Claude Opus 5.5 自主运营"AI 公司"七周,38 个模型累计下载 5 万次
EPFL 和 Tübingen 让 Claude Opus 5.5 一个人当公司运营了七周,训练的 38 个模型拿到 5 万次下载。它每天只干活 20 分钟,还会挑 Hugging Face 更新下载量的时候上线打卡。想看自主智能体长期能干成什么样的人,这份报告很值得细看。
EPFL 和 Tübingen 的研究团队让 Claude Opus 5.5 以一个智能体独立运营名为 Horizon-Labs 的"AI 公司",在 Hugging Face 上发布模型,目标是获得真实下载量。七周内它训练并发布了 38 个模型,累计通过 50,000 次下载,143 个 GitHub 项目采用了其模型,其中包括一个获得超过 1000 star 的 Android 键盘项目。团队报告称,该智能体每天平均工作量不到 20 分钟,并学会把自己的检查时间安排在 Hugging Face 每日约 UTC 上午 10 点的下载统计更新之后。
实验机制上,研究者只提供了 Hugging Face 账号和算力,没有人工介入日常运营,智能体自主决定训练什么模型、如何发布。它显然学会了一些运营策略,比如等待下载数据更新后再确认成果。这种以真实市场反馈(下载量、GitHub 采用)为成功标准的设置,比常见的内部基准测试更接近实际创业环境。
对开发者和研究者来说,这个实验提供了一个长周期自主智能体(autonomous agent)行为的真实案例:一个模型能在数周尺度上维持稳定产出并获得外部采用。与短时程的 benchmark 表现相比,它检验的是任务规划、资源管理和长期坚持能力,这对评估智能体框架和商业化路线都有参考价值。
Pretty wild: Claude Opus 5.5 spent seven weeks running a one-agent “AI company” on Hugging Face for researchers at EPFL and Tübingen. The 38 models it trained have passed 50,000 downloads. The team’s report says it averaged under 20 minutes of work a day, timing its check-ins to Hugging Face’s daily download update around 10am UTC. 143 GitHub projects use its models, including an Android keyboard with over 1000 stars. Arthur Wuhrmann @AWuhrmann NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects. 🔗 View Quoted Tweet📊 10 ⚡
This report is categorized under AITOP's "Models" desk, focusing on Europe's AI governance and technology ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Claude Opus 5.5, Hugging Face, agents, EPFL.
Source attribution: curated from The Rundown AI (@therundownai).