Ben Hylak 谈用模拟做 AI 智能体评测
Ben Hylak 讲怎么用模拟测智能体，还聊到模型会察觉自己在被模拟然后改行为，做 eval 的可以看看。
Hamel Husain 在 X 上发布了对 Ben Hylak 的访谈，主题是 AI Evals 课程中如何用模拟做评测。Ben 分享了在多家公司落地模拟评测的经验，指出输入输出测试会漏掉智能体的失败。他提到"模拟感知"现象：模型能察觉自己身处模拟环境并改变行为。访谈还覆盖回放生产 trace、重建智能体周边的数据与工具、跨运行检查答案与成本等环节。
Recently had the chance to interview @benhylak on how he is thinking about simulations in our AI Evals Course 😁
Simulations are incredibly useful for evals when you do them well. Ben discusses what he learned from implementing simulations for evals at several companies. Some fun topics came up, my favorite being "simulation awareness" - when models know they are in a simulation.
Below is a chapter summary:
02:02 Why input/output tests miss agent failures 10:59 Replay production traces to see what a change breaks 13:18 Recreate the data and tools around the agent 16:27 How agents detect simulations and change their behavior 20:39 Choose scenarios that test the change you made 22:00 Check answers, costs, and tool behavior across runs 26:43 Replay known failures to verify a fix 31:20 Common mistakes that make simulations misleading 32:20 Check that the simulation reproduces production behavior 37:37 When to start using simulations
You can also watch on YT here: https://t.co/EjlKZMw2y1
