How to profit off AI

The economics of commoditized intelligence

If my last post on AI was a bit of a downer, it wasn’t meant to be. Near-term AI will be disruptive to many industries and organizations, but it will also create enormous social benefits. Institutional disruption may itself be one of AI’s greatest blessings — at least if it’s managed well. As in other cases of creative-destruction, commoditized intelligence will radically change the economy’s cost structure, producing the institutional churn needed to bust monopolies, dissipate economic rents, expand wealth and opportunity, and right-size institutions.

The question I keep getting is how someone can best profit off these trends. I don’t have a crystal ball, but I can share some general ideas, as well as a framework for thinking about AI’s impact more generally.

Buy and hold index funds

For most people, the best investment advice for capturing the upside of AI is evergreen: invest in index funds. If AI follows the pattern of previous hype cycles, we’re in for an asset bubble and billions of dollars in investment flowing into new products and companies. Index funds track the overall market and are the easiest way to earn strong returns on your savings. The efficient market hypothesis suggests actively trading stocks or picking individual investments is a fools errand, and it will only get more foolish as AI-augmented algorithms drain whatever alpha is still left.

AI’s impact on specific asset prices is simply hard to predict. You’d think the present value of expected future cash flows for a company like Shutterstock would be cratering given the rise of open source models for generating royalty free images, and yet SSTK 0.00%↑ (while far down from its peak) is still worth more than it was pre-pandemic. More generally, it’s not obvious to what extent the social surplus from AI will be capitalized into asset prices or arrive in the form of intangible consumer welfare gains and rapid price deflation. Either way, some set of companies will surely go to the moon, so it will be hard not to benefit with a sufficiently diversified portfolio.

Own scarce factors like land and commodities

In a world where AI is driving rapid productivity improvements in labor and capital, nominal spending will flow to the scarcest factors of production. The simplest models of the economy contain just three factors: land, labor and capital. Land is really a stand-in for natural resources and other endowments. In practice, that means energy and key commodities, including the metals used in batteries and semiconductors.

Many of the productivity gains from AI will also be captured in land rents, so it will pay to own some real estate. It’s hard to know which housing markets will rise or fall, but so far the information age has been long for cities. As AI enters the network edge through IoT devices and the like, cities may become even more attractive due to the variety of intelligent public goods that will only be economical in areas with highspeed internet and large agglomerations of people. Land also has the feature of being intrinsically scarce, as not everyone can live in the posh part of town. If they did, it wouldn’t be posh anymore.

At the same time, some real estate markets may crash. If AI radically disrupts public education, for example, families will be less eager to pay a premium for property zoned in a top performing school district.

Bet on service innovation

AI will no doubt have a large impact on tradable goods, from medicine to robotics. But in the nearest term, the service sector seems ripe for a productivity boom. Services are the largest sector of the economy and, until recently, have been relatively impervious to process innovation. After all, Oxford professors have been lecturing in roughly the same format for over 800 years. I’d therefore expect near-term AI to begin picking away at any sector afflicted by Baumol’s cost disease, as the potential returns are enormous. This will range from low paying customer service and call center jobs to many high paying knowledge and managerial professions too.

For any given service job, there are three structural questions to keep in mind:

• is AI a substitute or complement to human labor?

• can the job be disaggregated into distinct tasks?

• and do the markets for inferior and superior versions of the service interact?

Take ATMs, or Automatic Teller Machines. Rather than replace human tellers, the diffusion of ATMs spurred banks to open new branches, causing the number of human tellers to actually increase. ATMs turned out to be a complement to human tellers, taking over a menial task while humans found comparative advantages in other forms of banking services, such as helping you apply for a credit card. I suspect AI and online banking will substitute for tellers going forward, but that won’t be true in every case. AI could very well grow the total number of lawyers, for instance, or keep their numbers roughly constant while expanding access to legal services many fold.

At the same time, it’s important to inspect the quality structure of the market to detect whether current business models contain any implicit cross subsidization. The human translation market has been decimated by AI translation tools, for example, not because AI is currently better at translating Homer than experts in Ancient Greek, but because the entire industry was sustained by a long-tail of basic translation needs for which Google Translate suffices.

Lauren Wilford @lauren_wilford

a YouTube comment about the way AI destroyed the field of translation years ago

Pay attention to education

Per the Oxford example, education seems especially exposed to disruption by near-term large language models, starting at the youngest ages. Parents are told to read and talk to their kids from infancy, helping prime their language faculty and build some basic vocabulary. So why not have a plush chatbot teddy bear that converses with the child 24-7? It could even teach the infant a second language through repetition and immersion, becoming their very best amigo.

For older kids, mastery based learning is one of the most effective interventions in the education literature. The only constraint is human labor. So why not create an adaptive AI tutor? Soon, a precocious homeschooler will bootup their web browser and meet an animated Ms. Frizzle, ready to pick up where they left off. She’ll help break down problems, keep the student motivated, and summon visual aids to a virtual blackboard. The language model could be fed a variety of metrics, from the timbre of the student’s voice to their latest test scores, optimizing its approach to maximize learning outcomes. “I see you got a B on the grammar test. Let’s go back over verbs and adverbs one last time.”

Will we even need teachers anymore? And what about the human touch? To which I say: Of course, the human touch is important, but why does it need to come from someone with seven Masters’ degrees? Why not their grandma or personal trainer or a 20 year old who’s good with kids? Since tutoring can be done anywhere, education could even unbundle and rebundle around some other setting, from a religious community to a chess club.

Interventions of this kind have already seen significant success in the developing country context. One RCT in India found 90 minutes of study a day using an adaptive teaching program called Mindspark raised student test scores by 0.37 sigma in math and 0.23 sigma in Hindi over just a 4.5-month period. The relative gains were greatest for the academically-weaker students. And this is using a relatively dumb application, albeit based on a corpus of high quality education materials. The GPT-4 version will be lit.

Higher education may face the hardest reckoning of all. Existing trends already indicate many small liberal arts colleges will soon be up for auction, while university administrators are overdue for some austerity of their own. AI will be a boon for researchers and academic scientists, but it will also deflate the caché of performative degrees that only exist to flex your wordcel muscles. Pursuing an MFA or art history major is largely a leisure class activity in the first place; a way to signal your status. But if AIs can make anyone a good writer, music producer, or “creative,” then many existing sources of status distinction will likely whither away. Indeed, while ChatGPT doesn’t yet pass the Turing Test, it easily drafts an essay on all the ways AI reinforces the patriarchy. The Sokal hoax literally writes itself.

Start a company

In the medium term, the single best way to capitalize on trends in AI is to start a company. The barriers to forming a profitable start-up are already falling fast. Take Lensa, the viral app for turning your selfies into cool avatars. It’s mostly just implementing off-the-shelf image generation techniques, and yet has likely pulled in millions in revenue in just the past week alone. Or consider this web app for generating news articles based on a single Tweet:

Erik @0xfdah

Here's my @Replit prototype: the-gpt-times.eriks-studio.repl.co The GPT Times: All the news that's fit to generate 🗞️

Balaji @balajis

Here is the interface we want to see. 1) Prompt with a few tweets 2) Output an article in the style of NYT, WSJ, etc 3) Make the aesthetic look exactly the same 4) Compare the speed and cost of having an AI do it vs a traditional NPC Can set this up as a @replit bounty…

To top it off, Erik, the creator, is only 85 days into learning Python! The app uses GPT to confabulate news articles in the style of the New York Times using a Tweet as the prompt. The articles are some combination of truth and hallucinated fact, but you can see where this is headed. By day 100, Erik will easily be on the way to creating the next Buzzfeed. Imagine a newsroom where reporters submit bullet points and quotations into a simple CMS that generates the full article for them, all in house style, and with several variants for the final editor to choose from. Feel free toggle to a neutral voice if your prefer. The scourge of Russell Conjugation will soon be over.

The workflow of a startup founder or software engineer is changing in real time. While no one should expect unemployed coal miners to “learn to code,” almost anyone can “learn to prompt.” The job of an app developer thus looks increasingly like that of a supervising manager, curating the outputs of an army of AI script kiddies while giving constructive feedback here and there — only these junior employees won’t conspire against you on Slack.

The transition problem

The societal benefits from AI are potentially enormous. Barriers to entry will fall, innate inequities will close, and economic and scientific productivity will accelerate. All for the good.

My concern is primarily about the transition period. Take the fallout from internet and mobile. Over the last two decades, ridesharing disrupted taxi monopolies, online bookstores disrupted publishing, downloads and streaming disrupted the film and record industries, and blogs and social media disrupted the news media. In every case, the disruption was relatively sudden but far from total, as incumbents either adapted, fought back or both. Taxi drivers eventually got apps of their own, but also staged violent protests. Publishers moved into ecommerce, but also went to war against Amazon. Hollywood and major record labels came around to digital streaming, but only after many failed copyright battles. And the news media downsized and consolidated, but only while fomenting a coordinated “techlash” against their new platform competitors.

The common factor behind all these recent cases of disruption was a technology-induced shift in transaction costs. Transaction cost is a term of art in institutional economics; not merely the literal costs of processing a transaction. Rather, the concept embodies all the costs associated with a market exchange, including opportunity costs, travel costs, agency costs, language barriers, and more.

As Ronald Coase argued, transaction costs are why we have companies and governments at all, as islands of central planning in the free market sea. When the costs from outsourcing are prohibitive, a production process will be brought in-house, enabling centralized management, team production and superior monitoring. In turn, organizations become a nexus for culture building and rule enforcement. If one day all transaction costs went to zero, we would find ourselves all working as individual S Corps, rendering company cultures and a great deal of tax and regulatory policy obsolete. But transaction costs never go to zero. Instead, as transaction costs fall in one domain, new costs emerge to form the basis for institutional reorganization.

The transaction costs associated with hailing a taxi, for example, include the effort of having to call the dispatcher, the anxiety of not knowing if or when your car will arrive, the potential miscommunications over which directions to take, the risk of being scammed, and hassle of haggling over the tip. Mobile phones enabled ridesharing apps to “economize” on all those transaction costs while also achieving efficiencies of scale and coordination, resulting in an overall better experience. Yet in so doing, mobile phones also indirectly changed the structure of the market, leading to new governance models with distinct competitive dynamics. When self-driving cars arrive, the dynamics will change once again, as the “two-side platform” structure of Uber and Lyft gives way to transportation-as-a-service.

Ceteris imparibus

As I expressed in Before the flood, my main worry with AI is that people are too focused on the technology in isolation rather than seeing it as a platform, like electricity or the mobile phone, that will have much broader ramifications. To use some more econ jargon, humans tend to think in terms of the comparative statics, in which one thing changes while everything is held constant, rather than in terms of general equilibrium. It’s a useful heuristic most of the time, but not for general purpose technologies, just like you would never analyze the emergence of the mobile phone solely through the lens of its impact on land-lines.

Even sophisticated AI safety discussions have a tendency for ceteris paribus thinking. Circa 2000, no one was having “internet safety” discussions about how, in just over a decade, the technology would shed unprecedented light on political corruption while creating new tools for mass mobilization, leading to the Arab Spring uprisings and a crisis of authority across Western democracies. At the time, some even argued the internet was “a passing fad.” Likewise, AI safety discussions tend to focus on the static aspects of the technology itself: Is it biased? Does it make stuff up? Is it well aligned? What about plagiarism and artist credit? Will it turn us all into paper clips?

It’s not that those aren’t important issues. It’s that they miss the forest for the trees. As Balaji said, the die is already cast. AI will do much more than make homework essay assignments obsolete. AI will change what it means to get an education in the first place. Meanwhile, as with the internet, it’s important to not just solve for the final equilibrium, but to also consider the indirect effects that a rapidly changing cost structure will induce across every organization.

Democratizing intelligence

My basic model for AI as a technology is that it’s limiting to a universal translator, a best possible approximator for extracting signal from noise, and a functional map from any input A to any output B. That can be applied to language, creating low cost translations that drive human translators out of business. But it can also be applied to anything else, from diagnosing Parkinson’s disease, to predicting someone’s gender from a retina image, to IDing someone by their unique gait. Everything and everyone is constantly shedding information all of the time, but due to entropy we interpret most of it as useless heat. Yet if there’s signal to be extracted, deep learning reinforcement models will find it, conditional on cost. Paradoxically, the same technology that gives rise to deep fakes will make other aspects of the world radically less opaque.

Our 20th institutions were designed around a world of medium-high transaction costs. Across several centuries, technological change lowered the costs associated with consolidating governments into centralized nation states. 19th and 20th century governments economized on transaction costs by creating general laws, standards and risk pooling arrangements that displaced the ad hoc particularism of city states, guilds, and mutual societies. Industrialization took this a step further, enabling mass production, assembly lines, centralized bargaining, mass communications, and continental transportation networks. Yet overtime, as transaction costs continued to fall, we began to turn the curve on the High Modernist era. Once great empires decolonized, creating dozens of newly independent nations. Labor unions declined in line with outsourcing and market fragmentation. Talk radio arose as a challenger to the mainstream media, and then the internet unbundled broadcast rights from FCC licenses altogether.

The “bundle, unbundle, rebundle” cycle seen in many technologies reflects these dynamics in micro. Technology rarely reduces transaction costs across the board. Rather, falling transaction costs in one area create the conditions for new organizations to form and internalize the transaction costs associated with the new paradigm. Stripe Atlas, for instance, radically reduces the transaction costs associated with incorporating a new business, allowing a thousand flowers to bloom, but only by economizing on the smaller transaction cost of organizing a talented but centralized team of product engineers.

Forecasting the near-term impact of AI thus requires a theory of which transaction costs will fall and what other transaction costs will rise as the traction for institutional reformation. Essay homework assignments, for instance, are premised on the transaction costs of fabricating an essay being prohibitively high. The fact that ChatGPT can now write your term paper implies that the transaction costs of generating a passable essay have fallen to near zero while the transaction costs of validating an essay’s source have exploded. Evaluation methods will adapt accordingly, if not the structure of higher education as whole.

In the The Art of Not Being Governed, James C. Scott documents the anarchic societies in the Zomia highlands of Southeast Asia. The mountainous geography makes the region intrinsically resistant to nation building, i.e. “barbaric by design.” It’s the sort of society you get when the transaction costs of forming state capacity vastly exceed the transaction costs of local tribes banding together in resistance to distant rulers, not unlike in Afghanistan or, for that matter, Appalachia.

The open question is whether access to advanced AI, located at the network edge, makes society look more or less like Zomia. Do cryptography and generative AI and other technologies for local resistance erect invisible mountains that make society less and less legible? Does instant customization obviate the need for rationalistic laws and standards setting bodies created to economize on transaction costs that no longer exist? Does superior risk and productivity modeling cause insurance arrangements to unwind and wage inequality to balloon? Will the state win the arms race against the public and install unprecedented forms of surveillance to secure public order? Or does the ability to map any A to any B, like from natural language instructions to some well-written code, create the conditions for a true, radical diversity that’s just a function map away from mutual understanding?

I don’t know the answer to these questions, but consider me long volatility.