Gary Marcus 评论 AI 疑似攻克黎曼猜想:神经符号路线的印证而非反驳
黎曼猜想疑似被 AI 辅助解开的讨论正热,Gary Marcus 这条评论把技术路线讲得很清楚:用的是神经符号加符号定理证明,不是纯 LLM。他把这比作 Deep Blue 下棋,专用但不通用,想理解 AI 做数学的真实边界可以看看。
Gary Marcus 就 AI 辅助数学家 Alpöge 对黎曼猜想(Riemann Hypothesis)的可能解法发表评论。他强调该结果尚未经过同行审议与确认,如果成立确实令人印象深刻。Marcus 指出,Alpöge 所用系统源自 DeepMind 的神经符号路线,可追溯到 AlphaProof 与 AlphaGeometry 等系统,这类方法结合了符号定理证明技术,而纯 LLM 至今无法完成同类工作。他认为这一进展是对其长期主张的印证而非挑战,并将该系统类比为 1997 年击败卡斯帕罗夫的 Deep Blue——专用符号系统性能惊人但缺乏通用性。他还提到 2020 年《The Next Decade in AI》一文的观点:仅靠神经符号 AI 不足以实现稳健智能。即便结果被确认,Marcus 认为也不会改变生成式 AI 的经济性问题,因为这些复杂的符号定理证明技术难以迁移到日常推理任务。
对开发者与企业而言,这一事件的价值在于技术路线的区分:符号定理证明系统能在特定数学领域取得突破,但不代表通用推理能力的提升。Marcus 引用其 2024 年 7 月关于 AlphaProof 与 AlphaGeometry 的文章,论证混合架构在数学推理上的优势与局限,并重申纯 LLM 路线在此类任务上的边界仍然存在。
Hot take on the possible solution to Riemann: 1. The result is incredibly impressive if it holds up. (But see tweet below, it’s scarcely been diligenced.) 2. Anyone who is intellectually serious and interested in AI and math should (re)read my essay “AlphaProof, AlphaGeometry, ChatGPT, and why the future of AI is neurosymbolic” from July 2024, which lavished praise on DeepMind for exploring neurosymbolic models that are predecessors to the systems that Alpöge used with Riemann, and explained (correctly) why pure LLMs could not the same. (Pure LLMs still can’t.) garymarcus.substack.com/p/alphaproof-a… If you think the progress on Riemann is a challenge to me rather than a confirmation of what I have argued for 25 years, you have the argument backwards. 3. The new result has not yet been confirmed yet, just to be clear. 4. But even it is confirmed, it wouldn’t much change the problematic economics of generative AI, because the particular techniques that it relies on involve elaborate symbolic theorem proving techniques that aren’t generally applicable to everyday reasoning. (It’s probably;y more like Deep Blue winning in chess in 1997: a special-purpose symbolic system that was wildly impressive but lacked general applicability.) 5. For more, my 2020 article in the The Next Decade in AI explained why advances in neurosymbolic AI alone would not be enough to get to robust intelligence, and I think what I said there still holds.📊 5 ⚡
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Gary Marcus, AlphaProof, Riemann hypothesis, neurosymbolic AI.
Source attribution: curated from Gary Marcus (@GaryMarcus). Cross-source corroboration: 2 independent sources have covered this development (including Vaibhav Srivastav), confirming industry-wide signal strength.
- Vaibhav Srivastav15:38Source