OpenAIのAstraが数学問題を10件解決 — なぜGary Marcusはハイプだと主張するのか?
OpenAIの社内モデル「astra」は数学と理論計算機科学における10件の未解決問題を解きましたが、批評家たちはこのブレークスルーが過大宣伝されていると主張しています。
openai's 'astra' model generated proofs for ten major open problems in fields like quantum complexity and theoretical computer science, with researchers estimating the cost at around $2000 in api credits. while figures like noam brown tout this as a major step for scientific reasoning, commentators such as gary marcus and dean ball suggest the implications are being exaggerated. marcus argues that the capability to solve theoretical math problems does not inherently translate to general intelligence or ability to solve mundane tasks. the debate highlights a persistent disconnect between benchmark performance on specialized tasks and the broader expectations for agi.