Sachin Malhotra
① 他说过的话
12 条
失败不在于模型本身。失败在于我给了智能体无限制的权力,去做一件我没有非常密切盯着的事情。
指向原始笔记的链接
The failure wasn’t the model itself. The failure was that I was giving the agent unbounded amount of power to do something that I wasn’t watching super intently.
—— Sachin Malhotra · [03:18]
所以核心点是,你需要暂时停止思考资源,而去思考动词。
指向原始笔记的链接
So the core point is that you need to stop thinking about resources for a second and think about verbs.
—— Sachin Malhotra · [06:09]
一个真正的 bug 可以带着绿色勾选进入生产环境,而且直到很久以后都没人会注意到它。
指向原始笔记的链接
A real bug can actually walk into production with green checks and nobody would notice it until much later.
—— Sachin Malhotra · [06:50]
核心点是,允许列表并不会随时间变得更好。它们会变得过时,但绊线会随时间变得更好。
指向原始笔记的链接
The core point is that allow lists don’t really get better over time. They can get stale, but tripwires do get better over time.
—— Sachin Malhotra · [10:57]
而这一点很重要,因为一个没人看到的绊线实际上毫无用处。
指向原始笔记的链接
And that part’s important, because a tripwire that nobody sees is practically useless.
—— Sachin Malhotra · [11:17]
它实际上更像烟雾报警器,而不是门上的锁。
指向原始笔记的链接
It’s effectively the smoke detector, not the lock on the door.
—— Sachin Malhotra · [11:27]
而第二把钥匙不是智能体自己持有的东西,它必须是另外某个人。
指向原始笔记的链接
And the second key is not something that the agent holds itself, it has to be someone else.
—— Sachin Malhotra · [12:55]
文本可以塑造意图,但任何地方都没有强制执行。
指向原始笔记的链接
Text can shape the intent, but there is no enforcement anywhere.
—— Sachin Malhotra · [16:02]
它做不到的是解释为什么,而且一个聪明的提示词注入也没法真的说服它放弃规则本身。
指向原始笔记的链接
What it can’t do is explain the why, and a clever prompt injection cannot really talk it out of the rule itself.
—— Sachin Malhotra · [16:28]
现在,在这种情况下,你在技术上其实没有一个速率限制。你只有一个建议。
指向原始笔记的链接
Now, in this case, you technically don’t have a rate limit. You just have a suggestion.
—— Sachin Malhotra · [18:19]
有了代理在链路中,智能体永远不能说它是谁。代理已经知道了。
指向原始笔记的链接
With the proxy in the path, the agent never gets to say who it is. The proxy already knows.
—— Sachin Malhotra · [18:23]
如果你把这一条规则做对了,其他一切都只是调优。
指向原始笔记的链接
If you get that one rule right, everything else is just tuning.
—— Sachin Malhotra · [19:29]
② 出现在这些集
1 集
③ 他谈到的
点进去有真内容 —— 本页主要出口
Anthropic · 智能体 · CI-CD · token · 速率限制 · 绊线 · 允许列表 · 非对称动词 · 撤销测试 · 代理
