System 1 and 2

Update: Kahneman actually talks about this in his NeurIPS 2021 invited talk: Perception.

I forget where this was mentioned (either in one of the AI podcast episodes, or this numenta video), but basically we can think of GPT-3 as being the first almost perfect copy of system-1 human thinking, which is how Kahneman chose to dichotomise how our brains work---essentially, system-1 is the fast, intuitive thinking, while system-2 is the deliberate, rational, logical thinking.

Pattern recognition is basically system-1, and it's where all the problems of correlation ≠ causation occur, since it's just focused on predicting things by association. And that's basically what GPT-3 is capable of doing.

The question is then how do we get to system-2 thinking, which is pretty much our competitive edge---deliberate thought. Here's a random #idea that I had on my run, and I suspect someone has already thought about: what if system-2 = system-1 + simulation? It seems to me that the crucial piece of the puzzle is basically being able to simulate the world, or at least some very crude model of it. Once you have the capacity to simulate the world, then you can run your system-1 inferences, and see how things compare to the truths of your simulation, while also making sure to update your model of the world against reality.