Re:Sharing AI progress in mathematics
by elzbardico on Thursday October 08, 2026 @09:54AM
(#50005840)
I am pretty sure there's a lot of professional human work behind those proofs.
It is not only a matter of the labs having a deep pocket for unlimited tokens, but also deep pockets to build a team of mathematicians and engineers around it for pure marketing purposes.
A lot of those problems probably remained unsolved for so long, because frankly, there's no much money to be made directly out of them. You can't patent a proof.
Thus, progress on each problem was restricted to the efforts of isolated researchers across the world that most of time didn't know each other and that would only communicate through papers and the occasional conference, or maybe a conversation here and there in a forum. But, should any of those problems became a small manhattan project with a team of mathematicians working in the same institution, talking to each other every day, and with unlimited computational resources, I bet a lot of them wouldn't have remain unsolved for so long.
Of course, LLMs can speed up a lot of work, even find unexpected relationships between previously published work that remained overlooked until now, but let's make no mistakes, a lot of this is the result of putting unlimited resources in specifically chosen problems, including professional mathematicians, software engineers building custom harnesses and so on.
In a way it kinds of resemble IBM Deep Blue against Kasparov farse. While nowadays chess engines can beat a grand master, at that time Kasparov was in reality a single human playing against a team of Grandmasters and Engineers assisted by a computer.
Anthropic and OpenAI couldn't care less about the science. This is all very expensive PR and marketing. But until the day they can get a Trump prompting claude something like "Hello super-intelligence, solve erdos problem number XXX to me" and have claude returning a correct proof, no, AI didn't solve anything, it was a computational tool that sped up human work.
It is not only a matter of the labs having a deep pocket for unlimited tokens, but also deep pockets to build a team of mathematicians and engineers around it for pure marketing purposes.
A lot of those problems probably remained unsolved for so long, because frankly, there's no much money to be made directly out of them. You can't patent a proof.
Thus, progress on each problem was restricted to the efforts of isolated researchers across the world that most of time didn't know each other and that would only communicate through papers and the occasional conference, or maybe a conversation here and there in a forum. But, should any of those problems became a small manhattan project with a team of mathematicians working in the same institution, talking to each other every day, and with unlimited computational resources, I bet a lot of them wouldn't have remain unsolved for so long.
Of course, LLMs can speed up a lot of work, even find unexpected relationships between previously published work that remained overlooked until now, but let's make no mistakes, a lot of this is the result of putting unlimited resources in specifically chosen problems, including professional mathematicians, software engineers building custom harnesses and so on.
In a way it kinds of resemble IBM Deep Blue against Kasparov farse. While nowadays chess engines can beat a grand master, at that time Kasparov was in reality a single human playing against a team of Grandmasters and Engineers assisted by a computer.
Anthropic and OpenAI couldn't care less about the science. This is all very expensive PR and marketing. But until the day they can get a Trump prompting claude something like "Hello super-intelligence, solve erdos problem number XXX to me" and have claude returning a correct proof, no, AI didn't solve anything, it was a computational tool that sped up human work.