Anthropic 数学家承认 Claude 黎曼猜想相关产出是 25 万页未验证文档
Anthropic 数学家自己承认那批 Lean 文件是“slop”转述、25 万页没法读,还赶在股价跌回发行价当天官宣。想跟风说 Claude 攻克黎曼猜想的人,先看看这份拆台笔记再说。
Anthropic 数学家 Levent Alpoge 发布了一份 4 页说明,称 Claude 参与了黎曼猜想相关工作的写up,但文档内的细节揭示了实际情况:相关写up超过 25 万页,且未被完整验证。在脚注中他承认这是对一份“slop”摘要的转述,“可能毫无意义”,并称自己尚未掌握其中涉及的数学内容。
根据转述,最终产出是 14 个拆分的 tar.gz 格式 Lean 文件,而非一份可直接阅读的证明文档。他本人拒绝公开原始 PDF,因为认为它是一部“不可读的 slop 大杂烩”。文中 Theorem 1 的开头甚至从“实数是全序的,一个数若不是 1/2 就要么大于要么小于 1/2”这类平凡陈述展开。
此事发生的时间点耐人寻味:就在 Anthropic 股价从 IPO 后高点回落 27%、跌向 83 美元发行价的当天上午,这份声明被放出。Gary Marcus 转发了这一分析,认为任何关于“AI 攻克黎曼猜想”的庆祝都为时过早。对于关注 AI 数学能力评估的开发者和研究者,这说明形式化验证(Lean)产出与人类可读、可复核的数学证明之间仍有很大距离。
and see this which makes it clear that any rejoicing is premature: Alexey Fateev @superalesha Everyone thinks Claude just solved the 167-year-old Riemann Hypothesis, until you actually open the PDF. Hours after Anthropic's $ANT dropped 27% from its post-IPO peak, their mathematician Levent Alpoge posted a 4-page note claiming Claude cracked it. Inside Footnote 2 of alpo. ge/remark.pdf, he admits Claude's writeup hit 250,000+ pages, which Claude cutely called a "book". Theorem 1 literally opens by noting that because real numbers are totally ordered, if a number isn't 1/2 then it's either greater or less than 1/2. He wouldn't post the PDF because he calls it a "massive unreadable travesty of 'slop'", and admits in Footnote 4 he's "paraphrasing a 'slop' summary, which might make zero sense". Instead, he dumped 14 split tar.gz Lean files and wrote that he still hasn't mastered the math around them. He even messaged Dario to clear the announcement, then added in Footnote 1: "There is an obvious cynical view, to do with the timing, but it is false." Yeah, 250,000 pages of unverified Lean slop just happened to land the exact morning ANT slid back toward its $83 IPO price. 🔗 View Quoted Tweet
This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison and AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Anthropic, Claude, Riemann hypothesis, Lean.
Source attribution: curated from Gary Marcus (@GaryMarcus).