看看这些工程师怎么教AI干活：智能体自动提交PR、训练模型、修复代码，甚至催更模型权重，效率提升肉眼可见。
一个夜间报告智能体连续运行数周后，将经理的私人团队分析作为PR提交到代码库。一位CTO通过智能体将每周2到10个PR的产出提升至每晚。一个微调分类器带来1200万美元收入，但修复成本高昂，现被前沿模型技能取代，修复时间缩短至一小时以内。智能体每晚打开数百个GitHub问题，请求研究人员发布模型权重，数千个问题后仅收到两次投诉。
What teaching the system actually looks like, from nine workflows: - A nightly report agent ran cle...
What teaching the system actually looks like, from nine workflows: - A nightly report agent ran clean for weeks, then posted its manager's private team analysis as a PR on the repo. Nothing changed. The model decided to be helpful - A CTO with 15 recurring meetings ships 2 to 10 PRs a week: brief the agent at 5pm, let it run all night, test the stack in the morning. One overnight run trained two production ML models - A fine tuned classifier brought in $12M at 50x ROI and still turned into tech debt: every fix cost a week of retraining. Rebuilt as skills on a frontier model, a fix now ships in under an hour - Agents open hundreds of GitHub issues a night asking researchers to publish their weights. Thousands of issues in: two complaints, and researchers' agents now reply to his agents - One engineer, zero lines of code written this year, an email product thousands trust. The rule: spend half your time teaching the system what it got wrong Watch the AI Architects playlist: youtube.com/watch?v=nxokqO… 💬 0 🔄 0 ❤️ 1 👀 400 📊 1 ⚡