Jonathan's Blog

This is the best article I’ve seen so far summarizing the incredible results that OpenAI and Google DeepMind got on this year’s IMO competition. They far outdid even the betting market expectations as the article points out. I am a little skeptical of OpenAI’s announcement though. I glanced through both OpenAI’s and DeepMind’s submitted answers, and OpenAI’s answers are so spartan that I can’t believe they were written by anything close to a normal LLM. I would suspect they probably RL’d their model within an inch of its ability to even produce sensible text before setting it loose on the IMO problems. Maybe DeepMind tidied up their model’s answers before presenting them but at first glance they seem much closer to what I would expect as an output from a more normal LLM.

I get some amount of schadenfreude watching each new field realize that AI models are much better than the vast majority of people in their field. First it was the spam SEO blog post writers, then the artists, and now the mathematicians. So far I think the programmers are handling it the best.