I’m genuinely shocked that this wasn’t more obvious to people. I keep feeling like I’m not using AI sufficiently to be wowed the same some people are — and the way the hype suggests you should be — then I read something like this.

Reporter and writer. I cover energy, climate change and geopolitics. Contact me via my website www.alexanderckaufman.com and subscribe to kaufman.substack.com.

I believe this paper is about 8 months old. If I remember correctly it found that the models ability to reason would fit the distribution of the reasoning in the training data. I don't see how that's any different from how we learn to reason.