The argument that "AI is just an autocomplete parrot that can never discover anything new" just took a massive hit.
In 1946, Paul Erdos posed the unit distance conjecture, predicting a strict sub-polynomial bound on how many point pairs can sit exactly 1 unit apart in a 2D plane.
For 80 years, human mathematicians tried to prove or disprove it using symmetrical grids and lattices. They failed because human brains have an innate symmetry bias.
An OpenAI reasoning model just disproved it by discovering an asymmetric counterexample using high-degree algebraic number fields, formalized by Princeton's Will Sawin on arXiv at n^1.014.
Why did humans miss it for eight decades?
Because the point configuration looks like shattered glass. It possesses zero visual symmetry.
link is here
Top comments (1)
Some comments may only be visible to logged-in visitors. Sign in to view all comments.