GPT-5.6 Sol Ultraが50年前の数学の予想を解決 — しかし、単に過去の研究成果を組み合わせただけなのか?
OpenAIのGPT-5.6 Sol Ultraが、64個の並列サブエージェントを使用して、50年前の「cycle double cover予想」を1時間未満で証明しました。数学者のThomas Bloomはこの証明を検証しましたが、AIが先行文献を引用しなかったことを批判しており、モデルが本当に創造的なのか、それとも単に粘り強いだけなのかという疑問を投げかけています。
openai's new model, gpt-5.6 sol ultra, reportedly generated a complete proof for the cycle double cover conjecture—a problem unproven for 50 years—in under an hour by coordinating 64 subagents. mathematician thomas bloom of the university of manchester confirmed the proof's validity, describing it as elementary and short, but criticized openai's paper for omitting citations to prior work from 1983. bloom argues the ai likely found the solution not through creative insight but by executing a 'counterintuitive twist' that humans would have abandoned too quickly, reflecting a pattern where models succeed by applying extreme patience to existing theories. this development reinforces the debate over whether reasoning models produce genuine novelty or simply recombine known literature, raising ethical concerns about attribution in ai-generated research.