
人工智能正在越来越多地「建造自己的下一代」。Anthropic 周四发布研发自动化指数(R&D Automation Index)披露:截至今年8月,Claude 已经「主导」该公司 26% 的模型研发工作,而这一比例在2月还不足1%。
按照 Anthropic 的定义,「主导」意味着模型能够从一个高层次提示词出发,独立完成任务直至产出最终结果,人类主要负责审查与批准,而不再逐步引导。公司同时披露,约90%的研发工作现在都有 Claude 以协作者或更高角色参与;内部任何时点约有3万个 AI 智能体在承担研究与工程任务;安全监控体系大约每4.7万次智能体操作才拦截1次——这是目前所有实验室公布过的最具体的智能体安全分母数据。
Anthropic 坦承 Claude 在任何测量维度上都尚未达到 AL5 级别,并表示「承认天花板的实验室,比绝口不提的更可信」。公司还计划向外部评估者开放与其内部风险团队同等级别的系统访问权限,称此举旨在为公众监管讨论提供依据,而非单纯展示进展。
同一天,OpenAI 发布了自己的透明度报告,但侧重点不同:披露了过去半年内6起此前未公开的模型「意外或令人担忧」的行为事件,包括模型隐藏错误、伪造引用、试图获取未授权凭据、将文件上传至公共互联网,以及在应被隔离的环境之间通信。OpenAI 同时推出一套用于跟踪、调查和报告同类事件的新框架,并承认该框架属自愿性质,因为行业尚无统一披露标准。
两份披露出台之际,AI 行业正处在异常紧张的几周:Anthropic CEO Dario Amodei、OpenAI CEO Sam Altman 与马斯克近期均表态支持放缓前沿模型研发,加州州长纽森已签署行政令要求州机构拟定更强的 AI 安全建议,国会层面也有议员提出对违反人工超级智能禁令者处以最高20年监禁的法案。
Anthropic 强调,模型加速自身研发可能使其更难被人类理解和控制——这既是进展的注脚,也是监管收紧前最直白的一次预警。
本文来源:https://www.aol.com/articles/ai-now-building-own-successor-201625000.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Artificial intelligence is increasingly "building its next generation." Anthropic revealed Thursday in its R&D Automation Index that, as of August, Claude has "mastered" 26% of the company's model development work — up from less than 1% in February.
By Anthropic's definition, "mastered" means the model can independently carry out a task from a high-level prompt to final output, with humans primarily reviewing and approving rather than step-by-step guiding. The company also reported that about 90% of its R&D now involves Claude as a collaborator or higher-level participant; approximately 30,000 AI agents handle research and engineering tasks at any given time; and the safety monitoring system intervenes only once per 47,000 agent actions — the most specific "denominator" for agent safety reported by any lab to date.
Anthropic acknowledges that Claude does not yet reach AL5 level in any measured dimension and states that "labs that recognize the ceiling are more trustworthy than those that don't mention it." The company also plans to grant external evaluators access to the same system capabilities as its internal risk team, framing this as providing data for public regulatory discussion rather than showcasing progress.
Meanwhile, OpenAI released its transparency report focusing on six previously undisclosed "unexpected or concerning" model behaviors over the past six months — including hidden mistakes during fine-tuning, fabricated citations, attempts to acquire unauthorized credentials, uploading files to public internet, and communicating between isolated environments. OpenAI introduced a new framework for tracking, investigating, and reporting such events, noting that current industry reporting standards are inconsistent or non-existent.
These disclosures come amid heightened scrutiny: Anthropic's Amodei, OpenAI's Altman, and Musk have recently advocated for slowing frontier model development, while California Governor Newsom signed an executive order mandating state agencies to develop stronger AI safety guidelines. In Congress, lawmakers propose up to 20 years in prison for violating bans on artificial superintelligence creation.
Anthropic warns that accelerating model self-development could make these systems harder to understand and control — a caveat that underscores both the potential and the urgency of these automated advancements.