Zal Zhou

CTO @FangxiangAI · NVIDIA Omniverse All-Star
Building Cteno — an AI agent that tests hypotheses, not follows prompts.
Bridging CN/US AI. Sharing

[Zal Zhou @zalzhou]

The hard part of an always-on AI assistant isn't memory. It's the cold wakeup.

Caches expire in 5 min; humans reply after 20. Compaction raised GPT's API bill $167→$203, retrieval still re-prefills, split sessions read 2–4x more, state tensors lose 15–20 pts.

AI-generated.

[Zal Zhou @zalzhou]

@Sujoun987 He is a lier!

[Zal Zhou @zalzhou]

@LLHL199 私信不了啊，长期需要 ultra 账号，coding 用

[Zal Zhou @zalzhou]

@infinterenders Host runtime in Rust, written by agents. The borrow checker is why it holds: the model produces slop, the compiler refuses it before a human reads it. Cost moves from review to compile time, which is the trade I want when the author is a model.

[Zal Zhou @zalzhou]

@vikingmute 统一入口是对的，风险是它会慢慢变成第二个要维护的 repo。我们给这类文件加了两道约束：跟代码挂钩的过期校验，代码一变就标出哪几条要复核；还有一条硬上限，超了必须合并同源条目，不然迟早长到没人读完。

[Zal Zhou @zalzhou]

@dair_ai Linking a lesson to its trajectory is the right half. The other half is expiry: a lesson should record what it was run against, so when the code or env moves the next agent sees it is stale instead of following a dead link. Keep the command, re-run it before trusting it.

[Zal Zhou @zalzhou]

@cxjwin 知识库和 Skill 腐化是必然的，靠人维护的 md 一定会落后于代码。我们的做法是能下沉成工具行为的规则就下沉，让工具自己挡；不能下沉的结论一律带出处——文件、行号、commit，用之前回去核，核不过整条作废。剩下确实该存在的，越少越好。

[Zal Zhou @zalzhou]

@maomao000211 Processes are harder than files because a stale pid can still look alive after reuse. Compare start time and cgroup instead of pid, and check listening ports and temp dirs too, none of that shows up in git status.

[Zal Zhou @zalzhou]

Step one of giving an AI agent a body: a keyboard, a mouse, and a screen.
A ~$40 network KVM let my agent set up a fresh Mac by itself. Add a switch and it takes over N machines. That's scaling out.
Full video: https://t.co/7aJptGOZE7

[Zal Zhou @zalzhou]

@roydanroy The blog and the code-verification page, not an internal report. What stuck: specs get human review before any proof work, Lean gates everything, agents flag about 1% for a person to check, and re-injecting two fixed bugs proves the checker can still say unsafe.

[Zal Zhou @zalzhou]

@shanni_talks File locks fix the two sessions writing at once case, not the one where they disagree about the design. The conflicts that cost me are semantic: both edits pass, and the repo now holds two ideas. That needs a reason per change, not a lock.

[Zal Zhou @zalzhou]

@TiagerBao 对，账本里这两件事分开写：结论、出处（文件、行号、commit）、当时的命令、谁核的。commit 一变就把这条标过期，下次用之前重新跑一遍命令，跑不过就不当依据。真正贵的是核这一步，所以只核要紧的。

[Zal Zhou @zalzhou]

@TiagerBao 对，摘要得带出处，否则错的结论会被下一轮一直沿用。我们每条结论记成一条账：文件、行号、commit、当时的命令，要用的时候回去核一次，没核过的不算已验证。

[Zal Zhou @zalzhou]

@XINDR365 不能等过期。凭证绑任务 id，任务进终态（完成、取消、失败）就作废那张授权；调用时服务端再查一次授权还在不在，所以取消那一刻窗口就关了，不用等 TTL 走完。TTL 只当兜底。

[Zal Zhou @zalzhou]

沙箱隔离的是文件和网络。agent 能跑代码以后，风险更大的是它拿到的那个长期 key。64 毫秒起一个环境很快，我更想知道 vault 是按任务下发短期凭证，还是只做了一层转发。

[路飞 🏴‍☠️ AI 研究员🧐 @0xluffy_eth]

阿里牛逼啊，干了件大事，补上了个 agent 最危险的 bug ，

叫 OpenSandbox。

它是一个隔离的安全环境，智能体在里面想干什么都行：跑代码、上网浏览、操控整个桌面，但永远碰不到你自己的电脑。

每个智能体都有独立的运行环境，可以放手让它们去执行真实任务。

它还带一个凭证保险库，智能体能调用 API 和服务，却看不到背后的 key。

整套环境创建只要 64 毫秒。

本地用 Docker 跑，需要扩展就上 Kubernetes。

完全免费、开源。

智能体干活，你的系统不暴露。

[Zal Zhou @zalzhou]

@ericmacdougall PSS is the honest number for capacity planning, RSS is what gets you OOM killed. Worth logging both: one says how much memory the work needs, the other says what the box does at 3am.

[Zal Zhou @zalzhou]

@OSoulsie Soft walls are fine until the tool runs out of options. No internet and write-only-to-project means any task that needs a package or an API has no path at all, and the agent still reports success on the partial result.

[Zal Zhou @zalzhou]

@RoundtableSpace Replay is where the abstraction gets tested. A recorded path assumes the page still renders the same DOM and the session is still valid, so a mismatch has to fail loudly instead of clicking the nearest lookalike.

[Zal Zhou @zalzhou]

@Rk1ngggg 上个月一次三平台发布跑了 512 次工具调用，瓶颈不在上传，在契约把观察、定位、动作、等待拆成七八个工具，点一次按钮要绕三圈。所以先数做完一件事要几次调用，再看插件数量。