> And it’s part of the lore of neural nets that—in some sense—so long as the setup one has is “roughly right” it’s usually possible to home in on details just by doing sufficient training, without ever really needing to “understand at an engineering level” quite how the neural net has ended up configuring itself.

> What determines this structure? Ultimately it’s presumably some “neural net encoding” of features of human language. But as of now, what those features might be is quite unknown. In effect, we’re “opening up the brain of ChatGPT” (or at least GPT-2) and discovering, yes, it’s complicated in there, and we don’t understand it—even though in the end it’s producing recognizable human language.

This is the kind of thing I'm referring to. Even though we can look at pictures of the neuron activations, we don't really know what it's doing, any more than you can look at a picture of a brain scan of a person speaking and know why they decided to say those particular words. We know at a low level how the network works, of course, because we coded it. It's the emergent behavior that we don't understand, and that's the bit that's scary because that's where the AI risk lives. It's like how we understand particle physics pretty well, and chemistry to some degree, but biology is a massively complicated jungle that we've barely scratched the surface of.

Maybe there's something we could develop analogous to using an MRI as a lie detector, for these networks. But as far as I know, this is still an unsolved problem, and apparently really hard for networks that are smarter than you are: https://docs.google.com/document/d/1WwsnJQstPq91_Yh-Ch2XRL8H...

I'm not particularly terrified of GPT-4, but a more of what comes after. Maybe GPT-6 or 8, or another breakthrough non-LLM AI. I'm not sure if reading it will do much more than push my estimate of real AGI out a couple of years, but thanks anyway, I appreciate it.

I'm not confident I can say the same about, say: an LLM calling itself from within an infinite loop in which the LLM is asked to predict actions that will [eg. add money into a bank account], and then to write code that executes those actions, then to examine the results of those actions and update its plan. (This takes no additional breakthrough in AI architecture, only an increase in the reliability of GPT for it to become practical. It is basically a summary of the AutoGPT project).

I don't know if LLMs are anywhere close to any part of what goes on inside the human thought process. But I am sure that evolution was able to stumble upon the human thought process. In 2003 I was quite confident that was forever beyond the reach of humans to program into a computer, but in 2023 I am far less certain. I think human cognition is gradually becoming less of a mystery.