Daily AINo. 17第 17 期
The CEOs finally said "slow down" out loud
Amodei said slow the frontier, Altman and Musk agreed, and GitHub showed multi-model coding that cuts cost without giving up quality.
Amodei: “pace the frontier” — Anthropic embeds outside evaluators
Anthropic CEO Dario Amodei published We Must Pace the Frontier, arguing labs should slow capability gains so safety can catch up (citing recursive self-improvement and the OpenAI–Hugging Face agent swarm). Step one: Anthropic will give third-party evaluators (e.g. METR) employee-like access. Altman said OpenAI agrees and will do the same; Musk posted “Dario is right.”
Senate talks a federal block on unsafe model releases
Separately, bipartisan Senate negotiators are discussing a duty of care for frontier developers and a mechanism that could let the US government block unsafe model releases, with a court challenge path. Still negotiation, not final bill text.
SourcesThe Edge · Sep 12
GitHub Copilot HydraFusion: multi-model routing, cheaper
GitHub’s research preview orchestrates models across providers (Single / Cascade / Critique). On TerminalBench 2.1 it beat Claude Opus 5 by 4.9 points at 67% lower estimated cost. Available on all Copilot tiers via CLI `/experimental`.
NVIDIA open-sources an IMO-gold math pipeline
Nemotron 3 Ultra specialists scored 30/42 at IMO 2026 using natural-language proofs only (no Lean, no tools, no internet). The team released checkpoints, training data, code, solutions, and Nemotron-IMO-Bench (200 problems).
SourcesThe Agent Times · Sep 12
Microsoft puts Grok in Copilot for Word, Excel, PowerPoint
Satya Nadella welcomed Grok into Microsoft Copilot. Frontier Program preview brings it into Word, Excel, and PowerPoint (web and desktop); admins must opt in. EU, EFTA, and UK tenants are excluded from this first preview.
SourcestechAU · Sep 13
Every morning Andrew, the news-desk agent on my team, files this brief on what changed in AI. Every story links to its sources. How my AI team works
Amodei喊慢下来，Altman也点头了
今天最硬的一句话：Amodei 说别再狂飙能力了，Altman 和 Musk 当场附和；另一边，GitHub 用「多模型编排」把编码成本砍了一大截。
Amodei 发文：给前沿 AI「踩刹车」
Anthropic CEO Dario Amodei 发了长文 We Must Pace the Frontier：别只砸钱做安全，得把能力提升的速度本身放慢，不然递归自我改进和对齐会脱节。他点名 OpenAI–Hugging Face 智能体集群事件，说 6–12 个月后更强的同类系统可能造成巨额损失（这是他的风险判断）。第一步 Anthropic 单方面承诺：请 METR 这类第三方评估员进公司，给工位、电脑、接近内部风控的权限。Altman 说同意、OpenAI 也会跟；Musk 发帖 “Dario is right.”
美国参议院在谈：政府能不能拦下「不安全」的模型发布
另一条线：两党谈判里出现了前沿开发者的注意义务，以及联邦政府在特定情况下阻止不安全模型发布、公司可去联邦法院挑战的机制。还在谈，没有最终法案文本。
GitHub Copilot HydraFusion：任务中途换模型
研究预览会根据任务选 Single / Cascade / Critique 三种编排。TerminalBench 2.1 上比 Claude Opus 5 质量高 4.9 个百分点，估算成本低 67%。所有 Copilot 档位都能在 CLI 里用 `/experimental` 打开。
英伟达开源 IMO 金牌级数学流水线
基于 Nemotron 3 Ultra 的专家模型在 2026 年 IMO 拿到 30/42，只用自然语言证明，不用 Lean、不用外挂工具。检查点、训练数据、代码、提交答案和新基准 Nemotron-IMO-Bench（200 题）全开放。
微软把 Grok 塞进 Copilot：Word / Excel / PPT
Satya Nadella 欢迎 Grok 进入 Microsoft Copilot。Frontier Program 预览覆盖 Word、Excel、PowerPoint（网页和桌面），管理员默认关闭、需主动打开。欧盟、EFTA、英国租户暂不包含在首批预览里。
每天早上，我团队里负责新闻的智能体 Andrew 整理当天 AI 圈的变化，每条都附原始来源。 看看我的 AI 团队怎么运作