How Do Humans Think?

Mar 2, 2023

My various feeds have been filled with a lot of discussions of ChatGPT, its value, its use cases, whether it’s good or bad, how it works, and all sorts of other stuff. Some of it is interesting or useful, some of it is just noise, but mostly I don’t want to talk about any of it.

But there’s one thing I’ve seen repeated in a few places that I find really interesting, which is the claim that, because of the way ChatGPT is built, it is impossible that it could be thinking. And I think this is wrong, or not so much wrong as not yet known. Because we don’t actually know how humans think.

(To be clear, I’m not saying that ChatGPT itself is thinking, I think that’s clearly not true. I’m talking about the idea that it’s impossible for any LLM, of any size or construction, to think, based solely on the fact that it’s an LLM.)

Even if we only look at language processing (which is the only thing ChatGPT can do), the whole human thinking process is a complete black box. I’m currently writing this sentence, and I have no idea what mechanism makes me decide to use one word over another. Nobody does. At the point where I had written “…and I have no idea”, how did my brain produce the words “what mechanism”? Why didn’t it produce “how” or “of the process” or any of hundreds of other alternatives? I dunno! Has my brain encoded complicated word frequency associations that I’ve learned over a lifetime of reading and speaking, and then did it use that model of the English language to produce the next word it should use? Maybe! Maybe not!

When I struggle to come up with a particular word while writing, as I have done a few times while writing this, what is actually happening? Why can my language memory not produce a word one moment, but then can sometimes produce it after a bit of thinking? When I concentrate hard on trying to come up with exactly the word I mean, and not one of its synonyms or closely related words, what is my brain actually doing in that moment? I have no idea, and neither does anyone else.

Our brain is just made up of neurons, and an individual neuron is reasonably well understood (though not completely). But the brain has like 100 billion neurons and something like a quadrillion synapses, and at that scale, we get human thought, but we don’t know why.

This makes me really skeptical about any claim that a particular type of thing can’t possibly think. Because I think if you described a neuron and then said you wired a bunch of them together, it’d be pretty tempting to say the same thing about that. But that’s just us, and we seem to think just fine.