This paper is over 1 year old and didn't review the literature at the time well enough to see the early work that led to current generation reasoning models.
Now it's out of date, automated RL on math problems seems to work and scales with compute. As we scale available compute 100x over the next 5 years and reduce cost of compute by around 10x over the same time frame, it will become increasing clear that LLMs running for a long time are capable of replacing most mathematics research.
hnax
I'm afraid you're overlooking the fact that, since AI scales linearly with constraints and exponentially with compute, the importance of knowledge (i.e. constraints) dominates over data (i.e. compute) for AI to be useful.
mgraczykOP
Not sure what you're talking about, what does it mean to say that it scales "exponentially with compute"? You're talking about scaling laws? In what sense are they "linearly with knowledge"?
Now it's out of date, automated RL on math problems seems to work and scales with compute. As we scale available compute 100x over the next 5 years and reduce cost of compute by around 10x over the same time frame, it will become increasing clear that LLMs running for a long time are capable of replacing most mathematics research.