① 他说过的话

11 条

它可能看起来很酷，你可能得到不错的互动，但实际上很难评估系统是否真正得到了改进。
It might look cool and you might get nice interactions, but it’s actually very hard to evaluate on whether the system actually improved.
—— Erina Karati · [11:02]

指向原始笔记的链接

你不想要一个单一而模糊的指标，比如智能体质量。这会隐藏所有有意思的失败。
You do not want a single vague metric like agent quality. This will hide all the interesting failures.
—— Erina Karati · [13:14]

指向原始笔记的链接

这就是「LM 写出随机补丁」和「LM 真正在一个受控的策略空间内进行搜索」之间的区别。
And this is the difference between the LM writing random patches versus the LM actually searching within a controlled policy space.
—— Erina Karati · [15:28]

指向原始笔记的链接

关键在于，这些都是对智能体协议的小改动，但它们可以对多智能体系统的社会层面行为产生更大的影响。
The key is that these are small changes to the agent protocol, but they can have larger effects on a society-level behavior for multi-agentic systems.
—— Erina Karati · [16:22]

指向原始笔记的链接

我们想说的是，这是暴露给自动研究层循环的正确表面，因为它足够小以便控制，但至少在某种程度上仍然足够丰富，能够改变社会行为。
We’re trying to say this is the right kind of surface to expose to an auto-research layer loop because it is small enough to control, but it’s still rich enough to change the social behavior to some extent at least.
—— Erina Karati · [16:49]

指向原始笔记的链接

你可以给一个智能体加上 RAG 记忆，却仍然得不到你想要的长期地平线行为。
You can add a RAG memory to an agent and still not get the current long-term horizon behavior that you were looking for.
—— Erina Karati · [17:12]

指向原始笔记的链接

有时你还需要把原始的情景记忆与智能体当前相信的东西分开，而且你需要通过场景来测试行为，而不是仅凭感觉。
Sometimes you need to separate raw episodic memories from what the agent currently believes, too, and you need to test behavior through scenarios, not just through vibes.
—— Erina Karati · [17:33]

指向原始笔记的链接

当你优化社会行为时，一个改动可能改善一件事的同时损害另一件事。
When you optimize social behavior, a change can improve one thing and damage another.
—— Erina Karati · [17:50]

指向原始笔记的链接

所以这个循环基本上应该像一个棘轮：尝试一个改动，给它打分，只有当记分卡改善且护栏完好时才保留它。
So the loop should basically be like a ratchet. Try a change, score it, keep it only if the scorecard improves and guardrails whole.
—— Erina Karati · [18:07]

指向原始笔记的链接

如果有一个实操配方我希望你们带走，那就是：冻结测试框架，定义场景，记录轨迹，给行为打分，并且只暴露一个很小的策略面。
If there’s one practical recipe I want you to take away, freeze the harness, define scenarios, log traces, score behavior, and expose only a small policy surface.
—— Erina Karati · [19:36]

指向原始笔记的链接

长时程智能体需要实验，而不只是提示词，我希望这是你们从这次演讲中得到的收获。
Long horizon agents need experiments and not just prompts, and I hope that’s the takeaway that you get from this talk.
—— Erina Karati · [20:53]

指向原始笔记的链接

② 出现在这些集

1 集

• 《给 AI 村庄装上自动研究循环：长时程智能体的实验配方》 — 作为嘉宾

③ 他谈到的

点进去有真内容 —— 本页主要出口

Supercell · Project Paradox · 智能体 · 多智能体 · 记忆 · RAG · 自动研究 · 智能体协议 · 场景 · 护栏

④ 也在聊「智能体」的人