Thinking out loud, or talking in your head?

The language of thought is language

ChatGPT is surprisingly good at carrying out complex tasks if you first ask it to “think step by step” and to reflect on its subsequent actions. This is the basic idea behind AutoGPT — a project built on GPT-4 that bootstraps the language model into a kind of autonomous agent. It’s generating some vigorous discussion within the AI alignment community, and for good reason. But to me, the most intriguing thing about the power of chain-of-thought prompting is what it says about the nature of human thought.

Not everyone has an inner monologue, but for the majority of us who do, it’s tempting to conflate the voice in our head with the language of thought itself. This is the essential mistake behind the Cartesian influence on analytical philosophy, leading to confused debates over how mental states could possibly have propositional content.

If you’ve read my piece on LLMs and Wittgenstein, you’ll know I reject this point of view. Instead, I think propositions gain their content from how they’re used in a language game. That irreducible intersubjectivity — the shared understanding of the rules of language-use as embedded in the norms of linguistic practice — is what underlies Wittgenstein’s argument on the impossibility of a private language. Yet what could be more private than the language of our inner thoughts?

The pragmatist response is that, rather than view talking as “thinking out loud,” we should instead view thinking as “talking in our heads.” That is, language starts out as “externalist” — a way of converting the ineffable thoughts that bubble out of our unconscious into a symbolic medium for external communication — and only later do we “internalize” language as cognitive scaffolding.

This is why symbolic approaches to AI were always doomed to fail. Symbols are just an external output; a serial code for sending and receiving thoughts that arise out of inscrutably parallel neural processes. Looking for symbolic architectures in the brain itself is akin to looking for desktop icons in a computer’s binary code: it confuses the external interface for the internal representation.

It’s fascinating that LLMs seem to mirror humans in this respect. They seem to become better reasoners when they “talk to themselves” and use their serial, symbolic outputs as a kind of scaffolding for subsequent thoughts. It makes you think! Maybe dual process theory is wrong; the so-called “thinking fast, thinking slow” model of human cognition. Instead of looking for a dedicated module in the brain for slow, analytical thinking, perhaps our “System 2” was just internalized language all along.

There seems to be evidence in this direction. When neuroscientists use fMRIs to look for the neural correlates of syllogistic reasoning, the inferior frontal cortex lights up. This is also the part of the brain involved in speech recognition and language processing. That’s why when you listen to your inner monologue you literal hear the words, and may even find yourself mouthing along to them.

So maybe we should stop saying inner-thoughts and start saying “inner-speech.”