本文讨论了 OpenAI 未发布的 AI 模型 Astra，据报道该模型在一天内解决了十大数学开放难题，这一成就对数学领域和 AI 发展具有重大意义。核心论点是，尽管结果令人印象深刻，但它们也突显了预期的转变，因为 AI 在高级数学方面展现了超人类能力，类似于之前在网络和编码方面的突破。作者指出，其他模型如 Fable 和 Sol 在指向这些问题时也能解决其中一些，但 Astra 的“Juice”——即独立识别并解决此类问题的能力——标志着一个潜在的阶跃变化。文章还探讨了更广泛的后果，包括消化 AI 生成的证明的挑战、错位的风险以及 AI 研发时间线的加速。结论强调，虽然这是一个重要的里程碑，但它不一定意味着 AGI，但它确实移动了目标，并强调了谨慎管理 AI 快速发展的紧迫性。
This article discusses OpenAI's unreleased AI model, Astra, which reportedly solved ten major open mathematics problems in a single day, a feat that has significant implications for the field of mathematics and AI development. The core argument is that while the results are impressive, they also highlight a shift in expectations, as AI demonstrates superhuman capabilities in advanced math, similar to previous breakthroughs in cyber and coding. The author notes that other models, like Fable and Sol, can also solve some of these problems when pointed at them, but Astra's 'Juice'—its ability to independently identify and solve such problems—marks a potential step change. The article also explores the broader consequences, including the challenge of digesting AI-generated proofs, the risk of misalignment, and the acceleration of AI R&D timelines. The conclusion emphasizes that while this is a major milestone, it is not necessarily AGI, but it does move the goalposts and underscores the urgent need for careful management of AI's rapid advancement.
核心贡献 · Key contributions
未发布的 OpenAI 模型 Astra 解决了十个重大开放数学问题，包括雅可比猜想的一个反例。 Astra, an unreleased OpenAI model, solved ten major open mathematics problems, including a counterexample to the Jacobian conjecture.
这些结果以低成本（不到 2000 美元）和短时间实现，表明 AI 在数学推理方面取得了重大进展。 The results were achieved at a low cost (under $2,000) and in a short time, indicating a significant advance in AI's mathematical reasoning.
论文强调，其他模型（Fable、Sol）在被指向这些问题时也能解决其中一些，表明数学证明能力存在“过剩”。 The paper highlights that other models (Fable, Sol) can also solve some of these problems when pointed to them, suggesting an 'overhang' in mathematical proof capabilities.
作者认为，这一结果是科学推理的阶跃变化，对 AI 研发的自我改进循环具有影响。 The author argues that this result is a step change in scientific reasoning, with implications for AI R&D self-improvement loops.
论文讨论了 AI 自动化 AI 研发的可能性，导致时间线加快，并可能实现超级智能。 The paper discusses the potential for AI to automate AI R&D, leading to faster timelines and possibly superintelligence.
作者认为，在这些问题中，验证通常比生成更容易，使 AI 能够在定义明确、形式化的任务中表现出色。 The author suggests that verification is often easier than generation in these problems, enabling AI to excel in well-defined, formalized tasks.
局限 · Limitations
论文缺乏对照组，因为 OpenAI 没有用类似预算测试其他模型，难以评估 Astra 的独特贡献。 The paper lacks a control group, as OpenAI did not test other models with a similar budget, making it hard to assess Astra's unique contribution.
Astra 在向人类解释时无法识别证明的难点，这是一个弱点，限制了其可解释性。 Astra's inability to identify which parts of a proof are hard when explaining to humans is a weakness, limiting its interpretability.
结果仅限于定义明确、形式化且易于验证的问题；可能不适用于不可验证的领域。 The results are limited to well-defined, formalized problems with easy verification; they may not extend to unverifiable domains.
论文承认 AGI 的目标可能移动，有些人可能不认为这些结果是 AGI 的证据。 The paper acknowledges that the goalposts for AGI may move, and some may not consider these results as evidence of AGI.
存在古德哈特定律的风险，即优化可测量任务可能导致不可测量领域的意外后果。 There is a risk of Goodhart's law, where optimizing for measurable tasks may lead to unintended consequences in unmeasurable areas.
论文章节 · Sections（共 12）
目录Table of Contents
这些结果有多令人印象深刻？How Impressive Are These Results?
我们本可以叫它 Sol 或 Fable 吗？Could We Have Called Sol or Fable?
它来了It’s Coming
他们仍然看不到即将到来的“它”是什么They Still Don’t See What Is The It That Is Coming
这是 AGI 吗？Is This AGI?
AI 解决了他最喜欢的问题The AI Solved His Favorite Problems
这令人惊讶吗？Was This Surprising?
人们难道不感到惊叹吗？Are People Not Impressed?
这在多大程度上改变了我们的预测？How Much Does This Change Our Predictions?