Mathematician says Astra can already make real math breakthroughs
Constantin Kögl says he cracked the result of paper 148 with Astra in a six-hour session and believes OpenAI's internal model is stronger but not dramatically better than Astra.
Controversial Opinion: The internal model by @OpenAI that solved all these math problems is better but not crazy better than Astra. I am an expert on papers 148 and 153. The result of 148, which is stunning, I got with Astra myself after a 6-hour session on September 27 and maybe relying on my local context. The main problem was solved after the model thought continuously for around 3 hours. (See the paper we wrote on it that is now on the arXiv.) 153 is impressive, but I am quite convinced I would have gotten that too if I had tried as hard as with 148. I am sure that for the Quasi-Riemann result they used significantly more compute. They claim that for most of the 4000 problems, they used 3 hours, which aligns with my experience for 148. What I am saying is that Astra already can make real breakthroughs in mathematics. The internal model is surely better, but at least in my area of expertise it doesn't seem to strongly outperform Astra. What are your opinions?