There you have it! An AI model just disproved a math conjecture that stood unsolved since 1946.
June 5, 2026 · 0 likes · 0 comments
Workforce AI
There you have it! An AI model just disproved a math conjecture that stood unsolved since 1946.
Let me be precise, because the headlines aren’t.
No, AI did not “solve the unit distance problem.” That problem is one of the hardest in combinatorial geometry and it is still wide open.
What happened is narrower — and more interesting.
A specific conjecture Paul Erdős posed in 1946, sitting inside that larger problem, was just proven false. The work came out of an OpenAI model. Humans then refined the proof and tightened it into something rigorous.
Here’s the part that should stop you.
Thomas Bloom — a working mathematician who publicly tore apart an earlier fake OpenAI “Erdős” claim — validated this one and put his name on it. The skeptic became a co-author.
And Tim Gowers, a Fields Medalist, called it a milestone.
Let that sink in.
This isn’t a chatbot writing a haiku. This isn’t autocomplete. This is a machine producing first-rate symbolic mathematics — the kind of work that was supposed to be the last bunker of elite human cognition.
A few caveats, because I don’t do hype:
It’s a preprint. arXiv, not peer-reviewed yet.
The broader problem remains unsolved.
And humans were in the loop refining the result — this was collaboration, not a machine working alone in the dark.
But the trend line is the only thing that matters here.
For two years the comforting story was: AI takes the routine stuff, the “real thinking” stays human. Coding, then law, then medicine — and pure mathematics was supposed to be the final wall.
That wall just developed a crack.
When the frontier of abstract reasoning starts moving, every job built on “but a human has to think about it” is on notice. Not next decade. Now.
More in my upcoming book, REPLACEMENT, July 28th!
Now ask yourself which jobs are actually safe.
Let me be precise, because the headlines aren’t.
No, AI did not “solve the unit distance problem.” That problem is one of the hardest in combinatorial geometry and it is still wide open.
What happened is narrower — and more interesting.
A specific conjecture Paul Erdős posed in 1946, sitting inside that larger problem, was just proven false. The work came out of an OpenAI model. Humans then refined the proof and tightened it into something rigorous.
Here’s the part that should stop you.
Thomas Bloom — a working mathematician who publicly tore apart an earlier fake OpenAI “Erdős” claim — validated this one and put his name on it. The skeptic became a co-author.
And Tim Gowers, a Fields Medalist, called it a milestone.
Let that sink in.
This isn’t a chatbot writing a haiku. This isn’t autocomplete. This is a machine producing first-rate symbolic mathematics — the kind of work that was supposed to be the last bunker of elite human cognition.
A few caveats, because I don’t do hype:
It’s a preprint. arXiv, not peer-reviewed yet.
The broader problem remains unsolved.
And humans were in the loop refining the result — this was collaboration, not a machine working alone in the dark.
But the trend line is the only thing that matters here.
For two years the comforting story was: AI takes the routine stuff, the “real thinking” stays human. Coding, then law, then medicine — and pure mathematics was supposed to be the final wall.
That wall just developed a crack.
When the frontier of abstract reasoning starts moving, every job built on “but a human has to think about it” is on notice. Not next decade. Now.
More in my upcoming book, REPLACEMENT, July 28th!
Now ask yourself which jobs are actually safe.