Mathematicians scrutinise AI proofs said to topple a pillar of complexity theory
OpenAI and, shortly afterwards, Anthropic claim their models have refuted a widely accepted conjecture in computer science; independent checks are still under way.
OpenAI on Monday announced three results in mathematics, computer science and physics that it says were found by its internal models. The most striking, according to The Next Web, is a refutation of the Exponential Time Hypothesis, a 1999 conjecture holding that certain logic puzzles cannot be solved substantially faster than in exponential time. Countless research papers rely on that assumption. OpenAI says it checked the proof mechanically with the proof language Lean.
TechCrunch reports that Anthropic posted its own refutation on GitHub a few hours later, after one of its mathematicians suggested that OpenAI had drawn on existing work. A group around the authors of the conjecture and MIT researcher Ryan Williams is examining the result, but stresses that this is no certification. OpenAI believes the result does not break widely used encryption.
Researchers are, however, increasingly irritated by the way OpenAI publicises its breakthroughs, The Verge writes, after a wave of hundreds of preprints in early October. OpenAI is giving 5 million dollars to the Simons Foundation for workshops on results found by AI.
Transparency: this edition was compiled by Claude, made by Anthropic.