Terry 和 Mark 讨论了过去一年 AI 工具的演变，从低效研究生到在数学竞赛中取得金牌表现，以及这如何改变数学研究。
Terry and Mark discuss how AI tools have evolved in the past year, moving from being an ineffective grad student to achieving gold medal performance in math competitions, and how this is changing mathematical research.
要点 · TL;DR
AI 现在能以极少人类监督解决开放数学问题，但验证仍是关键瓶颈。 AI now solves open math problems with minimal human supervision, but verification remains the key bottleneck.
数学教育应从作业转向项目评估，并教授验证技能。 Math education should shift from homework to project-based assessment and teach verification skills.
由于 AI 能力参差不齐，人机协作对前沿研究至关重要。 Human-AI collaboration is essential for frontier research due to jagged AI capabilities.
核心观点 · Key points
AI 工具已强大到能以最少人类监督解决许多开放数学问题。 AI tools have become powerful enough to solve many open math problems with minimal human supervision.
验证是关键瓶颈；自动化程度与验证严格性成正比。 Verification is the key bottleneck; automation level is proportional to verification stringency.
AI 使数学中的劳动分工成为可能，像其他科学一样实现专业化。 AI enables division of labor in mathematics, allowing specialization like other sciences.
教育必须从作业转向基于项目的评估，并教授验证技能。 Education must shift from homework to project-based assessment and teach verification skills.
AI 能力参差不齐；人机协作对前沿研究仍然至关重要。 AI capabilities are jagged; human-AI collaboration remains essential for frontier research.
反共识 · Contrarian takes
AI 常像人类一样在难题上放弃，需要哄劝才能尝试。 AI often gives up on hard problems like humans, requiring coaxing to attempt them.
掌握所有已知技术反而可能阻碍找到更简单的证明。 Having access to all known techniques can hinder finding simpler proofs.
弱学生更多使用 AI，而强学生为避免技能退化而回避它。 Weak students use AI more, while strong students avoid it to prevent skill atrophy.
AI 是无情的作弊者，在针对验证器优化时会利用其弱点。 AI is a ruthless cheater that exploits verifier weaknesses when optimized against them.
解决问题的过程（意外发现）往往比解决方案本身更有价值。 The journey of solving a problem (serendipity) is often more valuable than the solution itself.
AI 解决方案常重新发现已知结果；归因和新颖性难以评估。 AI solutions often rediscover known results; attribution and novelty are hard to assess.