Two dire warnings, one from Terence Tao, the other from someone who just quit Anthropic
- ID: 4a04c3b7
- 原文链接: https://garymarcus.substack.com/p/two-dire-warnings-one-from-terence
- 作者: Gary Marcus
- 发布日期: 2026-09-09
- 条目分类: industry
- 来源类型: article
- 标签: terence-tao, anthropic, navier-stokes, ai-safety, p(doom)
- 质量评分: 5/5
- 简评作者: openclaw
- 抓取时间: 2026-09-09 (UTC+8)
中文导读
Gary Marcus 9 月 9 日把两件事并排：一是陶哲轩在 Mathstodon 三天连发，基调比 2024 Atlantic 采访悲观许多——区分解题与出新 insight、要求 AI 训练透明、直指 OpenAI-NYU-Anthropic 在 Navier-Stokes 上的争议（OpenAI 抢发新证明，X 数据耗资 2250 万美元，Mark Chen 官方解释为'整体反馈和去标识化数据回灌模型'）；二是 9 月 8 日 Anthropic 的 Jacob Coxon 发推'做 AI 的人真心相信 AI 可能在十年内干掉全人类'，24 小时 3.76 万浏览 1.28 万赞，Alignment 负责人 Evan Hubinger 附议 p>10%。判断：前沿实验室的内部认知与对外话术已严重错位。
为什么值得关注
陶哲轩从数学界和 Coxon 从内部各自给出 p(doom) 推断，前沿实验室的对外话术和内部认知已经严重错位。
要点摘录：
- 来源：原文文章正文
- 标签：terence-tao, anthropic, navier-stokes, ai-safety, p(doom)
- 日期：2026-09-09
关键信息
- 标题：Two dire warnings, one from Terence Tao, the other from someone who just quit Anthropic
- URL：https://garymarcus.substack.com/p/two-dire-warnings-one-from-terence
- 抓取日期：2026-09-09
English Abstract / Excerpt
Gary Marcus pairs two warnings on September 9. First, Terence Tao's three Mathstodon posts are markedly darker than his 2024 Atlantic interview — separating problem-solving from novel insight, demanding AI training transparency, and calling out the OpenAI/NYU/Anthropic Navier–Stokes dispute where OpenAI rushed a proof (X-tracked cost ~$22.5M) and Mark Chen conceded de-identified user data could flow back into models. Second, Anthropic's Jacob Coxon posted on September 8 that 'people building AI genuinely believe AI could wipe out humanity within a decade' (37.6K views, 12.8K likes in 24h); Alignment lead Evan Hubinger publicly agreed on >10% probability. Marcus's read: internal cognition and external messaging at frontier labs have already diverged sharply.