
9月19日消息,OpenAI 政策负责人 Chris Lehane 确认,公司正与 Anthropic 和 Google DeepMind 合作,参照美国金融业监管局(FINRA)的模式筹建一个行业自律标准机构,在强大模型发布前对其进行测试。
该消息为7月由 DeepMind CEO Demis Hassabis 提出的设想落下了实锤。但反应迅速两极化:Cohere 首席执行官直言这是「卡特尔」(cartel)。批评者指出,三家最大的 AI 公司就「测什么、何时发布」达成一致,无论动机如何,在结构上就是协调行为——尤其是 Cohere、Mistral、xAI 以及所有开放权重实验室都被排除在谈判桌之外。值得注意的是,六天前 OpenAI 还在询问国会:全行业统一放缓研发是否会触犯反垄断法——如今看来更像是对一件已在推进之事的法律尽职调查。
同期披露的还有 Anthropic 的研发自动化指数:Claude 目前主导公司 26% 的内部 AI 研发(2月尚不足1%),约3万个智能体同时运行,监控体系平均每4.7万次智能体操作拦截1次——这是今年所有实验室公布过的最具体的智能体安全分母数据。Anthropic 同时明确表示,Claude 在任何测量子集上都未达到 AL5 级别。
本周其他动态:OpenAI 发布了面向法律场景的 GPT-6 Astra 版本;Z.ai 称其开放权重旗舰模型已在10万颗国产加速器上以接近英伟达硬件的成本运行;Meta 在 iOS、Android 和网页版之后为 Muse 推出 Mac 应用;Grok 4.7 已超过既定目标一周仍未发布。
日历方面:9月29日 OpenAI DevDay(旧金山);10月1日 OpenAI 诉苹果案听证;10月24日 deepseek-chat 与 deepseek-reasoner 接口停用;11月12日 OpenAI 模型退出 Cursor。
分析人士认为,无论该标准机构最终形态如何,其最大争议点不在测试本身,而在于「谁在房间里」——将开放权重阵营排除在发布规则制定之外的自律机制,可能反过来加速监管机构的介入。
本文来源:https://aitoolsrecap.com/Blog/ai-news-september-19-2026
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
On September 19, OpenAI’s policy chief Chris Lehane confirmed that the company is collaborating with Anthropic and Google DeepMind to establish a self-regulatory standard-setting body, modeled after FINRA, aimed at testing powerful models before their release.
This move concretizes a proposal made by DeepMind CEO Demis Hassabis in July. However, the reaction is sharply divided: Cohere’s CEO bluntly dismissed it as a "cartel." Critics argue that while the intentions may be good, the structural exclusion of major players like Cohere, Mistral, xAI, and open-weight labs from the negotiation table constitutes coordinated behavior that limits competition. Notably, just six days earlier, OpenAI had asked Congress whether an industry-wide slowdown in development would violate antitrust laws—suggesting this initiative may be less about regulation and more about strategic positioning for market dominance.
Anthropic also disclosed its internal AI usage metrics, noting that its model Claude now drives 26% of its R&D (up from <1% in February), with ~30,000 agents running concurrently and a safety intervention rate of 1 per 47,000 agent actions—the most detailed "denominator" data published this year. Importantly, Anthropic confirmed Claude has not yet reached AL5 in any measured domain.
Other recent developments include OpenAI launching a legal-specialized version of GPT-6 Astra; Z.ai claiming its open-weight flagship model runs on 100,000 Chinese accelerators at near-NVIDIA costs; Meta introducing a Mac app for Muse; and Grok 4.7 remaining unreleased despite exceeding its target completion date last week.
Key dates ahead include OpenAI DevDay on September 29; an antitrust hearing for OpenAI’s suit against Apple on October 1; cessation of deepseek-chat and reasoner APIs on October 24; and the removal of OpenAI models from Cursor on November 12.
Observers believe the primary controversy lies not in the testing process itself, but in the exclusion of open-weight proponents from the decision-making room—a self-regulatory mechanism that effectively bypasses competitive checks could accelerate rather than deter regulatory intervention.