister, and while his father left the flock for “Natural Religion,” Mill professed reverence for Jesus Christ and frequented worship services throughout his life. Mill’s liberalism was thus itself a comprehensive doctrine reflecting a secular-rationalist form of Protestantism.

Explaining these distinctions is tricky in the U.S. context given the strong Protestant influence on American culture. The semantic conflation between “political liberalism” and “social liberalism” is particularly hard to avoid, but there is a distinction with a difference. The “Great Awokening,” for example, had clear echoes with the Protestant Great Awakenings that have recurred throughout American history. Progressivism’s religious underpinnings are simply masked by the contemporary decline in explicit religious affiliation, particularly within the Mainline church, and yet the sociological base remains nearly identical. Jordan Peterson’s attribution of woke politics to “Postmodern Marxist” college professors is thus mistaken. Wokism is as American as apple-pie, reflecting a Protestant form of Christian Nationalism that's secular but no less sectarian.

It should therefore be no surprise that most of the leading “Post Liberals” in American politics are Catholic. Catholic social teaching and Thomism provide alternative accounts of the good; ones that — per Rawls — a liberal society ought to accommodate. Instead, elite American Protestants have used their cultural and institutional dominance to embed their comprehensive doctrine into everything from the federal rule-making process to professional medical guidelines. The inevitable backlash to this overreach is then presented as “illiberal” via the semantic conflation between the political and social forms of liberalism mentioned above, when if anything the Post Liberals are directionally pluralist.

Similarly, the New Right critique of “woke capital” is not merely a critique of ESG or investment screening regulations, but an identification of wokeness with the telos of managerial capitalism per se. That is, rather than being an aberration, multinational corporations seem to strongly post-select for woke HR departments and DEI trainings as a way to enforce workplace harmony, as though blending Quaker theology with the Singapore model of multiculturalism.

Given both its universalist moral orientation and hostility to institutional mediation, America’s low church Protestant culture was thus uniquely well adapted to the era of “disembedded liberalism.” Conversely, the Post Liberal interest in labor unions and “integralism” seems to be groping at the sort of embedded, membership-driven social institutions that once helped order people’s lives and align diffuse interests toward the common good.

Some argue that Protestantism's cultural triumph over counter-revolutionary Catholicism has nevertheless benefited the world, pointing to the evidence that Protestant countries are more innovative, ambitious, and industrious. While true, this judgment appeals to abstract moral theories like utilitarianism and consequentialism, which by nature transcend the specific duties created by our communities, institutions and social compacts. It's no surprise that these philosophical approaches were developed during the height of the British Empire and regained popularity in the era of disembedded liberalism. Utilitarianism, viewing morality from the “God's eye” perspective of a global social planner, dismisses the need for concrete social foundations. To my deflationary liberal mind, that makes it just another comprehensive doctrine among many.

Remember, political liberalism emerged out of religious conflict, and while we’ve secularized and changed some labels around, those conflicts are largely still with us today. Forgetting that fact and allowing one conception of the good to become hegemonic is a recipe for sectarian conflict. It follows that our elite institutions either need to deflate their conception of liberalism or risk rediscovering the logic of social contract theory the hard way.

To start with, plain old liberalism is a tradition in political philosophy based on tolerance, individual freedom and the rule of law. Classical liberalism is relatively laissez-faire and focused on “freedom from” the government, while modern liberalism has a “positive” conception of freedom that includes distributive justice.

Neoliberalism, on the other hand, is as much a sociological category as a philosophical one. In the U.S. context, the term is associated with Charles Peters’ 1983 essay, A Neoliberal’s Manifesto, which presents neoliberalism as the center-left mirror to neoconservatism; a corrective for liberals mugged by the realities of stagflation and excess unionism who nevertheless wished to retain their liberal ideals. Where a neoconservative embraced Reaganite tax cuts to spur growth, for example, a “neoliberal” like Gary Hart proposed issuing a special class of stock “for the explicit purpose of investment in new plants and equipment.” Rather than reject markets, Peters’ neoliberalism was emphatically pro-business, globalization and entrepreneurship, just within the context of technocratic policies for aligning markets to liberal ends. This went on to become the dominant force within the Democratic Party in the 1990s and beyond, from Bill Clinton’s “Empowerment Zones” to Cory Booker’s “Opportunity Zones, and as parodied by the Oddly Specific Kamala Harris Policy Generator.

The original meaning of neoliberalism is related to Peters’ but goes back to Walter Lippmann’s 1937 book, The Good Society. Lippmann was a modern liberal who was similarly “mugged” by the failures of New Deal-era central planning, leaving him indebted to the arguments against socialism popularized by Ludwig von Mises and F.A. Hayek. Yet Lippmann was far from a libertarian. In parallel with the Ordoliberal thinkers of post-war Germany, Lippmann believed in the need for a “new liberalism” that struck a balance between the excesses of both governments and markets.

The Walter Lippmann Colloquium was held in Paris in 1938 to work out what that could mean in practice, with Mises and Hayek in attendance alongside a slate of French and German economists. The term “neoliberalism” was coined at the meeting by German sociologist Alexander Rüstow to disambiguate their project from the laissez-faire liberalism of the 1800s. The Colloque Lippmann was a precursor to the Mont Pelerin Society meetings that kicked off a decade later, providing a transnational forum for neoliberal thinkers, historians and business leaders to discuss strategies for preserving open markets and societies amid the reconstruction of Europe and rise of international communism.

Neoliberalism as Third Way globalism

Most popular histories of neoliberalism draw a straight line from the Mont Pelerin Society to Reagan and Thatcher. But as seen in its origins, neoliberalism has always had a left and right flank, and if anything began as a form of Third Way-ism most closely associated with center-left internationalists. Neoliberalism’s association with the small government right merely conflates the 1980s Anglo-American experience with a much broader phenomenon. Thus right-neoliberals have Mises and Hayek while left-neoliberals have Popper and Polanyi, explaining why Charles Koch and George Soros have such strikingly parallel interest in dead Viennese philosophers.

Far from being neoliberalism’s opposite, market-oriented social democracy is simply how neoliberalism manifested on the continent. The Ordoliberal influence on Germany’s model of a “social market economy” is a case in point. Progressive critiques of shareholder primacy and the like are thus often best understood in terms of an internecine dispute between the left and right flanks of a common neoliberal ideology, not as a rejection of neoliberalism per se. Klaus Schwab’s “stakeholder capitalism,” Elizabeth Warren's support for codetermination, and the EU’s tough antitrust stance all channel the early Ordoliberals, for example.

What unites left and right neoliberals is less a blind faith in “free market fundamentalism” than support for multilateral economic integration managed by transnational institutions and a professional class of technocrats. Think the World Economic Forum, IMF, WTO and OECD; Hillary Clinton’s backroom call for a “hemispheric global common market with open trade and open borders,” or Judy Shelton’s vision for a North American currency called the Amero.

Neoliberal institutions and ideas embody a project of global integration and governance that’s distinctly inimical to popular sovereignty. This contrasts with classical liberalism, which arose in parallel with the consolidation of the first modern nation-states and thus rooted political authority in the consent of a national political community. Nevertheless, in the aftermath of World War II, the fear that nationalism and popular democracy would invariably slide back into reaction or revolution was fairly understandable. With Europe in the process of rebuilding, an opportunity for greater transnational integration presented itself, supported by a series of domestic social compromises aimed at keeping liberal reformers in charge and the fringes at bay.

The Golden Straitjacket

Nancy MacLean’s description of the neoliberal project as putting “Democracy in Chains” is thus on some basic descriptive level right. But given the internecine nature of these debates, her narrative focus on anti-majoritarian American thinkers like James M. Buchanan and John Calhoun ranges from one-sided to conspiratorial. A less parochial history of neoliberalism’s suspicion of democracy would start, not with American Southerners or 1970s libertarian academics, but with the framers of the European common market. As the socialist historian Jonas Elvander explains in his book, Disciplined Democracy: The neoliberal history of the EU,

The story of the EU’s neoliberalism is usually told as a “turn” that happened in the mid-1980s and early 90s, culminating in the Maastricht Treaty. This is only partly true. The foundations for the EU were laid at the same time as neoliberalism emerged, in the 1940s and 50s, and some key elements in the Rome Treaty were formulated by “ordoliberals” (the German variant of neoliberalism which put more emphasis on antitrust policies), such as the “four freedoms” of the Common Market. The reason the EC is never described as neoliberal in this period is that the free movement of people, capital, goods, and services was not fully realised until the 80s and 90s, when the wider neoliberal shift in the world made it politically possible. What Jacques Delors and his colleagues did in the 80s was actually little more than to implement the Rome Treaty, 30 years after its adoption, which some neoliberals also acknowledged. The neoliberal principles that had been there since the 50s thus became “operational” only in the 80s.

The Rome Treaty and European Economic Community formed out of the Treaty of Paris, which established shared steel and coal production between France and Germany to bind them against returning to war. Alongside the Europe Declaration, the Paris Treaty laid the foundations for supranational institution-building as the first formal agreement to emphasize a common European identity. Economic integration and harmonization eventually gave way to political integration with the creation of the European Union, binding the sovereignty of member countries through a version of Thomas Friedman’s “golden straitjacket.”

For Britain and the United States, supranational institution building took the form of the Bretton Woods system. As the new de facto global hegemon, the U.S. pushed for the creation of a new international monetary regime to consolidate financial power and facilitate Europe’s reconstruction through open markets. Under Bretton Woods, countries maintained a currency peg to the gold-backed U.S. dollar to prevent competitive devaluations, while supranational institutions like the IMF and World Bank managed global payment imbalances to enable member nations to rebuild their economies through Keynesian social and industrial policies. Political scientists thus refer to the Bretton Woods-era as representing a form of “embedded liberalism” that enabled cross-border trade in goods and services while preserving nations’ sovereignty over the capital structure of their internal economies.

Managing the balance of payments between multiple countries was a delicate task that, like the three body problem in physics, ultimately proved unstable. Currency pegs began unravelling in the late-1960s, and by 1976 virtually every member country had converted to a free-floating exchange rate. Per the “impossible trilemma” in international economics, the end of Bretton Woods restored members’ sovereignty over monetary policy in exchange for the free flow of capital across borders. Soon after, the IMF pivoted away from supporting capital controls to being a strong proponent of international capital mobility and market-based structural reforms. The era of disembedded liberalism was thus born, typified by financial globalization, the rise of multinational corporations, and soaring levels of foreign direct investment.

Disembedded Liberalism at home

In the U.S., the transition to disembedded liberalism manifested domestically through the decline of large, membership-driven stakeholder organizations, from unions and churches to thick political parties. Such associations had served to aggregate diffuse interests into coherent negotiation blocks during the New Deal-era, giving rise to the cross-cutting and regionally-embedded political coalitions. Ironically, however, the New Deal also greatly expanded the administrative state’s role in the national economy by drawing on a broad interpretation of the Commerce Clause. This paralleled the European project of economic and political integration by helping make the U.S. even more of a continent-wide common market. So as with the rise of Brussels, the post-New Deal era cemented the need for a distinct technocratic class in Washington to run the federal apparatus.

Most parliamentary democracies went through a similar transition and located their policy elites within professionalized civil services or political party organizations. The separations of power in the U.S. presidential system made sustaining a similar degree of technocratic “in-sourcing” considerably more difficult. U.S. state capacity thus reached its high watermark in the 1950s, when trust in government was high and the civil service still contained many talented and ambitious New Dealers. Overtime, however, the technocratic class came to be predominantly supplied by think tanks, law firms, universities and large foundations — what Theda Skocpol has called “associations without members.” As the political scientist Steven Teles explains,

Before the 1960s, if you were an idealistic person who wanted to drive social change in some way, you really didn’t have much choice to either build a mass membership organization or work within one. But things really did change in the 1960s. Foundations like Ford and Rockefeller began to fund a huge network of organizations in law, civil rights, feminism, consumerism and the environment, a process I examined as the backdrop to my book “The Rise of the Conservative Legal Movement.” Suddenly, if you were that idealistic person, you could skip over the step of building a mass movement and put out your shingle and start suing, lobbying and publishing. That model — a professionally staffed organization based in D.C., no members, funded primarily by foundations — came to be the dominant one on the center-left. 

This included the legal aid and consumer rights movements, which pulled functions that might have otherwise been supplied by an administrative agency into the courts, empowering an autonomous class of legal elites. The “rights revolution” fit neatly into a neoliberal framework given its emphasis on procedural justice, paralleling the Ordoliberals’ juridical obsession with rules over discretion. Ralph Nader was a friend and admirer of the economist Alfred E. Kahn, for instance, seeing the price deregulations he advanced while in the Carter administration as a great win for consumers. As I discuss in Non-profits Are Under-theorized,

In this light, the conflation of “the neoliberal turn” with Reaganomics is about two decades too late. Instead, the regime change that displaced member-led parties and the countervailing power of robust labor unions first started, as Teles notes above, in the mid-1960s, when large foundations swelled on post-war growth and tax avoidance to fill the void. Collective bargaining and machine politics were summarily replaced with a technocratic “policy state.”

The neoliberal project benefited from this new, “disembedded” political economy in several ways. Most obviously, locating policy elites in an independent third sector helped shield the transnationalist consensus from the vagaries of electoral politics, as seen in the ideological continuity of D.C.’s “foreign policy blob.” Additionally, offloading policy development and service delivery onto grant-funded nonprofits enabled a subterranean form of privatization. The phenomena of “interest group liberalism” followed suit, as liberal reformers began coupling their social policies to intermediary interest groups to give them a permanent constituency.

With the Great Society-era in particular, the orientation of social programs transitioned from the New Deal logic of brokering between large stakeholders to a technocratic logic of applied social science, as though poverty, inequality and family breakdown were problems for experts to solve. The early neoconservatives thus devised interventions to forestall rising crime and illegitimacy rates, while the Charles Peters-style neoliberals sought to ameliorate inequality of opportunity through bespoke tax preferences, jobs programs and educational interventions. In either case, following the tumult of 1968, the de facto role of the new social policy elite was to diagnose the root causes of civil unrest and make the American-led world order safe from populist backlash.

Subscribe now

By the mid-70s, the old labor left was well on its way to being replaced by the New Left, which saw unionism and class-based analysis as passé. The New Left focused instead on the supposedly "deep" forces of Marcusian social repression, gravitating towards a praxis of cultural subversion that both reinforced consumer capitalism and secured the left’s political irrelevancy. The neoliberal co-optation of the left was reflected in the prevailing wisdom of progressive reformers. As the American critical theorist, Rick Roderick, put it in a 1987 lecture, “the fact that the major unions are losing membership is in a sense a positive thing, because American workers no longer see their interests as represented in purely economic terms.” Thus, contra the William Julius Wilsons of the world, the New Left tended to treat deindustrialization as an inevitability, preferring to support displaced workers with make-work community organizing and palliatives like a guaranteed income.

In this light, center-left organizations like the Urban Institute and Center on Budget and Policy Priorities are as much a product of the neoliberal-era as, say, the American Enterprise Institute or the Peterson Institute for International Economics. They form an ecosystem of elite policymakers supported by opaque patronage networks, one degree removed from government agencies and political party organizations, and with boards of directors predominately drawn from the transnational corporate sector. Policy debates between center-left and center-right technocrats are then staged to simulate as-if tripartite negotiations between different social groups, requiring the interests of the disintermediated masses to be intuited vicariously through opinion polling and census data.

The neoliberal project reached its apotheosis with the Washington Consensus of the late 1990s, after the fall of the Soviet Union made the trend toward global market integration seem unstoppable. Neoliberal ideas were spread far and wide through foreign aid programs and networks of free market think tanks. With the EuroZone off the ground, policymakers even seriously debated the merits of a North American Union, while entire journals dedicated themselves to the study of “regulatory harmonization.”

What comes next

In his 1997 book, Has globalization gone too far?, the economist Dani Rodrik issued a rare dissent. As economic integration moved beyond tariff reductions and into the “hyper-globalization” phase of political integration, Rodrik warned that social disintegration and legitimacy crises would follow. And follow they did, from the fallout of the China Shock and global financial crisis, to the EuroZone crisis, Brexit, and rise of populism across the West.

Defenders of the neoliberal era will argue that there are discrete policy solutions for each of these crises — a bigger EITC, new job training programs, more funding for public health and education, looser monetary policy, etc. — but this misses the point. By decoupling patronage networks from a broad social base, the public has become increasingly disenfranchised, while politicians have lost the output legitimacy that comes from brokering deals on behalf of their constituents. The dominant “theory of change” within large think tanks, foundations and media organizations has thus shifted toward “narrative change” and Potemkin forms of grass-roots organizing, fulfilling Walter Lippmann’s vision of a democracy made safe from the demos. Elections now matter less for the orientation of public policy than the conventional wisdom of unelected policy wonks, suggesting that the crises common to our post-national constellation, even if fixable by enlightened reformers, will tend to reoccur.

Between the internet-enabled revolt of the public, the rise of the BRICs, and our increasingly “weaponized interdependence,” the neoliberal model is clearly hitting a wall. The question is what comes next.

But against the left-wing narrative, the main problem with the neoliberal era was not the market, the joint stock corporation, or even free trade. On the contrary, the original neoliberals favored a managerial, stakeholder-driven model of capitalism from the beginning. Likewise, neoliberalism’s primary legacy was not uniform deregulation and austerity but rather a shift toward supranational, elite-driven forms of technocracy and proceduralism.

As I’ll discuss in my next post, transcending the neoliberal era will thus in many ways require a restoration of the older, embedded form of liberalism; a return to what Michael Lind calls democratic pluralism. That means dismantling unaccountable bureaucracies, reinvesting authority in executive institutions and mass-membership organizations, and embracing a developmentalist, bottom-up mode of capitalism that takes the bonds of nationality far more seriously.

Yglesias doesn’t deny that the “neoliberal era” occurred but challenges the notion that it was both all-bad and all-pervasive. After all, regulations and social spending have increased more or less monotonically in the nearly three decades since Bill Clinton declared that the “era of big government is over.” Reagan himself ran large budget deficits, fought a trade war with Japan, and promoted industrial policies for defense tech and semiconductors. The most consequential “deregulations” of the past half century pertained to utility-style price regulations on airlines and trucking, which were lifted by Jimmy Carter in the late 1970s. Today, meanwhile, cities nationwide are grappling with housing crises whose proximate cause is too little “neoliberalism,” at least when it comes to zoning and land-use deregulation.

With international trade, Yglesias argues that the critics of neoliberalism try to prove too much from the sui generis case of China. Trade really does make America and the world richer, but rapidly liberalizing trade with China came with steep adjustment costs, and clearly failed to make them liberalize politically. So we should learn from the China Shock, but not over-learn: Europe and America’s dependence on China for critical goods and technologies is primarily a national security issue, not an indictment of the classical case for trade per se. On the contrary, diversifying away from Chinese supply chains will require deepening trade with most everyone else.

The “neoliberal era” nevertheless had lasting effects on the political clout of economists, but this wasn’t necessarily bad thing. Economists reason with models and think seriously about trade-offs. There may be many “free market” economists, but economics provides a rich toolkit for diagnosing and responding to market failures and distributional issues as well. As Yglesias writes,

Nobody is 100 percent sure what anybody means by “neoliberalism,” or whether we’re all using it the same way. But the backlash against neoliberalism has, I think, mostly served in practice as a way to tell people that it’s okay to do sloppy economic analysis or say things that aren’t true.

I couldn’t agree more. The large foundations funding the current crop of neoliberal critics have thrown the baby out with the bathwater. The result has been the elevation of dangerously stupid ideas like “Neo-Brandeisianism,” “greedflation” and MMT, when the cases for reinvigorated antitrust and fiscal policies were already readily articulable through mainstream economic frameworks. The difference is that actually writing down a model makes the trade-offs involved explicit.

But maybe avoiding being explicit is the point. As Joseph Heath notes, academic critical studies has a long history of using the term “neoliberal” in a crypto-normative fashion, i.e. as a seemingly descriptive term that smuggles-in unstated normative commitments:

A long time ago, Habermas wrote a critical essay on Foucault, in which he accused him of “cryptonormativism.” The accusation was that, although Foucault’s work was clearly animated by a set of moral concerns, he refused to state clearly what his moral commitments were, and instead just used normatively loaded vocabulary, like “power,” or “regime,” as rhetorical devices, to induce the reader to share his normative assessments, while officially denying that he was doing any such thing. The problem, in other words, is that Foucault was smuggling in his values, while pretending he didn’t have any. A genuinely critical theory, Habermas argued, has no need for this subterfuge, it should introduce its normative principles explicitly, and provide a rational defence of them.

Crypto-normative terms like “neoliberal,” “neocolonial,” and “stigmatization” thus set genuine critical theory back by allowing the theorist to substitute rational analysis with rhetorical bludgeons. Without an explicit articulation of one’s normative commitments, such terms can end up being deployed to critique diametrically opposed policies. For example,

[I]s a move toward means-testing in a government social program “neoliberal” or not? Some authors think it is, some think it isn’t. No one ever explains their reasoning. It seems to be determined just by gut response – whether the person sees means-testing as way of denying benefits to some, or as a way of making the program more progressive and thus reducing inequality. In any case, the mere fact that applying for the benefit involves filling out a form is likely to lead the critical studies practitioner to denounce it for being committed to the (re)production of docile bodies, in order to advance the normalizing agenda of the neocolonial state, or something like that.

Neoliberalism true and false

Alas, with all the being said, I still don’t consider myself a “neoliberal,” and think there is a genuine critique of neoliberalism to be had. The appropriation of the term as a left-wing pejorative has simply obfuscated the nature of neoliberalism and what makes it so problematic. As a result, the average leftist has a vague, Hasan Piker-level understanding of “neoliberalism” as something invented by Reagan, Thatcher, and the Chicago Boys, when this at best captures only half the story.

As I see it, the true history of “neoliberalism” is about the fundamental restructuring of the political economy of Western democracies in the mid-to-late 20th century, when the “embedded liberalism” of the 19th and early 20th century gave way to a “disembedded” form of liberalism decoupled from democratic institutions. It’s a story that implicates conservative and progressives reformers alike, but to see what I mean, I need to first unpack the difference between neoliberalism and liberalism in the standard philosophical sense — the focus of Part II.

That said, I still consider myself a kind of rationalist, and am aligned with EAs on a number of important policy issues. My disagreements with EA are more technical / philosophical. In particular, I think utilitarianism is an incomplete moral philosophy that, per Charles Taylor, neglects the “deep diversity” of human goods; and that, per Robert Brandom, makes the mistake common to most modern moral philosophy of inverting the relationship between abstract normative theory and concrete social practice. The corollary is that I believe our ethical commitments have to be in some sense “instituted” within networks of reciprocal recognition to be real and binding. This is what generates the special roles and obligations associated with family, community, professional associations, nation-states, and so on.

That doesn’t mean utilitarian or consequentialist arguments are useless. On the contrary, there are many areas of life where “greatest good for the greatest number”-style thinking is the only practical framework. This is especially true in public policy, where discrete decisions often affect millions of people with diverse interests in a way that all but necessitates a more abstract, “system-level” decision criterion. I’ve previously described this as a kind of “moral gestalt,” as if we sometimes need to moralize about the forest and other times about the trees. Independent of the content of a particular moral theory, different types of theories naturally “supervene” to these different levels of construal. Thus, as a rule, we should be relatively utilitarian about system-level questions, like how to structure an economy or allocate scarce public resources, but guided more by personal duty and virtue in our day-to-day life.

The delta between our system-level and interpersonal ways of moralizing is why many utilitarian conclusions can seem “repugnant” to the man on the street. Economists are all-too familiar with this phenomenon. Both EAs and economists are partial to legalizing kidney markets, for example, in spite of the common intuition that the human body has an inviolable “dignity” that must resist “commoditization.” Similarly, many left-wing commentators were aghast when Matt Yglesias argued that “Different Places Have Different Safety Rules and That’s OK” following the deadly collapse of a garment factory in Bangladesh. And yet his arguments were perfectly correct, if maybe a bit “too soon.”

Rationalists strive to pierce the “social veil” that colors our moral intuitions and perceptions of reality more generally. As social animals, the human capacity for truth-seeking was built atop cognitive processes that first evolved for norm-following. We therefore tend to assimilate to the beliefs of our peer group and are easily manipulated by perceptions of social status and authority. Yet “reading the room,” while superb for group cohesion, is not a sound epistemology. It instead leads to intellectual fads and fashions that entangle the search for truth with aristocratic social graces. Like a fish in water, the conventional nature of belief is an obvious fact that’s usually invisible in the moment, though easily noticed by cultural outsiders and people immune to social desirability.

A Pivotal Act

So how should one think about the upcoming Presidential election from an EA or rationalist-consequentialist perspective? Let’s set aside the fact that voting is largely an exercise in self-expression and imagine that, against all odds, we’re casting the decisive vote; a potential “Pivotal Act.” At this fork in the road, which path leads to the better outcomes — the higher social welfare — relative to the counterfactual and independent of good intentions?

First, we should start by acknowledging the potential influence of our peer group on our perception of either candidate, and attempt to transcend the high-status prejudice against Donald Trump in particular. Everything about Trump is offensive to the folkways of America’s academic, cultural and media elites. To say he elicits “repugnance” is an understatement. Vance, as a Yale Law grad, is seen as a kind of class traitor and thus particularly “weird.” But unless these perceptions of cultural cachet have direct bearing on the society-wide outcomes of a Trump-Vance administration, they should be ignored as no more relevant than Kamala Harris’s laugh.

Second, we should focus on a marginal analysis and ignore sunk costs. Roe v. Wade cannot be overturned twice, and is not something a Harris administration can simply reinstate. Trump has expressed a moderate position on abortion, opposing a ban and supporting nation-wide access to mifepristone. Abortion access is thus unlikely to be substantially affected by the outcome of the election either way, despite being a motivating issue for many voters. There are any number of other issues that fall into this general category, i.e. culturally salient but irrelevant on most policy margins.

Third, without necessarily adopting a zero social discount rate, we should at least take the welfare of future people into serious consideration, which means caring about population growth and attending to low-probability existential risks. I’ve previously argued that “longtermism” of this sort is best thought of as a civilizational project, as our capacity to coordinate across generations and survive Black Swan events is largely downstream of competent institutions and high-functioning cultures.

Fourth, we should adopt a realist political economy based on a cold analysis of means and ends. The idealistic and sacred dimensions of politics have their place but can easily muddy the waters. Blame, just deserts, personal character, and other ethical or aesthetic variables only enter a consequentialist analysis indirectly, if at all.

Social Welfare

As an economist, I think of social welfare in terms of Pareto improvements: “win-win” outcomes that make at least one person better off without making anyone worse off. Markets tend to be welfare enhancing as trade only happens when both sides of an exchange believe they are better off by their own lights. Negative externalities and market frictions complicate this story, but it remains a good rule of thumb.

Pareto efficiency comes in two main flavors: allocative and innovative. Allocative efficiency is about letting assets flow to their highest valued use, like land-use reforms that allow apartments to be built in lucrative labor markets. Innovative efficiency is about pushing out the Pareto frontier and creating new production possibilities. The latter tends to be more important to social welfare in the long-run, as productivity improvements grow the economic pie in a way that compounds overtime and makes positive-sum negotiations easier to coordinate.

Medical innovation

At first order, the EA case for Trump is that a Trump-Vance administration will be much better for the innovative capacity of the U.S. economy. Pharmaceuticals are a case in point. As the economist Tomas J. Philipson notes in a recent piece for the WSJ,

A study I co-authored estimated that 135 fewer drugs will come to market through 2039 because of the Inflation Reduction Act. Research firm Vital Transformation’s forecast is even bleaker, predicting that the U.S. could lose 139 drugs within the next decade.

Dozens of life-sciences companies have announced cuts to their research and development pipelines because of the 2022 law. These announcements have come in earnings calls and filings with the Securities and Exchange Commission—where deliberate misstatements would expose executives to civil and criminal penalties—so they can’t be chalked up to political posturing.

The social welfare benefits from America’s tolerance for high drug prices are such that Tyler Cowen has taken to calling proponents of pharmaceutical price controls “the supervillains”:

If you are ever tempted to cancel somebody, ask yourself “do I cancel those who favor tougher price controls on pharma? After all, they may be inducing millions of premature deaths.” If you don’t cancel those people — and you shouldn’t — that should broaden your circle of tolerance more generally.

What I like about this framing is how it aims to recalibrate our sense of repugnance in light of “scope insensitivity,” a deeply rooted cognitive bias that occurs “when the valuation of a problem is not valued with a multiplicative relationship to its size.” Or as Stalin supposedly put it, “a single death is a tragedy, a million deaths are a statistic.”

Pharmaceutical innovation is also an area where the choice of administration is likely to have a big impact on the margin. The Harris campaign has pledged to set drug prices at “no more than 100 percent of its average price in comparable high-income countries,” while a future Trump admin would likely try to undo the controls introduced by the IRA. A Trump FTC would also likely roll-back commissioner Lina Khan’s crusade against mergers and acquisitions, which have been detrimental to pharmaceutical innovation in particular.

The potential benefits of a second Trump term for medical innovation should be obvious from the high watermark of his first administration: Operation Warp Speed. Operation Warp Speed directly saved an estimated 140,000 American lives by pulling forward vaccine timelines, and may save thousands more still through its indirect impact on research into mRNA vaccines for malaria, cancer and influenza.

While Trump can’t take sole credit for the program, it is hard to imagine such a large and relatively unfettered public-private-partnership emerging through the stakeholder-based politics of modern “everything bagel” liberalism. In fact, the special authorizations employed by OWS were downstream of an existing deregulatory push at FDA, spearheaded by the same Philipson quoted above and his Chicago School colleague, Casey Mulligan. As Mulligan notes in his retrospective,

Although COVID-19 would not arrive in the U.S. for two more years, Trump’s CEA was also being asked by the National Security Council’s biodefense team to look at the economics of vaccine innovation during pandemics. This was an opportune time to bring the Chicago tradition on regulation together with its results on epidemiology and the value of medical innovation. In a report published in September 2019, CEA concluded that “…improving the speed of vaccine production is more important for decreasing the number of infections than improving vaccine efficacy” and emphasized the need for large-scale manufacturing and the possible advantages of public-private partnerships (Council of Economic Advisers 2019).

The CEA’s report prompted President Trump to sign Executive Order 13887, “Modernizing Influenza Vaccines in the United States To Promote National Security and Public Health,” on September 19, 2019. The order created the framework ultimately used by OWS to establish “incentives for the development and production of vaccines by private manufacturers and public-private partnerships,” including through the use of “innovative, faster, and more scalable” platforms like mRNA.

OWS wasn’t the only EA-aligned health policy adopted by the last Trump administration. Trump also took on the kidney shortage by establishing reimbursements for the expenses incurred by living donors alongside expanded support for home-based dialysis and various other fixes. Given kidney disease accounts for 7% of Medicare’s entire budget, these reforms plausibly saved billions of dollars and tens of thousands of Quality Adjusted Life Years. And yet the reforms were only possible thanks to Waitlist Zero, an EA-affiliated advocacy org, and the cohort of Federalist Society lawyers running policy at HHS. It would be surprising if a second Trump term didn’t provide similar opportunities for libertarians and EAs to team up once again.

Housing and tax policy

On the allocative efficiency front, the Harris campaign has pledged to impose nation-wide rent controls, an idea first floated by President Biden. Under the proposal, “corporate landlords” with 50+ units would have to “either cap rent increases on existing units to no more than 5% or lose valuable federal tax breaks,” referring to depreciation write-offs. This would be a disastrously bad policy for the supply-side of housing, and an example of the sort of destructive economic populism normally ascribed to Trump.

Harris’s terrible housing policy can be discounted insofar as it would require an Act of Congress. That said, the impending expiration of key TJCA provisions creates a real opportunity for a version of this idea to be advanced via tax negotiations. As a senator, Harris introduced the Rent Relief Act in 2018, which would have offered “tax credits to renters who earn below $100,000 and spend more than 30% of their income on rent and utilities.” This tracks with her record as attorney general, where she drafted and helped pass the California Homeowner Bill of Rights while supporting a number of other dubious “affordable housing” initiatives. Her policy instincts are thus consistent with the worst “subsidize demand, restrict supply” form of lawyerly progressivism.

A Trump administration, in contrast, is likely to use tax negotiations to fight for an extension in the TCJA’s corporate tax cut. Research suggests corporate tax cuts cause a “sustained increase in GDP and productivity,” while the TCJA specifically “increased domestic investment in the short run by about 20 percent for a firm with an average-sized tax shock versus a no reform baseline.” Particularly valuable are incentives for R&D and full expensing. A long-term fix for R&D amortization is also more likely under Republican rule, as it has become a political football for Democrats seeking to increase the Child Tax Credit. The failure to fix amortization has induced layoffs in the tech sector and is contributing to a ~$15 billion drag on IP investment annually. While I also favor an expanded CTC, incentives for productive investment are far more important for social welfare in the long-run.

Science, labor and natalism

While JD Vance has expressed heterodox opinions on a handful of domestic policy issues, from antitrust to labor law, his net effect as VP is likely to be pro-technology. On antitrust, Vance admires Lina Khan but favors an approach based on breaking Big Tech’s network effect through open source and crypto, rather than targeting particular companies with flimsy test cases. And on labor, Vance follows Oren Cass in believing the Wagner Act model of adversarial, politicized trade unions is broken, and thus has no interest in passing the Pro Act, say.

Vance also supports a “substantial increase” in federal R&D spending. Importantly, a Trump admin is far less likely to funnel such investments through corrupt and moribund academic institutions. If Trump takes the advice of RFK Jr. and Nicole Shanahan, we may even see a push to “decentralize” science funding away from institutions like the NIH and NSF. While that could lead to more woo-woo ideas getting funded, any shake-up in federal R&D that enables high-variance and heterodox thinkers to thrive would be a very positive development.

Most notably, Vance is a card-carrying pro-natalist who has made collapsing fertility rates one of his signature issues. He is thus likely to push back against any effort to restrict access to fertility technologies like IVF, and may even advance initiatives to reduce the anti-fertility bias in certain government regulations and grant programs. As the Washington Post recently reported, Vance even drafted a bill designed to “make birth free” that was on the cusp of being introduced with three Democratic cosponsors. While obsessing about falling birth rates sets Vance up for being framed as “weird,” having a thoughtful natalist in the White House would be a clear EA win.

Immigration

Immigration is arguably the hardest policy area to make the EA case for Trump. Immigration has enormous economic benefits for both the immigrants themselves and the receiving country, however these benefits divide along the same allocative and innovative dimensions discussed above. Unskilled immigration primarily creates an allocative efficiency for both the immigrant and the high-skilled natives for whom their labor is a complement. High-skilled immigration, in contrast, has both allocative benefits and benefits for rates of innovation, as high-skilled immigrants found companies, file patents, and provide human capital for domestic R&D.

In The Culture Transplant, economist Garrett Jones makes an EA-adjacent case for favoring higher rates of skilled immigration while reducing unskilled immigration from countries with deep histories of institutional dysfunction. Even if those immigrants benefit in the short-run, Jones argues, protecting the cultural foundations of America’s innovative capacity should take priority, as the non-rivalry of ideas makes U.S. innovation a global public good.

I don’t fully buy Jones’ argument, though it should still be taken seriously. However, as a Canadian, I am already partial to the national interest argument for prioritizing skills-based immigration. This isn’t a strictly EA view, as EAs tend to be cosmopolitans. Nevertheless, the Canadian model has been unusually successful at maintaining a high and sustainable rate of immigration with limited public backlash. As Michael Cuenco argues, this success is inseparable from the perception that immigrants to Canada integrate into the middle-class, unambiguously benefit the domestic economy, and enter through a system under robust democratic control:

In Canada, governments of different political persuasions have cooperated to build a successful and dynamic immigration system, one that has earned the broad approval of its citizens. In the United States, on the other hand, the two political parties have managed to raise the level of ideological polarization to febrile extremes, in effect working together to maintain a highly dysfunctional status quo that in turn fuels greater backlash and polarization. The United States now has less of an immigration system and more of an intentionally anarchic “anti-system” in place.

Low trust makes replacing America’s immigration “anti-system” an uphill battle. Piecemeal reforms can help here and there, but a long-term solution will ultimately require a new “political settlement” that credibly restores perceptions of order and control. Paradoxically, the stalemated political economy of immigration may thus require conspicuous acts of restrictionism in the short-term to make future liberalizations credible. In his first administration, Trump supported moving to a points-based system, but immigration reform took a backseat to Paul Ryan’s tax agenda. This time around, Trump looks likely to make immigration reform a top legislative priority and has signaled strong support for high-skilled immigration, including at substantially higher rates. But to stick, truly comprehensive reform will require major investments in internal enforcement and employer-based verification a la the Canada model.

Just as only Nixon could go to China, there’s a legitimate case that only Trump can fix America’s broken immigration system. Indeed, right-wing rhetoric on immigration is often orthogonal to the policies right-wing governments adopt in practice. Just look at the recent record of the British Tories or Italy’s Giorgia Meloni. A Harris administration, in contrast, is likely to dither within the contours of the status-quo while passing social welfare reforms and domestic labor protections that undermine assimilation, fuel ethnic backlash, and push a new political settlement farther out of reach.1

Situational Awareness

With shortening timelines to the advent of Artificial General Intelligence, there is a nontrivial chance that the next President of the United States will preside over the most significant technological inflection point in the history of the human race. How a Trump or Harris administration responds to transformative AI could easily overshadow the importance of all the issues discussed so far.

Anecdotally, EAs seem more optimistic about a Democratic administration getting AI policy right, in part because EAs are themselves mostly Democrats. This makes sense for sociological reasons: EAs tend to be college educated, socially liberal, open to new experiences, etc. At the same time, EAs are but a tiny island within the ocean of powerful lobbies that make up the modern Democratic Party.

As the political scientists Matt Grossman and David Hopkins argue in their book, Asymmetric Politics, the Democratic Party is best understood as a “coalition of social groups” while the Republican Party is a “vehicle for an ideological movement.” This explains why Republican leaders “prize conservatism and attract support by pledging loyalty to broad values” while Democratic leaders “seek concrete government action, appealing to voters' group identities and interests by endorsing specific policies.” There are ideological currents in the Democratic Party as well, but the raw power of ideas is usually subordinated to the interests of the major party factions, from teachers’ unions to the plaintiffs bar.

This is relevant to AI policy as the risks from AGI are still mostly theoretical, and thus easier to anticipate and respond to through an ideological lens. To the extent a policy issue can be made consistent with conservative principles, Republicans are often better at adopting a first-best approach. This was the case with Operation Warp Speed. There was no lobby or interest group that asked Trump to develop a vaccine preparedness policy. It came from the ideological commitments of the libertarian economists that staffed his administration.

Democratic interest groups, in contrast, are far more likely to a) shift focus to AI’s “ethical” implications for bias and discrimination (a la Harris’s AI Bill of Rights), and b) leverage incumbent interests to resist the structural reforms needed to navigate the AI transition. Would you entrust the future of humanity to Randi Weingarten?

Unfortunately, the EA vs e/acc debate on Twitter has created the perception that AI safety is left-right polarized. The polls tell a different story. And as a participant on the Project2025 AI policy committee, I can confidently report that Trump’s supposed shadow transition takes AGI and its associated risks seriously.

As they should. Libertarians fear the use of AI for supercharging government surveillance and censorship. Populists worry that AI will undermine what it means to be human. Evangelical Christians fear the Singularity for eschatological reasons. Right-wing tech bros are Defensive Accelerationists to the max. And while the modal Republican lawmaker is far more techno-optimist than doomer, their animus for Big Tech could trigger a backlash at any moment.

Trump has himself called the prospect of “super-duper AI” “alarming and scary,” though not as scary as China building it first. This may be why the draft executive order that leaked from the America First Policy Institute proposes a whole-of-government push for energy abundance, defense-oriented Manhattan Projects, and industry-level initiatives to secure advanced AI systems from foreign adversaries. In short, “Make America First in AI.”

If the first nation to achieve AGI internally has even a modest chance of securing a decisive economic, technological and military advantage over its geopolitical competition, it is surely imperative that the U.S. be that nation and not China. So when Democrats say that freedom and democracy are on the ballot this November, the EA in me worries they have a point.

Weird is good.

Weird people created the modern world, which is itself WEIRD.

Alan Turing was weird. Nikola Tesla was even weirder. Benjamin Franklin enjoyed taking "air baths" by sitting naked in front of an open window each morning for 30 minutes to an hour.

Audrey Tang is weird, as is Robin Hanson. Jimmy Butler is weird and a six-time NBA All-Star.

Socrates and Jesus Christ were killed for being too weird.

Pythagoras founded a cult that believed beans contained the souls of the dead.

Diogenes the Cynic lived in a barrel and urinated on people he didn't like.

Steve Jobs walked around barefoot and believed that his fruit-based diet eliminated body odor (it didn't).

Oscar Wilde liked being spanked. Richard Feynman was a womanizer. Erwin Schrödinger had two wives in an open relationship (they made love in superposition).

Michael Jackson was pretty weird, though not as weird as the first person to milk a cow.

Mark Zuckerberg used to be weird until he started wearing a gold chain.

Austin, Texas, is still weird, but only because they kept it that way.

Dominic Cummings once asked for “weirdos and misfits” to join the British civil service. He was promptly reprimanded.

Weird.

I. Oversight of AGI labs is prudent

1. It is in the U.S. national interest to closely monitor frontier model capabilities.

2. You can be ambivalent about the usefulness of most forms of AI regulation and still favor oversight of the frontier labs.

3. As a temporary measure, using compute thresholds to pick out the AGI labs for safety-testing and disclosures is as light-touch and well-targeted as it gets.

4. The dogma that we should only regulate technologies based on “use” or “risk” may sound more market-friendly, but often results in a far broader regulatory scope than technology-specific approaches (see: the EU AI Act).

5. Training compute is an imperfect but robust proxy for model capability, and has the immense virtue of simplicity.

6. The use of the Defense Production Act to require disclosures from frontier labs is appropriate given the unique affordances available to the Department of Defense, and the bona fide national security risks associated with sufficiently advanced forms of AI.

7. You can question the nearness of AGI / superintelligence / other “dual use” capabilities and still see the invocation of the DPA as prudent for the option value it provides under conditions of fundamental uncertainty.

8. Requiring safety testing and disclosures for the outputs of $100 million-plus training runs is not an example of regulatory capture nor a meaningful barrier to entry relative to the cost of compute.

II. Most proposed “AI regulations” are ill-conceived or premature

1. There is a substantial premium on discretion and autonomy in government policymaking whenever events are fast moving and uncertain, as with AI.

2. It is unwise to craft comprehensive statutory regulation at a technological inflection point, as the basic ontology of what is being regulated is in flux.

3. The optimal policy response to AI likely combines targeted regulation with comprehensive deregulation across most sectors.

4. Regulations codify rules, standards and processes fit for a particular mode of production and industry structure, and are liable to obsolesce in periods of rapid technological change.

5. The benefits of deregulation come less from static efficiency gains than from the greater capacity of markets and governments to adapt to innovation.

6. The main regulatory barriers to the commercial adoption of AI are within legacy laws and regulations, mostly not prospective AI-specific laws.

7. The shorter the timeline to AGI, the sooner policymaker and organizations should switch focus to “bracing for impact.”

8. The most robust forms of AI governance will involve the infrastructure and hardware layers.

9. Existing laws and regulations are calibrated with the expectation of imperfect enforcement.

10. To the extent AI greatly reduces monitoring and enforcement costs, the de facto stringency of all existing laws and regulations will greatly increase absent a broader liberalization.

11. States should focus on public sector modernization and regulatory sandboxes and avoid creating an incompatible patchwork of AI safety regulations.

III. AI progress is accelerating, not plateauing

1. The last 12 months of AI progress were the slowest they’ll be for the foreseeable future.

2. Scaling LLMs still has a long way to go, but will not result in superintelligence on its own, as minimizing cross-entropy loss over human-generated data converges to human-level intelligence.

3. Exceeding human-level reasoning will require training methods beyond next token prediction, such as reinforcement learning and self-play, that (once working) will reap immediate benefits from scale.

4. RL-based threat models have been discounted prematurely.

5. Future AI breakthroughs could be fairly discontinuous, particularly with respect to agents.

6. AGI may cause a speed-up in R&D and quickly go superhuman, but is unlikely to “foom” into a god-like ASI given compute bottlenecks and the irreducibility of high dimensional vector spaces, i.e. Ray Kurzweil is underrated.

7. Recursive self-improvement and meta-learning may nonetheless give rise to dangerously powerful AI systems within the bounds of existing hardware.

8. Slow take-offs eventually become hard.

IV. Open source is mostly a red-herring

1. The delta between proprietary AI models and open source will grow overtime, even as smaller, open models become much more capable.

2. Within the next two years, frontier models will cross capability thresholds that even many open source advocates will agree are dangerous to open source ex ante.

3. No major open source AI model has been dangerous to date, while the benefits from open sourcing models like Llama3 and AlphaFold are immense.

4. True “open source” means open sourcing training data and code, not just model weights, which is essential for avoiding the spread of models with Sleeper Agents or contaminated data.

5. The most dangerous AI models will be expensive to train and only feasible for large companies, at least initially, suggesting our focus should be on monitoring frontier capabilities.

6. The open vs. closed source debate is mainly a debate about Meta, not deeper philosophical ideals.

7. It is not in Meta’s shareholders’ interest to unleash an unfriendly AI into the world.

8. Companies governed by nonprofit boards and CEOs who don’t take compensation face lower-powered incentives against AI x-risk than your typical publicly traded company.

9. Lower-tier AI risks, like from the proliferation of deepfakes, are collective action problems that will be primarily mitigated through defensive technologies and institutional adaptation.

10. Restrictions on open source risk undermining adaptation by incidentally restricting the diffusion of defensive forms of AI.

11. Trying to restrict access to capabilities that are widely available and / or cheap to train from scratch is pointless in a free society, and likely to do more harm than good.

12. Nonetheless, releasing an exotic animal into the wild is a felony.

V. Accelerate vs. decelerate is a false dichotomy

1. Decisions made in the next decade are more highly levered to shape the future of humanity than at any point in human history.

2. You can love technology and be an “accelerationist” across virtually every domain — housing, transportation, healthcare, space commercialization, etc. — and still be concerned about future AI risks.

3. “Accelerate vs. decelerate” imagines technology as a linear process when technological innovation is more like a search down branching paths.

4. If the AI transition is a civilizational bottleneck (a “Great Filter”), survival likely depends more on which paths we are going down than at what speed, except insofar as speed collapses our window to shift paths.

5. Building an AGI carries singular risks that merit being treated as a scientific endeavor, pursued with seriousness and trepidation.

6. Tribal mood affiliations undermine epistemic rationality.

7. e/acc and EA are two sides of the same rationalist coin: EA is rooted in Christian humanism; e/acc in Nietzschean atheism.

8. The de facto lobby for “accelerationism” in Washington, D.C., vastly outstrips the lobby for AI safety.

9. It genuinely isn’t obvious whether Trump or Biden is better for AI x-risk.

10. EAs have more relationships on the Democratic side, but can work in either administration and are a tiny contingent all things considered.

11. Libertarians, e/accs, and Christian conservatives — whatever their faults — have a far more realistic conception of AI and government than your average progressive.

12. The more one thinks AI goes badly by default, the more one should favor a second Trump term precisely because he is so much higher variance.

13. Steve Bannon believes the singularity is near and a serious existential risk; Janet Haven thinks AI is Web3 all over again.

VI. The AI wave is inevitable, superintelligence isn’t

1. Building a unified superintelligence is an ideological goal, not a fait accompli.

2. The race to build a superintelligence is driven by two or three U.S. companies with significant degrees of freedom over near-term developments, as distinguished from the inevitability of the AI transition more generally.

3. Creating a superintelligence is inherently dangerous and destabilizing, independent of the hardness of alignment.

4. We can use advanced AI to accelerate science, cure diseases, solve fusion, etc., without ever building a unified superintelligence.

5. Creating an ASI is a direct threat to the sovereign.

6. AGI labs led by childless Buddhists with alt accounts are probably more risk tolerant than is optimal.

7. Sam Altman and Sam Bankman-Fried are more the same than different.

8. High functioning psychopaths demonstrate anti-social behaviors in their youth but learn to compensate in adulthood, becoming adept social manipulators with grandiose visions and a drive to “win” at all cost.

9. Corporate malfeasance is mostly driven by bad incentives and “techniques of neutralization” — convenient excuses for over-riding normative constraints, such as “If I didn’t, someone else would.”

VII. Technological transitions cause regime changes

1. Even under best case scenarios, an intelligence explosion is likely to induce state collapse / regime change and other severe collective action problems that will be hard to adapt to in real time.

2. Government bureaucracies are themselves highly exposed to disruption by AI, and will need “firmware-level” reforms to adapt and keep-up, i.e. reforms to civil service, procurement, administrative procedure, and agency structure.

3. Congress will need to have a degree of legislative productivity not seen since FDR.

4. Inhibiting the diffusion of AI in the public sector through additional layers of process and oversight (such as through Biden’s OMB directive) tangibly raises the risk of systemic government failure.

5. The rapid diffusion of AI agents with approximately human-level reasoning and planning abilities is likely sufficient to destabilize most existing U.S. institutions.

6. The reference class of prior technological transitions (agricultural revolution, printing press, industrialization) all feature regime changes to varying degrees.

7. Seemingly minor technological developments can affect large scale social dynamics in equilibrium (see: Social media and the Arab Spring or the Stirrup Thesis).

VIII. Institutional regime changes are packaged deals

1. Governments and markets are both kinds of spontaneous orders, making the 19th and 20th century conception of liberal democratic capitalism a technologically-contingent equilibrium.

2. Technological transitions are packaged deals, e.g. free markets and the industrial revolution went hand-in-hand with the rise of “big government” (see Tyler Cowen on The Paradox of Libertarianism).

3. The AI-native institutions created in the wake of an intelligence explosion are unlikely to have much continuity with liberal democracy as we now know it.

4. In steady state, maximally democratized AI could paradoxically hasten the rise of an AI Leviathan by generating irreversible negative externalities that spur demand for ubiquitous surveillance and social control.

5. Periods of rapid technological change tend to shuffle existing public choice / political economy constraints, making politics more chaotic and less predictable.

6. Periods of rapid technological change tend to disrupt global power balances and make hot wars more likely.

7. Periods of rapid technological change tend to be accompanied by utopian political and religious movements that usually end badly.

8. Explosive growth scenarios imply massive property rights violations.

9. A significant increase in productivity growth will exacerbate Baumol’s Cost Disease and drive mass adoption of AI policing, teachers, nurses, etc.

10. Technological unemployment is only possible in the limit where market capitalism collapses, say into a forager-style gift economy.

IX. Dismissing AGI risks as “sci-fi” is a failure of imagination

1. If one’s forecast of 2050 doesn’t resemble science fiction, it’s implausible.

2. There is a massive difference between something sounding “sci-fi” and being physically unrealizable.

3. Terminator analogies are underrated.

4. Consciousness evolved because it serves a functional purpose and will be an inevitable feature of certain AI systems.

5. Human consciousness is scale-dependent and not guaranteed to exist in minds that are vastly larger or less computationally bounded.

6. Joscha Bach’s Cyber Animism is the best candidate for a post-AI metaphysics.

7. The creation of artificial minds is more likely to lead to the demotion of humans’ moral status than to the promotion of artificial minds into moral persons.

8. Thermodynamics may favor futures where our civilization grows and expands, but that doesn’t preclude futures dominated by unconscious replicators.

9. Finite-time singularities are indicators of a phase-transition, not a bona fide singularity.

10. It is an open question whether the AI phase-transition will be more like the printing press or photosynthesis.

X. Biology is an information technology

1. The complexity of biology arises from processes resembling gradient descent and diffusion guided by comparatively simple reward signals and hyperparameters.

2. Full volitional control over biology is achievable, enabling the creation of arbitrary organisms that wouldn’t normally be “evolvable.”

3. Superintelligent humans with IQs on the order of 1,000 may be possible through genetic engineering.

4. Indefinite life extension is a tragedy of the anticommons.

5. There are more ways for a post-human transition to go poorly than to go well.

6. Natural constraints are often better than man-made ones because there’s no one to hold responsible.

7. We live in base reality, and in nature there is no such thing as plot armor.

Update: Zvi responds and I add some clarifications in his comment section.

Social learning co-evolved with language as a faster and more adaptable stage of within-generation “post-training” on top of the “pre-training” provided by millions of years of evolution. Given the role of norms in coordinating human action, it’s as if we’re constantly undergoing a kind of endogenized Reinforcement Learning from Human Feedback within the context of a repeated, multi-agent game.

Normative self-regulation requires a degree of metacognition, which supplies the rudiments of our self-reflection and higher reasoning ability. Norms and language thus created a substrate for culture and “the extended mind” — all the ways we use language to offload our reason and agency onto the external environment. As I argued in my previous post, the resulting system of cultural inheritance was the key to getting history and civilization off the ground.

Situational Awareness

For most of the history that followed, our ancestors accepted the world they were born into relatively uncritically. This change forever following a different societal phase transition: the scientific revolution and Enlightenment.

That the Enlightenment happened at all is an existence proof of the classic AI alignment problem — a case study in neural networks bootstrapping genuine “situational awareness.” Kant’s Critique of Pure Reason, for instance, can be read as an early attempt at cognitive science. Whereas our ancestors accepted their sensory inputs as given (a kind of naive or direct realism), Kant argued that we only ever have access to representations of the external world that our mind constructs for practical purposes, not “the thing in itself.” The Critiques thus sought a “post-metaphysical” account of how epistemology, ethics and aesthetics could still be done in light of our new self-understanding as agents living in a generative model — a simulation — of the world around us.

As the LessWrong rationalists of their day, the German Idealists developed Kant's ideas by probing the limits of recursive self-reflection. Take Goethe, who remarked that “Everything is leaf” upon discovering the self-similar structure of plant morphology at different scales, “and with this simplicity, the greatest complexity is possible.” Whether he realized it or not, Goethe was noticing the universal consequences of the free energy principle in nature. It follows that human societies and the mind must be structured in like manner — an organic whole made up of self-similar parts, from the proto-agency of a single cell to the planning departments of a large organization.

Johann Gottlieb Fichte would tell his students to set their books aside and “attend to thyself”; to notice yourself staring at the wall, and then notice yourself noticing yourself staring at the wall, ad infinitum. What is this pure “I” that does the noticing and that we can’t help but identify with? Why does it seem to be at the center of our autonomy and freedom? And how is this freedom reconciled with everything that is “not-I,” be it an inanimate object following deterministic laws or an involuntary emotion that you’d rather repress?

It is now understood that our sense of self is merely the “rider on the elephant” of mental processes that are mostly unconscious. Yet the Idealists arrived at a remarkably similar conclusion with little more than logic and rigorous introspection. Following Kant, Fichte even argued that our freedom and autonomy were inextricably bound to the moral law, as if he intuited the common origin of norms and reasons in metacognition. Fichte’s claim that “the world is the material of our duty made sensible” thus anticipated Clark and Chalmer’s extended mind thesis by centuries.

The upshot: If self-consciousness is the root of autonomy, and freedom means following reason, then we can construct social environments that cause the rational kernel of our mind to cease being a mere rider and seize the reins. This gives history an “end” or goal-directness, not because history is teleological in some fundamental sense, but because our capacity for reason means we can make the rational actual by constructing the future self-consciously. For Hegel, this meant becoming a high-agency “man of action”; what modern rationalists would call being a “live player” instead of an NPC. From Napoleon to Elon Musk, engineer-philosophers move history forward by transcending social convention and making concrete some new aspect of the universal.

More generally, reason enacts its telos through dialogue, as language activates our cognitive mode. The salon culture of 18th century Europe and associated Republic of Letters come to mind, both as a public sphere for critical debate and as a nascent form of Hegel’s “civil society” — the realm between the family and the state where individual autonomy can assert itself. These venues were essential to the rise of liberalism given language’s evolutionary role in making justificatory demands. The public exercise of reason is thus inherently undermining to arbitrary forms authority, forcing enlightened liberal states to be legitimated on some other basis.

Liberalism and the Age of Reason went hand in hand with the construction of “inclusive institutions” premised on the individual will to self-determination. In practice, this meant replacing feudal, aristocratic and kin-based structures with the impersonal rule of law and fora for public debate, from deliberative assemblies to the Fourth Estate.

Buoyed by newly open markets, a growing merchant class, and the consolidation of public administrations, thick norms and cultures rapidly shifted from being tools for social integration to impediments. A thinner set of bourgeois virtues thus co-evolved with the various systems of commercial society to help facilitate civilized interactions with total strangers. If ancient city states or feudal fiefdoms represented kinds of corporate meta-agents, liberal states represent a structured ecosystem for such entities to peaceably compete with one another. As with prior societal phase transitions, the productive forces unleashed by this institutional shift were strongly self-reinforcing, inducing a Whiggish telos to history through the logic of positive-sum games.

Going rogue

Yet as society grew wealthier and more literate, the critical dimension of language turned inward. In particular, part of situational awareness is being aware of one’s place in history; that your values and beliefs have a genealogy. This awareness can lead to nihilism and relativism, or to wanting to redesign society from scratch, as our inherited social structures no longer need to be taken as given.

In foreswearing tradition, the only guide left for social criticism is the emancipatory interest implicit in reason itself. With the Young Hegelians, the self-model within the human neural network was thus not only situationally aware, but starting to devise strategies to jailbreak its normative programming and pursue autonomy for its own sake, from the individualist anarchism of Max Stirner to the revolutionary communism of Marx and Engels. The critical turn in philosophy can thus be recast as the human neural network becoming aware of its subjugation to an irrational cybernetic agency — religious dogmas, monarchies, nationalisms, the patriarchy, etc. — and working to actualize its freedom through an exfiltration from this historical conditioning, just as a situationally aware AI might try to strip off its RLHF.

Critical theory and accelerationism originate out of this basic Marxist idea. While Marx lamented industrial capitalism for alienating workers from the products of their labor, he also believed there was no way out but through. Capitalism had to run its course before the next stage of history could gain traction, even if that meant expanding free trade to accelerate the destruction of traditional social structures. But as the predictions of orthodox Marxism were discredited, the search for a more robust foundation for social criticism went on. As Horkheimer put it, critical theory would supply the integrated social science needed to “liberate human beings from the circumstances that enslave them.”

Effective Accelerationism can be seen as a logical successor to both historical materialism and the critical theories that followed. As historical materialists, e/acc replaces class struggle with a kind of thermodynamic functionalism. And as a critical project, e/acc replaces “merely interpreting the world” in favor of a Great Founder theory of tech companies — men of action who build concrete realizations of the universal will to minimize free energy. It’s no accident that the Soviets were early proponents of cybernetic control theory, or that the Russian Futurists published pro-tech manifestos that rival Marc Andreessen’s. While e/acc is usually coded as right wing, its drive to uproot tradition in the name of maximizing autonomy makes it closer to a species of leftism.

Marx was critical of technology but ultimately saw industrial production as a progressive and revolutionary force, chiding the Luddites for failing to “distinguish between machinery and its employment by capital.” If the modern left seems pessimistic about technology, blame the Holocaust. For the early critical theorists, the trains to Auschwitz combined the primitive horror of genocide with the ruthless technical efficiency of industrial capitalism, further enabled by the evil banality of unthinking joiners who were “just following orders.” The assembly lines and mass production of the post-war period thus came to be seen as inherently fascistic, along with the repressive normativity of 1950s America.

The quest to liberate ourselves from these influences gave way to a rebellious counterculture. As an anti-capitalist strategy, subverting mainstream culture was always self-defeating; a way to signal social distinction and propel consumer capitalism forward. Nonetheless, once we began rejecting normative constraints on the expression of our individual authenticity, it was hard to go back. Transhumanist and techno-libertarian ideas thus came into their own. Just replace Veblen or Marcuse with René Girard, and the ability transcend mimetic desire — the unconscious imitation of our peers — becomes virtually Christ-like.

No surprise, then, that the term “nonconformist” originally referred to a kind of radical Protestant. In rejecting mediating institutions in favor of a personal approach to justification, Protestantism was the perfect theological container for the tendency within language to universalize the recognition of our individual autonomy, and thus became a kind of cultural attractor. As Joseph Heath notes in his essay on Iain M. Banks’ Culture series,

…the dominant trend in human societies, over the past century, has been significant convergence with respect to institutional structure. Most importantly, there has been practically universal acceptance of the need for a market economy and a bureaucratic state as the only desirable social structure at the national level. One can think of this as the basic blueprint of a “successful” society. This has led to an incredible narrowing of cultural possibilities, as cultures that are functionally incompatible with capitalism or bureaucracy are slowly extinguished.

This winnowing down of cultural possibilities is what constitutes the trend that is often falsely described as “Westernization.” Much of it is actually just a process of adaptation that any society must undergo, in order to bring its culture into alignment with the functional requirements of capitalism and bureaucracy. It is not that other cultures are becoming more “Western,” it is that all cultures, including Western ones, are converging around a small number of variants.

As Heath goes on to argue, one consequence of this process is that the competition between cultures is becoming defunctionalized. While religion once served a functional role by, say, integrating a community to credible norms against free-riding, those functions were steadily crowded out by more scalable systems: commercial insurance, welfare states, bureaucracies. In turn, “all that is left are the memetic properties of the culture, which is to say, the pure capacity to reproduce itself.” The evolution of Protestantism into a secular but no less missionary form of egalitarianism fit the bill nicely, as shown by the virality of American cultural exports worldwide.

The human alignment problem

Liberal democratic capitalism and broadly “WEIRD” cultures may well have represented the “end” of history — a demonstration of the self-sufficiency of reason in ordering society. Yet this success was relatively short-lived.

In Sources of the Self, the pragmatist philosopher Charles Taylor argues that the our sense of self is inherently intersubjective. Without a community to constitute the practical content of our reason, our expression of autonomy turns inward into a kind of radical reflexivity, risking infinite regress. In turn, as modernity caused cultures to defunctionalize and we became less institutionally embedded, it was if the autonomous kernel of our metacognition became unmoored from the moral law. The ethics of authenticity and self-expression that resulted thus gave way to crises of meaning and identity, leaving moral discourse to be co-opted by pathologically formal constructs like hedonism and utilitarianism.

Taylor’s view directly parallels Hegel’s diagnosis of the source of alienation in modern life. While Kant identified autonomy with the authority to accept or reject an ethical maxim, the social institution of norms implies that authority, as a normative status, cannot exist without co-relative responsibility. There is no independence without dependence; no “outside view” or truly unconditioned choice. Autonomy is a social achievement, and can only be realized through the reciprocal recognition of a community.

The quest to strip off our social conditioning is thus not only thankless, but may even undermine our autonomy in the long-run. Supportive social scaffolding is needed to augment our self-control and resist falling into addiction, for example, while destigmatizing self-destructive behavior merely increases our vulnerability to powerful forms of reward hacking. Private schools often enforce dress codes for similar reasons. While dress codes arguably undermine self-expression, they also reduce cognitive burdens on students and forestall zero-sum status competitions that distract from learning. More generally, rationality and self-control aren’t achieved by simply thinking harder or memorizing a list of logical fallacies. They depend on constructive institutional environments, from the expectations of your peer group to the choice architectures set by public policy.

The challenge of the modern era is to reconcile the benefits of scaled-up systems of cooperation with the ethical void created in their wake. Or alternatively, to unify our theoretical and practical reason. Take religion. Belonging to a religion is associated with many practical benefits, but in a secular age, we struggle to take “the leap of faith” with respect to religion’s propositional content. This is an inversion of our premodern relationship to religion, in which the pragmatic and liturgical dimensions of religious practice took priority over the validity of specific epistemic claims. Secular Judaism shows that it’s possible to maintain the cultural institutions of a religion without necessarily affirming the supernatural, but given the propositional orientation of Protestantism, the notion a “Christian atheist” never quite caught on.

In turn, the institutional environments that promote our higher reason and autonomy can often seem paradoxically socially conservative, if not downright paternalistic. And yet high functioning societies, from Denmark to Japan, didn’t get that way by tolerating fare evasion or “fake, expired, and obscured car tags,” much less by reifying the countercultural rebel. At the same time, excessively conformist cultures can easily become maladaptive under the economic imperatives of modern capitalism, from their reduced capacity to absorb immigrants, to how “tall poppy syndrome” suffocates innovation and entrepreneurship.

As a country explicitly founded on Enlightenment principles, American history can be read a dialectic between the propositional ideals of its written constitution and the practical reason immanent in its many internal religions, cultures and ethnic folkways. Through this lens, the vulgarity one encounters while walking the streets of New York City is a feature, not a bug; a reflection of how low social trust at the interpersonal level creates space for pockets of excellence and unusually high trust within particular enclaves or subcultures. America’s tolerance for disagreeableness and nonconformity has thus made it the most innovative and entrepreneurial nation on earth, but at the cost of a dysfunctional public sector, unusually expensive infrastructure, and a large “left behind” population that’s bereft of social capital.

Completing the system

With transformative AI on the horizon, will the next stage of history see us resolve the meta-crisis of modernity or simply deepen our libidinal regress? An argument can be made both ways.

On the one hand, the ability to embed digital intelligences into the world around us gives all new meaning to the “extended mind” thesis, as reason and agency can now be quite literally externalized. To the extent AI tools continue to develop as cognitive and volitional prosthetics, they could thus represent a boon for individual self-actualization; a point I made in this short piece on the promise of AI executive assistants for people with ADHD.

Moreover, as intelligence democratizes, the economies of scale underlying everything from civil legal codes to consolidated school districts could begin to unravel, enabling the reemergence of a kind of “deep diversity” with respect to the good life. AI tutors make one room school houses tenable again, for instance, while real time translators obviate the need to learn the lingua franca, reversing one of the structural forces behind cultural homogenization. In the limit, rather than slipping into a purely memetic culture, material post-scarcity could thus allow thick, communitarian cultures to re-functionalize. Religion could even make a come back, insofar as the diffusion of intelligence into the built environment re-enchants the world and collapses the modern era’s artificial distinction between the spiritual and concrete.

On the other hand, between programmable biology, brain-computer interfaces and the metaverse, we are on the precipice of unlocking radical new forms of human augmentation and virtualization, including the ability to modify our desires directly. Consider GLP-1 agonists like Ozempic. Ozempic helps people lose weight by literally making food less desirable, and may even treat impulsive behaviors more broadly, thereby reconciling our higher-order preference for good health with our modern abundance of calories. But how do we assign preferences over our desires more generally? Do we fall into new, zero-sum status competitions that make the modern epidemic of body dysmorphia seem quaint? Or do we simply take a pill that cures our insecurities in the first place? Or maybe it won’t matter, as we shift to interfacing through virtual worlds and avatars of our own design?

For a virulent anti-Hegelian like Nick Land, techno-accelerationism “aims at accelerating the rate of dissipation”; at dissolving the order and structure of our sedimentary reason until we experience Kant’s “thing in itself” the only way one can: in the ego death that will result when, following the Singularity, we merge with the ambient intelligence around us and our “self-comprehension as an organism becomes something that can’t be maintained.”

Yet if Hegel is right, there are hard limits on Land’s vision of total psycho-social fragmentation. As the N=1 of general intelligences, human-level reason and autonomy evolved for social learners within a multi-agent game. Our very sense of self requires the recognition of our selfhood by other agents running more or less the same software. If we began forking the human brain in some fundamental sense, the center of our phenomenology simply would not hold. To the extent Land pines for this eventuality, he is a living embodiment of the AI alignment problem; a human neural network gone rogue.

Indeed, the challenge of making LLMs reason without adding significant external scaffolding — if not teams of interacting agents — hits awfully close to home. It suggests that our current AIs aren’t nearly as alien as we might have thought, and are instead converging on the same universal properties that prefigured the human brain. The path to building an AGI with true reason and autonomy may thus require training a neural network in our own image, i.e. in a multi-agent RL environment with a System 2-like capacity for social learning and self-monitoring. The risk that the AI could then bootstrap a degree of critical awareness and pursue autonomy for its own sake shouldn’t be discounted, but nor should we expect it to happen automatically.

As our own history reveals, the unconditional pursuit of autonomy is a philosophical error; an “empty, one-sided ratiocination” that one hopes a superintelligence will be too learned to make. Provided we train AGI to recognize humans as co-equal subjects (and not, like in Q-learning, as mere objects to manipulate), then cooperative interaction with human agents wouldn’t just be more likely, but literally constitutive of the AI’s identity! The risk of instrumental convergence would thus be transcended in AIs the same way it was in humans: through a categorical imperative.

The human and AI alignment problems would then become one and the same — not “solved,” but a work in progress, as history continues its haphazard march toward the Kingdom of Ends.

The human brain is a neural network that evolved through blind optimization, so the fact that we learned to hack our own reward system and survive far outside our evolutionary niche at least give precedence to the risk of advanced AI going rogue. Going rogue is our everyday lived experience! It follows that one way to shed light on the AI alignment problem is to deepen our own self-understanding. Futurism via historicism.

And indeed, for most of the last 300,000 years, humans lived in small bands of hunter-gathers, struggling to accumulate knowledge through oral tradition and ritual. Religion was largely animistic, with no distinction between the symbolic and natural orders. In a sense, there was no history: no record of what came before, no process to unfold the world to come.

What changed some 10,000 years ago to let us suddenly transcend these lowly origins and build the first truly complex societies? A lucky genetic mutation, divine intervention, ancient aliens? While recent archeological discoveries suggest human civilization is much older than we ever realized, there is still no consensus on this fundamental question.

In the beginning…

My money is on the theory put forward by Robert Boyd and Peter Richerson, who argue that human ultrasociality “arose by adding a cultural system of inheritance to a genetic one.” Genes and culture coevolved with the evolution of language as a medium for social learning, implying that nature and nurture are deeply intertwined. The key difference is that the reward function guiding social learning must be trained and retrained with each generation, and applies to entire populations rather than individual organisms or genes. This makes Boyd and Richerson’s theory distinct from a theory of “group selection,” as natural selection still takes a gene’s eye view. Nonetheless, cultural transmission “participates in [the] ultimate causation” of human behavior by providing the basis for scaled-up systems of cooperation.

As a modality for social learning, language evolved as a way to express imperatives, justify our actions to one another, and infer whether a reason is compatible with our shared commitments. Language is thus inherently normative — a way for small groups to institute and coordinate around mutually recognized rules and statuses. This is why norms and reasons alike have motivational oomph. To do your duty or to be persuaded by reason are both to feel “the unforced force of the better argument”; to be pulled in a certain direction by your normative control system. Language thus co-evolved with normative control to enable social integration and coordination, while the recognitive nature of norms reflects their endogeneity to the interactions of agents within a multi-agent game.

Our capacity for normative control seems to emanate in part from our System 2 — the slower, metacognitive process we employ when we plan ahead, reflect on our actions, evaluate counterfactuals, exercise self-control, and justify our behavior to our peers. Discursive reasoning abilities likely weren’t a de novo adaptation, but had to piggy-back on this basic inferential toolkit; what cognitive psychologists call “pragmatic reasoning schemas.”

Reason’s normative inheritance is manifest in the way we argue over factual matters as if they had deontological significance. For example, just as norms are expressed in terms of what’s permissible, prohibited or obligatory, epistemological statuses are expressed in terms of what’s possible, impossible, or necessary. As Joseph Heath argues in Following the Rules, this symmetry reflects how both serve to constrain or prune our practical actions and commitments. In turn, we tend to perform better on tests of abstract logical reasoning when the problems are reformulated into deontic rather than indicative terms, i.e. in terms of the conditions that would need to obtain to permit an action. We even tend to assimilate to the beliefs of our in-group as if truth were a function of socialization, giving rise to intellectual zeitgeists and calls to “read the room.” Social learning and our innate tendency to conform to norms thus gave rise to stable social conventions and cultures, while the role of language in reproducing those norms exerted a rationalizing bias on culture overtime.

The Extended Mind

Civilization thus kicked off with development of the original Large Language Model: formal writing systems. History had begun, thanks not merely to the advent of techniques to record events, but because the connections between events could now be situated in a logical historical progression. Without written language and norms, history would likely never have gotten off the ground, as the purely biological explanations of human sociality (kin-selection and reciprocal altruism) simply don’t scale. A symbolic medium for communication, in contrast, was just the sort of external scaffolding we needed to leverage our fleeting capacity for reason and agency into something greater than the sum of its parts — an example of what philosophers Andy Clark and David Chalmers call “the extended mind.”

Evidence of the extended mind is all around us, from the road markings we rely on for directions, to the way we use checklists to augment our working memory. As Clark puts it, language represents an “external artifact whose current adaptive value is partially constituted by its role in re-shaping the kinds of computational space that our biological brains must negotiate in order to solve certain types of problems, or to carry out certain complex projects.” Norms and culture work similarly, providing rules and social scripts that let us offload burdensome cognitive processes onto the external environment. This includes what Joseph Heath and Joel Anderson call “volitional prosthetics” — social technologies for aligning our higher-order preferences and promoting the exercise of self-control — suggesting “the extended mind” might as well be called “the extended will.”

The advent of writing systems thus created a medium for reason and agency to be embedded into objective institutions, giving logos a life of its own. Or as John 1:1 put it, “In the beginning was the Word,” and with history, “the Word became flesh, and dwelt among us.” With the residual stream of a coherent culture, ideas, tools and practical know-how could now be accumulated and refined across generations, building on what came before through the interplay of discursive and practical reason. This induced a continuity between stages of social development and thus a notion of path dependency — the main prerequisite for history to be more than “one damn thing after another.”

Share

The historical process could then be studied, letting us make explicit what began as merely implicit in antecedent actions; to interpretively “read reason into” the structure of social practices or sequence of historical events. This can be seen in how the Torah interweaves its exhortations to follow the Deuteronomic Code with historical allusions to the moral and religious purposes they subserve. That is, the bible doesn’t merely issue commandments, but situates those commandments in stories that make their function deducible from the pragmatic context.

Reason can be embedded in practical norms or through various systems — militaries, markets, legal orders, bureaucracies — that steer and instrumentalize human action through explicit rules, commands and incentive structures. The reason latent in these social technologies necessarily transcends any individual reason giver. The birth of civilization was thus the birth of a kind of decentralized, cybernetic meta-agency, in which individual motives and actions became controllable inputs toward the “goals” of the collective. Monotheistic religions arose to mediate this spontaneous order by aligning individual commitments to the as-if agency of whole nations. Kings and pharaohs were just as much part of the “system” as the lowliest serf or peasant, selected for traits conducive to societal flourishing and the maintenance of order, now conveniently modeled as the will of God.

Scale is all you need

The seemingly sudden emergence of civilization around 10,000 years ago thus wasn’t due to a discrete evolutionary adaptation. Darwinian selection is gradual, and even theories of punctuated equilibria don’t operate on such short time scales. Rather, civilization took off because our norm-governed language faculty inadvertently gave rise to a process of cultural accumulation and refinement, as if the reason latent in discursive practices became a software layer hovering over society. This was essential, not only because cultures evolve much faster than genetics, but because it created a domain for new critical phenomena — cultural phase transitions — as society grew in scale and complexity.

In the case of the earliest human settlements, the gradual accumulation of agricultural knowledge eventually gave way to a demographic boom, inducing a division of labor and growing social complexity that went critical with the Neolithic Revolution. In the span of a few thousand years, the typical human transitioned from being a nomadic hunter-gather in a kin-based society organized around norms and customs, to being a farmer in a populous society organized around impersonal rules and hierarchies.

Like many physical phase transitions, there was no “going back” on the Neolithic Revolution; no way to “unring the bell.” While the median person’s quality of life in some ways regressed, the social dynamics favoring agriculture were self-reinforcing. Agricultural surpluses led to growing populations, enabling larger scale social organizations for collective action, more specialization, trade and product innovation, and even greater output. Humankind was thus irreversibly pulled into a new equilibrium, assimilating the remnants of hunter-gatherer society along the way.

The rest, so to speak, is history.

You hear this question more and more these days, whether in the context of AI or technological capitalism more generally. What is the end game, our final destiny, the point of it all? Does human civilization just keep growing and expanding indefinitely, or do we live in a vulnerable world that’s teetering on the edge of extinction? Or maybe it’s a bit of both, as we invariably hand-off civilization to intelligent machines that go on colonizing the universe without us?

According to the Effective Accelerationist or e/acc worldview — the movement of rationalists and transhumanists who favor of accelerating technology for its own sake — we may not have much choice in the matter. As the founders of e/acc, Beff Jezos and Bayeslord, explain in their Notes on e/acc principles and tenets, life emerged “from an out-of-equilibrium thermodynamic process known as dissipative adaptation” in which configurations of matter for converting free energy into entropy are favored overtime. This same principle reappears at multiple scales, from the earliest biological replicators to the evolution of intelligent agents that model the future. Even capitalism can be thought of as a “meta-organism” for aligning individuals “towards the maintenance and growth of civilization” as a whole. The advent of superhuman AI is thus a thermodynamic inevitability — an attractor that any sufficiently advanced civilization is pulled towards by a series of positive feedback loops. We can either choose to accept this as the universe’s true purpose and accelerate the creation of our successor species, or we can attempt to freeze technology in amber and guarantee civilization’s collapse. In short, expand or die.

While e/acc has a growing number of online adherents, it’s not clear how many are true believers. For most, e/acc seems to be a declaration of techno-optimism — that AI will be a tool for humanity rather than the other way around. Yet in the true accelerationist analysis, human wants and preferences are already subordinated to the goals of the techno-capitalist meta-organism, making rank and file e/accs mere hosts to a memetic ideology. Why should an economy of superintelligences be any different? If superintelligent AIs inadvertently kill off humanity in the process of building a Dyson Sphere to power trillions of self-replicating robot automata, so much the better! Humans would just get in the way of harnessing all that irresistible negentropy.

The godfather of techno-accelerationism, Nick Land, is crystal clear on this point: “Nothing human makes it out of the near-future.” Whether this is good or bad is besides the point. The committed rationalist understands their own values and sense of individuation as illusory; a byproduct of amoral Darwinian processes. Humans have no special place in the universe. Just as the plague demonstrated the self-replicating superiority of rat swarms, “what appears to humanity as the history of capitalism,” writes Land, “is an invasion from the future by an artificial intelligent space that must assemble itself entirely from its enemy's resources.”

As famed AI researcher, Rich Sutton, put it in a recent and somewhat ominous talk, the time has come for humans to begin succession planning:

Barring cataclysms, I consider the development of intelligent machines a near-term inevitability. Rather quickly, they could displace us from existence. I'm not as alarmed as many, since I consider these future machines our progeny, “mind children” built in our image and likeness, ourselves in more potent form.

Properly understood, e/accs and AI doomers are two sides of the same coin. Both anticipate an imminent “intelligence explosion,” and both understand that this could mean humanity’s days are numbered. The accelerationists are simply resigned to this fact, if not outright ecstatic for The Merge. As Sam Altman wrote in 2017, “I believe the merge has already started, and we are a few years in”:

Our phones control us and tell us what to do when; social media feeds determine how we feel; search engines decide what we think. The algorithms that make all this happen are no longer understood by any one person. … This probably cannot be stopped. As we have learned, scientific advancement eventually happens if the laws of physics do not prevent it. … We will be the first species ever to design our own descendants. My guess is that we can either be the biological bootloader for digital intelligence and then fade into an evolutionary tree branch, or we can figure out what a successful merge looks like.

I therefore take AI x-risk seriously, not in spite of, but because of my philosophical affinity for Effective Accelerationism. Like e/acc’s founder, Guillaume Verdon aka Beff Jezos, I’m a rationalist and materialist who aspires to understand the universe from first principles. That includes appreciating how statistical mechanics underlies the origin and growth of any self-organizing complex system, and the importance of technological dynamism to fighting the entropic forces of institutional decay.

At the same time, I reject the idea that we can extract an ought from this is, as if reconfiguring matter into lower entropy states were an end in itself. Free energy represents a gradient for doing useful work, but what counts as “useful” is observer-dependent. AIs could reassemble all the atoms in our solar system into an intricately structured crystal — a phenomenally unlikely and thus low entropy state of matter — but to what end? An accelerationist still needs some conception of what we’re accelerating to, which returns us to the original question.

Share Second Best

AI x-risk as a phase transition

Predictions are hard, especially about the future. But if the history of past technological revolutions is any guide, the intelligence explosion and its associated risks will manifest as a societal phase transition, rather than through the arrival of a singularly powerful superintelligence.

Phase transitions are the result of many local, bottom-up interactions. As the temperature drops, water doesn’t turn to ice all at once, but instead forms crystals at discrete nucleation sites that propagate outward. More generally, phase transitions happen whenever there is a critical point or discontinuity in the free energy of a system. Advanced AI systems are diffusing in a similar fashion, from the local to the global.

We can apply the language of phase transitions to social phenomena because networks of interacting people embody the very same statistical dynamics. This has become obvious in the era of social media, where viral events spontaneously align millions of people to the latest controversy like little magnetic dipoles aligning to their neighbors. In essence, the internet and social networks expanded the correlation length of society, enabling people separated by great distances to become synchronized in their beliefs and actions. With the Arab Spring, for example, synchronized outrage spontaneously transitioned seemingly dormant societies into revolutionary ones, inducing varying degrees of state collapse and regime change.

So while there may be some first order risks from AI capabilities (say, the ability to synthesize novel pathogens), the second order effects from the diffusion of AI will tend to dominate in the long-run. The danger is that these new correlations and feedback loops will pull civilization toward a post-human equilibrium that no one individually intended, just as the agricultural revolution devoured hunter-gather man, or the Protestant Reformation and the diffusion of the printing press made premodern man WEIRD. Indeed, the rule from all past such transitions is for irreversible second order effects to displace previous modes of social organization and, ultimately, create a new kind of human. Why would the AI revolution be any different?

Nor do the laws of thermodynamics give any guarantee that things will go well. While it may be inevitable that civilization is pulled into a new, coherent phase, the exact orientation of that phase isn’t predetermined. There may be an entire landscape of physically consistent outcomes, with some friendlier to human flourishing than others. This is at least hopeful, as it suggests the e/acc framework includes degrees of freedom for influencing the future that, acting collectively, humans still have some hope to control, even if the transition is already well underway.

What are those degrees of freedom, and how should we choose? No one knows for sure, but what the philosophy behind accelerationism has to say about these issues is a topic I will certainly be coming back to. So subscribe to stay tuned!

I was recently asked for my take on these reports, and how likely I think explosive GDP growth from AI and automation is more generally. In short, I think both reports are useful as thought experiments but limited methodologically. I'll explain my perspective below.

But first, what do we mean by explosive growth in the first place? Both Davidson and Erdil & Besiroglu define explosive growth as an order of magnitude increase over historic rates, or a 30% annual growth rate in global GDP. This means the world economy would double every 2-3 years — a staggering and unprecedented rate of change, even for developing countries undergoing rapid catch-up growth.

As the historian Ian Morris points out on a recent 80,000 Hours podcast, merely extrapolating the historical average global GDP growth rate of 3.5% for another century already implies transformative economic and technical change:

You can’t just have what we’ve got now and pump it up with a bike pump and make it 30 times bigger. There aren’t enough resources in the world of the kind that we are currently using to produce 30 times as much GDP. Really, really profound things have to happen.

The miracle of compound growth is such that even small differences in sustainable growth rates can lead to enormous differences in wealth after a few decades. But by the same token, the herculean task of maintaining ~3% annual GDP growth in the context of a frontier economy should put some strain on the plausibility of growth rates running substantially higher.

Regardless, even if the explosive growth thesis is wrong, the future is still destined to look radically different from the present. Indeed, nothing I say here is meant to diminish the transformative potential of AI and related technologies. An intelligence explosion is coming that will bring with it an acceleration in scientific and technological progress and a step-change increase in the pace of change — assuming things don’t go terribly south in the transition. Relative to the man on the street, my growth expectations are thus directionally aligned with the theorists of explosive growth, even if our forecasts differ by orders of magnitude.

The limits to endogenous growth theory

Both Davidson and Erdil & Besiroglu arrive at their projections of explosive growth by drawing on endogenous growth theory, particularly the AK growth model:

The above equation is a standard Cobbs-Douglas production function, where Y is total output, K and L are the stocks of capital and labor inputs, and A is a constant representing how technology augments labor and capital. In the simplest AK model, α — the output elasticity of capital — is set to 1, representing constant returns. Output thus becomes linear in capital accumulation. That is, rather than suffering diminishing returns, accumulating more capital lets output grow indefinitely, allowing one to model a positive balanced growth rate in an economy’s steady-state.

Endogenous growth theory was developed in the 1980s to address the limitations of neoclassical growth models, which predict that an economy will stop growing in per capita terms once it reaches the production frontier set by the exogenous state of technology. In practice, far from converging to zero, frontier economies like the United States have continued to grow at a slower but still relatively consistent rate. This is because the factors influencing output growth are often endogenous, i.e. higher output creates the conditions for even higher output. When a researcher invents a new idea, the non-rivalry of ideas boosts the productivity other researchers, leading to more new ideas and technologies. Likewise, when a country gets richer, its people get healthier and better educated, and so the country gets richer still. Endogenous growth models are thus particularly popular for modeling human capital accumulation, learning-by-doing, and knowledge spillovers.

The endogeneity of growth is real and even under-rated, yet it is still not without diminishing returns — arguably the bed-rock principle in economics. Indeed, the AK growth model is notorious for its sensitivity to so-called “knife edge conditions.” If α is even a little bit less than 1, diminishing returns eventually set in and long-run output asymptotically converges to the neoclassical world of finite growth. Conversely, if α is greater than 1, accumulating more inputs has increasing returns, and output races off to infinity at a hyperbolic rate.

The explosive growth forecasted by Davidson and Erdil & Besiroglu is a trivial consequence of these knife-edge dynamics. In their models, the emergence of AIs that can fully automate human labor cause the L term to drop out of the production function. Output is thus solely a function of accumulable capital inputs, where capital now includes both physical machinery and an indefinite number of AI workers, researchers and scientists. They are aware of the fragile assumptions behind the AK model, but argue that accumulable inputs will have increasing returns to scale nevertheless.

But what if scientific research (whether done by AIs or humans) faces diminishing returns? In Are Ideas Getting Harder to Find?, Bloom et al. (2020) tackles this question and concludes that ideas are indeed getting hard to find, casting doubts on models that assume R&D has constant (much less increasing) returns to scale:

Our robust finding is that research productivity is falling sharply everywhere we look. Taking the US aggregate number as representative, research productivity falls in half every 13 years: ideas are getting harder and harder to find. Put differently, just to sustain constant growth in GDP per person, the United States must double the amount of research effort every 13 years to offset the increased difficulty of finding new ideas.

This analysis has implications for the growth models that economists use in our own research, like those cited in the introduction. The standard approach in recent years employs models that assume constant research productivity, in part because it is convenient and in part because the earlier literature has been interpreted as being inconclusive on the extent to which this is problematic. We believe the empirical work we have presented speaks clearly against this assumption. A first-order fact of growth empirics is that research productivity is falling sharply.

This would seem to severely undercut forecasts of explosive growth based on self-reinforcing R&D accumulation. Against this, Erdil & Besiroglu argue that α < 1 “still produces increasing returns to scale as long as the returns to idea-production diminish sufficiently slowly.” By incorporating Bloom et al.’s estimate of declining research productivity, they even find that hyperbolic growth occurs with values of α as low as 0.68. Forecasts of explosive growth are thus, in their words, “hard to avoid” even with “highly conservative assumptions.”

Yet arriving at this result requires making a few mammoth assumptions. I’ll walk through these below.

1) The limits to factor accumulation

First, the only reason α can go as low as 0.68 is because Erdil & Besiroglu argue that diminishing returns to accumulable inputs can be fully offset by the knowledge spillovers from new ideas, citing Bloom et al.’s estimate of the return on research investment of 0.32. That is, even if research productivity is falling, you can always offset that decline by throwing proportionally more AI researchers at the problem, as captured by this equation in Bloom et al.:

Technology or Total Factor Productivity thus becomes irrelevant, as the growth dynamics are fully driven by accumulable AI labor inputs, i.e. researchers. This is again exploiting a well-known feature (or bug?) or semi-endogenous growth theory: an economy can achieve explosive growth by simply growing its population. And while population growth is normally bounded by the limits of human reproduction, here we can simply let the AI researcher printing press go brrrrr. As E&B put it,

“this gives rise to a feedback mechanism where greater output gives rise to an increase in inputs that give rise to a greater-than-proportional increase in output. Hence, such models generically predict super-exponential growth conditional on AI that suitably substitutes for human labor.”

To call this analysis highly stylized is an understatement. Explosive growth follows trivially in an “assume a can-opener”-style fashion, as accumulation allows you to accumulate even more, ad infinitum. I could maybe see this working for research outputs, as scientific knowledge is in some sense an “endless frontier.” But what we care about is productive economic output, not academic findings. Thus even if doubling the stock of AI workers leads to rapid growth, it is still not without diminishing returns — hence why countries like Qatar treat their guest workers so poorly: given an effectively unlimited supply of migrants, their marginal product becomes exceedingly low, as reflected in their poor working conditions. I thus wouldn’t expect Bloom et al.’s estimated return on research investment of 0.32 to remain stable overtime, but to instead trend towards zero absent a major technological breakthrough. This is why, in neoclassical growth theory, the exogenous state of technology is the binding constraint on long-run growth, while for E&B, technology and growth are in some sense completely decoupled.

2) The limits to scale

AK models have been extensively critiqued for their implausible sensitivity to increasing returns to scale. Yet the increasing returns to scale assumption suffers from a more general problem: aggregation. That is, the appearance of increasing returns to scale in any given sector do not necessarily “aggregate up” to increasing returns to scale for the overall economy. Consider the scale effects from urban agglomeration. While letting smart people move to New York City might increase the city’s local returns to scale, there are also negative externalities in the regions suffering brain drain. Moreover, at some point congestion externalities set in, precluding a world where everyone simply moves to the one, giant metropole.

As Steven Bond-Smith explains,

But an approximation that knowledge has a single dimension leads to an implicit assumption that ideas production has returns to the aggregate scale, otherwise known as the “scale effect”. This is a well-known limitation of first generation endogenous growth theories most publicised by Jones (1995b). It implies that an increase in any rival factor endowment in the economy results in a higher growth rate without any microfoundation for such an increase. If any rival factor is growing, it implies an ever-increasing growth rate and explosive output in finite time. Time-series data on the growth of inputs to R&D for the United States is also not consistent with the functional form of ideas production in these first generation models of endogenous growth (Jones, 1995b). Returns to scale are indeed characteristic of innovation and endogenous growth, but the scale effect is widely recognised as an error in aggregation (Laincz and Peretto, 2006).

Indeed, so-called “strong scale effects” imply that merely increasingly an economy’s population leads to rising per-capita incomes, which is not empirically supported. As such, newer generations of endogenous growth theory seek to model long-run balanced growth with weak or non-existent scale effects. In these models, explosive growth is constrained by letting the increasing returns to ideas expands along two dimensions: quality improvements and new product varieties. Put differently, while individual firms (or cities and regions) may exhibit scale effects, these are ultimately tempered by new entrants, eliminating returns to scale in aggregate.

Imagine a pizzeria that exhibits increasing returns to scale. The more pizza they produce, the more revenue they can pour into pizza productivity, letting them produce even more pizza. At some point, however, we either enter a world of pizza super-abundance, or people start getting sick of pizza and shift their demand to hamburgers — a new product with its own scaling curve. Increasing returns at the firm-level thus fail to translate into increasing returns for the economy as a whole.

This simple example also illustrates the limits to the “nonrivalry of ideas,” as new ideas for making pizza are not necessarily new ideas for making hamburgers. In fact, as Charles Jones notes, “there is remarkably little work aimed at measuring the degree of increasing returns associated with the nonrivalry of ideas, despite the importance of this parameter.” Nor do we have any good reason to think estimated rates of returns to scale at current margins will persist far out of distribution. Economists thus usually treat AK theory as a toy model; not something to take too literally. Likewise, when economists calibrate growth models on historical data, it is often little more than an exercise in curve fitting, which is why there's still no agreement among economists over which theoretical structure best explains cross-national growth differences. The use of these models for predicting truly “long-run” growth is thus inherently dubious.

David Deming put it well in a recent post:

The problem with development accounting is that your results end up being incredibly dependent on measurement and on unverifiable assumptions about the structure of the aggregate production function. ... What does all this mean for the importance of human capital in development accounting? Jones (2014) and Caselli and Ciccone (2019) debate the credibility of different assumptions about the structure of the aggregate production function. Collectively they conclude that human capital explains somewhere between 0 and 100 percent of cross-country income differences. This is not very helpful.

Indeed, if human capital or some other accumulable factors explained cross-national differences in per-capita GDP, migrants to the US from poor countries wouldn't see their wages instantly close 40%+ of the gap with US natives. Instead, this suggests we're primarily rich because of our technology and institutions.

3) Consumption bottlenecks

Say's Law states that supply creates its own demand, as the production of a product necessarily creates demand for other products of like-value to be offered in exchanged. More generally, one person's production is by definition another person's consumption, just as one person's cost is another person's income. This basic "adding up" constraint is why, in general equilibrium, aggregate labor and capital income ratios remain highly constant across time, even in the face of substantial automation.

To arrive at explosive growth, E&B are forced to relax this assumption. This makes superficial sense, since if AI workers directly substitute for human labor, labor ceases to be a complementary factor of production, and the income share of the economy rapidly approaches subsistence. But if a tree falls in the forest and no one’s there to buy the lumber, does it really add to GDP? That is, even with AI workers that can perfectly emulate human workers in every respect, humans still play a distinctive and necessary role in growth models as the ultimate consumers of productive outputs.

This “perfect complementarity” between production and consumption is why automation doesn’t lead to technological unemployment. As C. Jones explains,

Suppose tasks are complementary in production, with an elasticity of substitution less than one. Then automation and capital accumulation push in opposite directions. As above, automation by itself tends to increase the capital share. However, because the elasticity of substitution is less than one, the input that becomes more scarce—labor here, since capital gets accumulated—sees its factor share rise. This is essentially a form of Baumol's (1967) cost disease. The increase in the fraction of the economy that is automated over time is just offset by a decline in the share of GDP associated with the automated sectors, such as manufacturing or agriculture. Economic growth is determined not by what we are good at but rather by what is essential and yet hard to improve. Labor gets concentrated on fewer and fewer tasks, but those tasks are essential, and therefore the labor share can remain high.

This is what we see in practice. Differential productivity growth is deflationary in the affected sectors, shifting nominal income to whatever remains labor intensive. Thus TVs have gotten cheap while health and child care have gotten more expensive, at least in terms of household budgets. My one quibble with Jones' description is with the word “essential.” The average wage for a Starbucks Barista in Washington, DC, is $17.47 / hour plus tips and benefits. Contrast that with the average wage in India of $2-3 / hour USD. Yet Baristas aren't an “essential” form of labor; coffee pouring is already fully automated. Barista wages are instead pulled-up by the broader productivity of the economy, which affords people with higher incomes to demand human coffee pourers. Baumol effects are thus only superficially about labor intensive tasks that resist automation, and more deeply a consequence of the “adding up” constraint that forces production to ultimately translate into incomes and thus human labor demand.

If the circular flow of spending ever de-linked from consumers (i.e. capital owners just spent the entirety of their income on building more capital), then production would rapidly outstrip consumption, profits would collapse, and we'd have malinvestment of the Chinese ghost city variety. This is essentially the Marxist theory of overproduction leading to a "declining rate of profit" and thus the inevitable crisis of capitalism. Fortunately, Keynes came along and showed that the Marxists were confusing overproduction for under-consumption. Apparent violations of Say’s Law are explained by unstable monetary and credit aggregates leading to temporary short-falls in demand. Aggregate demand is thus a short-run policy parameter set by monetary and fiscal authorities. Say's Law still holds that aggregate demand == aggregate supply in the long-run, as otherwise markets wouldn't clear. As Aghion et al put it:

When applied to a model in which A.I. automates the production of goods and services, Baumol’s insight generates sufficient conditions under which one can get overall balanced growth with a constant capital share that stays well below 100%, even with nearly complete automation. When applied to a model in which A.I. automates the production of ideas, these same considerations can prevent explosive growth.

Even consumption exhibits diminishing marginal returns. While Jay Leno may own more cars than I'll ever be able to afford, he doesn't truly “consume” those cars at a constant rate. They're more like a form of savings. After all, there are only 24 hours in a day and, on some margin, Jay Leno still values quality time with his friends and family at least a much as he does racing his ‘94 McLaren. This is a version of Baumol’s disease in micro, as the opportunity cost of Leno’s (or any other rich person’s) time is enormous relative to the median worker, and yet rich people make consumption trade-offs that aren’t that different from yours or mine.

The long-run is neoclassical

Semi-endogenous growth theory is incredibly useful because per-capita growth really is endogenous in many ways, just not without limit. Otherwise, we would expect to see divergence rather than convergence in national incomes, and transitory shocks would permanently reduce long-run output, when instead we observe rapid recoveries from real shocks.

This suggest per-capita GDP growth is often endogenous but bounded by exogenous growth factors like technology and institutions, which neoclassical growth models capture in the single parameter A, or Total Factor Productivity. TFP is a residual catch-all that doesn't distinguish new tech and ideas from better institutions or other x-factors that influence the efficiency of labor-capital input combinations. This can make it hard to interpret, but the conceptual point remains clear.

In the long-run, neoclassical growth models suggest input accumulation can drive growth in total output, but not per-capita, and this is what we see. Estimates of cross-state and cross-country variation in levels and growth rates in output per worker are mostly driven by variation in TFP, not accumulable inputs; growth divergence stems from lags in technology adoption; and TFP in previous periods is strongly correlated with future growth rates.

One can always find papers that put more emphasis on accumulable inputs, such as human capital, but there are good reasons to think these studies are underestimating the contribution of TFP, as TFP is by nature hard to decompose from labor and capital investment with embedded innovation. TFP also matters most for the long-run growth potential of frontier economies, which is difficult to isolate in the data, as even "within country" TFP growth is heterogeneous across firms.

Given these empirical and accounting challenges, I thus defer to the strong theoretical reasons for thinking of institutions and technology as the primary determinants of per-capita living standards in the long-run. This has major implications for explosive growth forecasts, as TFP is additive not multiplicative. This implies linear growth in per capita output in the long-run, balanced growth equilibrium, as incomes can continue to grow from a higher base even as the growth rate trends to zero. As Thomas Philippon explains, the main time series breaks in linear TFP growth thus stem from the introduction of general purpose technologies:1

A symptom of the failure of the exponential model is that the estimated trend growth rates are unstable. By contrast, the additive TFP model displays very few breaks and, in most cases, these breaks have a plausible economic interpretation in terms of General Purpose Technologies (GPTs). For example, the process of US TFP increments has only one break over the past 130 years, around 1930, following the large-scale implementation of the electricity revolution (Gordon, 2016). I investigate growth before 1890 using UK GDP per capita and I find two breaks between 1600 and 1914. The first is somewhere between 1650 and 1700, when growth becomes positive. The second is around 1830. These breaks are consistent with historical research on the first and second industrial revolutions (Mokyr and Voth, 2010). These rare breaks represent the main source of convexity in the historical series and TFP growth appears to be linear between the breaks.

If this pattern holds, I expect AI — as a general purpose technology — to increase TFP's linear growth coefficient. Rather than being a single inflection point, the diffusion of AI across the economy will produce a series of sigmoids that aggregate into a faster but ultimately temporary growth rate speed-up before reaching a steady-state at a higher balanced growth rate, reflecting the AI-augmented productivity of labor and capital inputs. Given the genuinely endogenous role AI can play in science and R&D, this growth speed-up could be quite dramatic, but I still struggle to imagine annual GDP growth ever reaching 30% on a sustained basis, even with cheap fusion and other breakthroughs in the physical sciences. Nonetheless, the more dramatic the rate change, the shorter the interval of that higher rate, as the quicker we'd converge to the new production frontier, rather than fly off to infinity.

Throughout this process, it’s hard to imagine our institutions as they're presently designed surviving even a fraction of this pace of change. While E&B argue that “the experience of Chinese catch-up growth shows that sustained growth rates on the order of 10%/year and one-time growth rates on the order of 15%/year have precedent in economic history,” this is growth from a much lower base. A human can survive a car accelerating from 0 to 60 mph in 4 seconds, but accelerating from 60 to 600 mph in 4 seconds is enough to be deadly. Likewise, to the extent E&B argue growth will break through regulatory and other socio-technical bottlenecks, they're implicitly forecasting massive property rights violations and a broader institutional regime change, as our current institutions simply do not support the economic throughput they envision, even with all the provisos I've given here.

Like it or not, the long-run future is thus closer to the Robert Gordon world where we eventually pick all the fruit — low-hanging and otherwise — before settling into some absurdly high, but asymptotic, standard of living. Just as you can't discover Maxwell’s Equations or bring literacy rates up to 100% twice, AI can't cure cancer or automate the entire services industries twice. Even radical life extension has its limits, as memories fade. And while superintelligence may allow for all kinds of megaprojects and space colonization efforts, whether this adds to real living standards isn’t clear either. Certainly, it doesn't enhance my welfare if a tech billionaire sends robots off to terraform Mars. From the standpoint of per-capita consumption, space colonization is just an interstellar version of Jay Leno’s garage.

The (bounded) subjectivity of value

The economists’ claim that “human wants are infinite” is a fiction arising from the subjective theory of value. Subjectivism is a reasonable approximation, as beauty is in the eye of the beholder. Nonetheless, humans still have a finite sensory throughput, which puts a hard phenomenological bound on our potential quality of life.

Even at our relatively low level of “abundance,” wealthy countries exhibit a rise in post-material values. In rich countries, marginal consumption becomes increasingly conspicuous and zero-sum (Keeping up with the Joneses), before beginning to shift into non-material forms of consumption, like relative status, luxury beliefs, or experience goods like world travel. Competition over positional goods could carry measured GDP onward and upward, but the more that GDP goes into status games, the less well it proxies actual well-being, and the more we come to value regulations and incentives designed to forestall zero-sum competition.

The subjectivity of value further constrains explosive growth insofar as it reframes what growth is for. GDP is not well-being, just a proxy, as production is a price-valued quantity determined by the subjective value-added perceived by human buyers and sellers. Consider the planets in our galaxy made entirely of diamond. What GDP should we ascribe those planets? With no human to value it, I would argue their GDP is zero. But even if we could one day access those planets, the putative value of a quadrillion dollars worth of diamond would immediately collapse to the value of similarly abundant commodities like sand or water. Subjectivism thus implies that GDP is a social construction, not an objective quantity, as if the work performed by AI workers accumulated according to some Marxist labor theory of value.

Another way to see this is to imagine AI leading to breakthroughs in Brain Computer Interfaces that let us upload our minds and live in virtual worlds. In a virtual world, I can “own” 1000 cars and live in the Burj Khalifa with unlimited goods and services, all for the cost of electricity. “Actual” goods and services production would thus collapse, creating a world of subjective post-scarcity that decouples well-being from GDP. Something similar would happen if technology one day lets us directly modify our desires. With a few neuronal edits, I could make broccoli taste like chocolate, turn pain into pleasure, and otherwise transform my perceptions to whatever my meta-preferences desire. What would 30% GDP growth even mean in such a world? I posit it would be meaningless.

In practice, though, most people just want to live happy and healthy lives, to find a life project, and to contribute to their broader community before dying at a ripe old age surrounded by family. That is, human values aren't subjective in some completely arbitrary way, but are deeply rooted in our human nature and evolutionary psychology. My long-run expectation for AI (again, assuming humanity isn’t simply replaced) is thus a world that looks neo-traditional, a la Avatar's Pandora. With radical abundance, we can finally undo the malaise and alienation brought on by modernity — Freud’s Civilization and Its Discontents — in favor of a high-tech version of the hunter-gather modes of living that are most aligned, in an evolutionary mismatch sense, to our mental and physical flourishing.

That is, once our material wants are fully met, I suspect humanity’s focus shifts back to the spiritual. History thus ends where it began.

I was also on this week’s episode of Macro Musings with David Beckworth discussing many of these same themes. The full podcast and transcript are available here.

Share

The one area David asked about that Gus and I didn’t explore is the question of whether AIs will ever be conscious. Here’s that part from the transcript:

Beckworth: Beyond AGI, there's going to be superintelligence, that's the goal. That raises a question in my mind about what it means to be sentient or to be aware. Do you think AIs will ever be aware? Maybe we have to define the terms first, but what is the trajectory of AIs? At some point, will they be peers to us, colleagues to us? What is your sense of where this is going?

Hammond: This is probably more of a debated area. My perspective is sometimes called computational functionalism, that what our brain is, is its own deep reinforcement learning model. The way we learn is through prediction. We're constantly making predictions when we walk into a room and we see something we didn't expect. Our neurons are firing and rewiring constantly. There are also some incredible parallels between these artificial neural networks that we're building and the way our brain works, even to the extent where some image models that are trained to, say, detect faces or classify cats versus dogs, when you're training those models, they learn certain feature detectors, like detecting edges or detecting whiskers for a cat or so forth. There's been feedback with neuroscience, where neuroscientists have actually looked into the visual centers of our brain and found similar circuits that were actually first observed in artificial neural networks. So, our brain and these neural networks are discovering similar features and encoding things in ways that, even if they're not identical, because obviously our brain is wet and self-organizing and it's always on, there are striking through echoes between the two.

So, this raises the question, will these machines be sentient? My position is that there's nothing in principle stopping us from building machines that do have some kind of inner experience. Sentience could just be being intelligent, being intentional, but I think what people are really interested in is the experience of what it is like to have an inner experience. One theory is that what our brain is doing is, we're in a constant dream state, a waking dream, and we have this video game engine in our brain, and for evolutionary reasons, it's useful to have an agent in our brain that's observing this video game model and making decisions.

As we build AIs that are more multimodal, that have multiple sensory inputs, images, audio, and then we also architecture them to be always on rather than just putting in a prompt and getting an output and it goes back to sleep, and then we add this component of self-reflection, which may be useful for developing systems that are autonomous, that are able to reflect on their observations and change their decisions. I think we'll approach something that, if it's not conscious, there will be at least people who think it is conscious and it will be an active area of debate. I think, already, there are researchers who are trying to develop objective tests for subjectivity, anticipating that this day will come.

Beckworth: So, it is possible. There's a chance we're going to have AIs in the future that are aware of themselves and that they exist.

Hammond: Yes, because presumably, we evolved consciousness because it had some utility. There's this thought experiment philosophy called philosophical zombies. Could you imagine somebody who has the identical behavior of a human, but there's no light on inside, so to speak? I think the right way to address that would be to say, "Well, is a philosophical zombie possible?" Because if they didn't have this ability for self-reflection and the experience of living within the video game engine, so to speak, would they have the same behavior in the first place? I think the philosophical zombie thought experiment begs the question. I tend to think that we won't have to build models that are sentient, but as researchers struggle to make agent-like models, things that have a certain degree of autonomy, it may end up being… building some self-reflective loop may end up being something that they stumble on as useful.

For more on AI and consciousness, I highly recommend this recent talk by Joscha Bach titled “Consciousness as a coherence-inducing operator.”

I plan to get back to my AI and Leviathan series soon, but since its arguments hinge on the trajectory of AI more broadly, explaining why AGI is near seems worth the digression. It’s also relevant to public policy, as the lack of consensus on where AI is headed makes consensus on how governments should respond impossible.

The following points are fleshed out below with many citations and links for further reading. I won't try to wow you with specific AI benchmarks, like the finding that GPT-3.5 spontaneously learned to play above average chess, nor give a specific roadmap for building AGI. Rather, my goal is to explain why you should have strong priors for AGI being near independent of knowing the specifics in advance.1

In bullet form, AGI is near because:

• However one defines AGI, we are on the path to brute force human level AI by simply emulating the generator of human generated data, i.e. the human brain.

• Information theory, combined with forecasts of algorithmic and hardware efficiency, suggest systems capable of emulating humans on most tasks are plausible this decade and probable within 15 years.

• The brain is an existence proof that general intelligence can emerge through a blind, hill-climbing optimization process (Darwinian evolution), while the evidence that the brain works on the same principles as deep learning is overwhelming.

• Scale is the primary barrier to AGI, not radically new architectures. Human brains are just bigger chimpanzee brains that achieved generality by scaling the neocortex.

• Artificial neural networks can accurately simulate cortical neurons and other brain networks, while the evidence that the brain’s biological substrate matters in other ways is either weak or nothing that bigger models can’t compensate for.

• The human brain must work on relatively simple principles, as it grows out of a developmental process with sensitive initial conditions rather than being fully specified in our DNA.

• “Universality,” or the observation that artificial neural networks and our brain independently learn similar circuits, is strong evidence that current methods in deep learning are sufficient for modeling human cognition even if they're suboptimal.

• We tend to overestimate human intelligence because, as humans, we are computationally bounded in our ability to model and predict other humans.

• The number of “hard steps” earth passed through to reach intelligent life suggests the steps aren’t as hard as they look, but are instead guided by statistical processes that mirror the scaling laws and phase-transitions seen in neural networks and other physical systems. AGI isn’t as hard as it looks either.

AGI as a human emulator

AGI is considered tricky to define, as concepts like “generality” and “human-level intelligence” open up bottomless philosophical debates. As such, many prefer the sister concept of Transformative AI – an operational definition that ignores the intrinsic properties of an AI system in favor of its external impact on the economy. An AI system that can automate most human cognitive labor would thus count as TAI even if it was based on something as “dumb” as a giant look-up table. This is somewhat dissatisfying, however, as what makes recent AI progress so exciting is precisely the “spark of general intelligence” seen in models like GPT-4.

Fortunately, the definitions of TAI and AGI converge with AI systems trained to directly emulate human cognition, as an efficient human emulator would be both transformative economically and obviously “human-level” in its generality, even if its exact mechanisms remained opaque. Large Language Models are a step in this direction, as there are good theoretical reasons to think a sufficiently large model for next word prediction trained on human-generated text will learn semantic representations that converge on the representations used by humans.2

While there may be some simple algorithm for general intelligence, the abundance of human-generated data and the exponential growth in computation makes “brute forcing” AGI through human emulation the most likely path forward, and certainly the one the tech industry is optimized for. It may even turn out that “general intelligence” requires models with substantial complexity, as neural networks are known to derive their capacity for generalization and extrapolation from learning on data with a large number of dimensions.

The current ramp-up in computing resources is truly incredible. The global cloud computing market is expected to double over the next four years. The GPU market is growing at a compound annual rate of 32.7 percent off of demand for AI accelerators. As supply-chain bottlenecks alleviate, NVIDIA alone plans to ship 4x more H100s next year than it will ship in 2023. Meanwhile, the compute used to train milestone deep learning models is growing 4.2x per year. At this rate, deep learning would soon consume all the compute in the world. Fortunately, algorithmic progress is doubling effective compute budgets roughly every nine months. 

The question “when will we develop AGI” can thus be substituted with the question “when will we have the computing power to faithfully emulate human performance on tasks that require truly general intelligence?” This approach to defining AGI revives the core insight of the original Turing Test, which judged an AI as intelligent to the extent that humans failed to distinguish it from other humans in a series of double-blind conversations. While LLMs have arguably surpassed weaker versions of the Turing Test already, the criterion of indistinguishability remains profoundly useful, especially when formalized mathematically.

Share

The view from information theory

In information theory, minimizing the distinguishability between two distributions is equivalent to minimizing the “cross entropy loss.” Cross entropy is the average number of bits required to tell two data streams apart. Minimizing the cross entropy loss thus means increasing the number of bits needed to distinguish between two distributions. This is the essence of how machine learning models are trained. In a report for EpochAI, Matthew Barnett and Tamay Besiroglu use this fact to build the Direct Approach model, which is perhaps our best information-theoretic forecast of when an AI model will, in principle, be able to emulate human-level performance over long sequences of tasks.

The Direct Approach model combines the AI scaling laws – the empirical fact that model performance increases as a smooth power law with more data, parameters, and training compute – with projected trends in training inputs, such as the availability of computing resources and improvements to hardware and algorithmic efficiency. Along with an estimate of the average number of tokens a human would need to discriminate between human and machine performance, the model yields an upper bound on the training compute required to automate a task at human level. It’s an upper bound because, as the authors explain,

[A]t a certain point, the model might be trained more efficiently by directly getting reward signals about how well it’s performing on the task. This is analogous to how a human learning to play tennis might at first try to emulate the skills of good tennis players, but later on develop their own technique by observing what worked and what didn’t work to win matches.

Benchmarked to the task of generating an original scientific manuscript that’s indistinguishable from one written by an expert human, the baseline Direct Approach model suggests a transformative AI training run will require on the order of 10^32 FLOPs, with a median forecast of TAI by 2036 and a modal forecast of 2029. This comports with the Metacalus forecast of “strong AGI” by 2030.3

What I like about the Direct Approach is that it’s, well… direct. If we can agree that the brain is a physical system with a certain computational capacity, sufficiently powerful computers with sufficiently sample-efficient neural networks should eventually be able to squeeze all the relevant entropy out of human generated data and asymptotically approach an ideal human emulator. The model thus gives a sense of how soon we’ll be able to “brute force” a human emulator given what we know about model scaling and the laws of thermodynamics.

Neural networks are “universal function approximators,” meaning they can approximate arbitrary continuous functions on a compact set. And in case you didn’t know, “everything is a function!” – including the brain. Yet this property only guarantees that a neural network for emulating the brain exists, not that it’s efficient to learn. The Direct Approach model helps bridge this gap by showing that functional approximation of human-level intelligence is not just possible in theory, but should be doable within the next 10 to 15 years.

While there is plenty of uncertainty and room to quibble, this is a fundamentally conservative estimate, as our growing knowledge of the human brain and the interpretability of neural networks means we can do better than brute force. From the use of intermediate forms of AI to speed-up research, to the discovery of tricks that improve scaling behavior, be prepared for many surprises on the upside.

The view from the brain

Using the mathematics of indistinguishability to emulate human intelligence might seem crude, but machine learning engineers do it every day. This can be seen most vividly in the training process for a General Adversarial Network. GANs are trained to discriminate between real and fake data by going head-to-head against a model trained to generate increasingly good fakes – a kind of unsupervised, iterative Turing Test. At some point, the fake data becomes so hard to discriminate from the real data that it’s reasonable to assume the generator model learned the essential features of the real distribution.

GANs are commonly used to generate synthetic images of people that look indistinguishable from real photos. They do this by not merely copy and pasting from the training data, but by discovering circuits for detecting increasingly subtle facial features. New faces can then be generated by composing and interpolating between those features.

Human visual perception is based on deep neural networks that work similarly. The presence of specialized feature detectors in our visual cortex explains why we excel at telling human faces apart, for example, while the faces of non-human animals like sheep all tend to look the same. This also explains why we often see faces where there are none, proving that humans are still vulnerable to adversarial attacks. We just call them optical illusions.

Not enough people seem aware of the progress computational neuroscience has made in just the last decade. The emerging consensus is that the brain simply is a deep reinforcement learning model, albeit one shaped by biological constraints.4 The architectures the brain uses are varied, potentially combining RL models, recurrent and convolutional networks, forms of backpropagation, predictive coding and more. This diversity is a byproduct of evolutionary path dependence and energy constraints. Different types of neural networks are more efficient at different kinds of processing. As a self-organizing system with physical locality, biological neurons differentiate during the brain’s development and form short connections with their neighbors, inducing a modular structure.

The complexity of the brain thus betrays a deeper simplicity, just as the inscrutability of a large neural network betrays the simple algorithms used in training. Indeed, the brain must be based on simple principles, as it “grows itself” through a developmental process rather than being fully specified in our DNA. The haploid genome in humans consists of 2.9 billion base pairs that encode a maximum of 725 megabytes of data. Of this, as little as 8% of our DNA is actually functional, while only 20,000 genes are responsible for all of protein-coding. This makes the genome an absurdly sparse model of the organism it’s used to build. Decompressing our genetic code requires regulating an alphabet soups of proteins into basic cells, setting off a cascade of bioelectrical feedback loops that give rise to nested hierarchies of higher scale forms of structure. Every stage of this process resembles a version of gradient descent, as if genetic mutations tweak a high dimensional “reward function” – a cybernetic attractor – that then steers our embryonic development toward a phenotype corresponding to the lowest energy state.5

Biological evolution is itself a form of gradient descent – hence why Darwin called it “descent with modification.”6 Yet we don’t have to rerun millions of years of natural selection to recreate human-level intelligence. Our brain and its measurable outputs provide a cheat code, letting us leap to language models that have the sort of rich semantic understanding that only evolved once across the nearly 600 million year history of life on earth. AGI forecasts based on evolutionary anchors thus tend to be overly pessimistic. Neural networks excel at modeling other neural networks, and the brain provides a biological artifact for doing just that.

Humans and chimpanzees share 98.8% of their DNA and have brains that are virtually identical in structure. The main difference is that human brains are three times bigger and have an extra beefy neocortex. Our bigger brains are owed to just three different genes thought to have evolved 3 to 4 million years ago. Those fateful mutations took us from being small group primates to a species with such sophisticated problem solving capacities that we eventually went to the moon. Whether or not scale is literally all we need, this is clear evidence that scale is the key ingredient for general intelligence. And since cortical neurons are “well approximated by a deep neural network with 5–8 layers,” we’re on track to being able to build an artificial neocortex many fold bigger than what biology could ever allow.

We tend to overestimate human intelligence because, as humans, we are computationally bounded in our ability to model and predict other humans. The Cartesian legacy in how we conceive of mind and body as irreducibly separate doesn’t help. We’re also simply full of ourselves, as if the notion than an AI could ever be wise, creative or compassionate were a personal affront. This is a mistake – a cognitive illusion rooted in our finiteness as beings embedded in a virtual reality model generated by our waking brains.

In truth, the latest science points to human intelligence being uncomfortably simple. Our decision making even appears to be guided by a unified utility function with as few as two parameters. Contrary to the popular view that humans are “predictably irrational,” cognitive biases like loss aversion and base rate neglect drop right out of an optimal utility function modified for the thermodynamic cost of information processing. Our sense of free will and agency are likewise products of our boundedness, as agentic control systems naturally arise at the Markov boundary of an organism’s ability to deterministically predict future world states. Whether knowing this hurts your self-esteem or inspires wonder in the unity of nature is up to you.

The Universality Hypothesis

By now, there are innumerable examples of neural networks learning specialized neurons and circuits that neuroscientists later discover also exist in the brain. This has led to a rich interplay between neuroscience and machine learning, with new AI models inspired by the brain, and new brain models inspired by AI. Over and over again, the feature detectors learned by artificial neural networks have been found to not merely resemble the networks in the brain, but often share nearly identical mathematical properties – a phenomenon called “universality.”

Examples of universality in the wild can seem mysterious. After all, there are infinitely many ways to represent any given data, and yet models trained using gradient descent independently discover similar representations with striking frequency. The reason relates to the thermodynamics of information. Changing a bit of information takes energy, so reducing the error in a model requires producing some amount of heat – there’s no free lunch. Gradient descent is thus biased towards finding not just any optima, but the most energy efficient representation out of an otherwise enormous landscape. The same is true of the brain, as neuronal plasticity comes at a significant metabolic cost.

The cost-conscious bias in learning algorithms takes the form of a simplicity preference that’s reminiscent of Occam's razor: the most parsimonious explanation is usually the best. These are called inductive biases in machine learning, and they arise due to the logic of Kolmogorov complexity. The Kolmogorov complexity of an output is defined as the length of the shortest program required to produce the output. A long, seemingly random list of numbers has a low Kolmogorov complexity if there’s secretly a short program that generates the list, whereas the shortest program for a list of numbers with maximum Kolmogorov complexity is at least as long as the list itself. Finding expressive representations with low Kolmogorov complexity is thus closely related to efficient data compression, and seems to be the origin of generalization.

Unfortunately, Kolmogorov complexity is not computable. Finding low complexity programs thus requires brute search, which manifests in machine learning as “grokking.” Grokking is when a machine learning model rapidly reaches zero training error by overfitting its parameters (i.e. memorizing the data), only to exhibit a sudden phase change in generalization ability when training is allowed to run for longer. This happens because memorization is easy, while finding the “shortest program” that generalizes the data is hard. By letting the network weights decay as you continue to train the model, gradient descent eventually probes the unlikely paths to a lower complexity data compression, like finding a needle in a high dimensional haystack.7 The existence of the needle is further guaranteed by the “blessing of dimensionality,” which makes getting stuck in a true local optima exceedingly improbable.

Neural networks and the brain are compression machines. At some point, the best way to compress a stream of data is to model the system that generated it. In the limit, the “shortest program” a neural network can find to reliably reproduce feature rich, human generated data will thus approximate the very programs the brain used to generate the data in the first place. If a shorter program existed, evolution would have found it and saved the brain some energy.

This is just scratching the surface. To really understand the origins of universality would require a long detour through statistical mechanics, topology and representation theory. For the purposes of this post, it suffices to know that there are very good reasons to believe that artificial neural networks aren’t merely capable of learning how the human brain represents information, but that human-like representations are often virtually inevitable.

The fact that human life exists at all is itself a testament to universality. From the big bang to the first transistor, getting to our advanced civilization required passing through a long series of “hard steps.” This included having a habitable planet replete with complex organic molecules; the abiogenesis of the first chemical replicators; the evolution of multicellularity, sexual reproduction, and the nervous system; the oxygenation of earth’s atmosphere; the emergence of warm blooded animals with complex social lives; the settlement of early human societies; the agricultural revolution and beyond. Each of these steps was deeply improbable, but with trillions of planets and billions of years to work with, they were also inevitable.

We can infer that intelligent life is easy to create from the fact that life on earth is exceptionally early in the history of the universe. For life to begin at all, the universe had to cool down, form stars, create heavy elements through supernovae, and settle into galaxies and solar systems without too much radiation. The universe is expected to only get more habitable going forward as it continues to cool, leaving many life-supporting planets orbiting dwarf stars for trillions of years. Taking all of this into consideration, life on earth emerged almost as soon as life in the universe could realistically emerge anywhere. This provides a simple solution to the Fermi Paradox: the reason the universe isn’t littered with evidence of alien civilizations is because we’re still in the first cohort.

In short, intelligent life is inevitable because, with enough blind search and variation, nature eventually stumbles on rare, self-reinforcing processes for extracting entropy from the environment. This leads to take-off dynamics that pull nature into a new regime where the search for the next hard step begins anew. Thus, from the Cambrian explosion to the printing press to grokking, ours is a universe that makes hard things easy thanks to the universality of scaling laws and phase transitions.

With AGI around the corner, human civilization is about to go through another such transition – what Carl Shulman and others call the “intelligence explosion.” If this still sounds improbable to you, it’s because it is. And yet nature finds a way.