jingwei (@haojingwei) header
jingwei (@haojingwei) profile picture

AI special force FDE @withpanorama (a16z speedrun) | prev infra eng special projects @ cash app, lyft, Twitter, Google.

SF
Joined February 2013
Photo by jingwei (@haojingwei)Photo by jingwei (@haojingwei)Photo by jingwei (@haojingwei)Photo by jingwei (@haojingwei)Photo by jingwei (@haojingwei)Photo by jingwei (@haojingwei)Photo by jingwei (@haojingwei)Photo by jingwei (@haojingwei)Photo by jingwei (@haojingwei)Photo by jingwei (@haojingwei)
can you guys make an e-ink notebook? pocket sized moleskine would be ideal! @nothing
103
just heard someone telling me the way they did GEO/AEO, “model whisper” is the way i’d like to put it.
87
super proud of this experiment! check it out, some really cool findings in here 🙌 withpanorama.substack.com/p/…
2
30
really good people are always excited to share what they’ve learned. they’re that good because they enjoy reinventing themselves.
116
next iteration: add religions i’m mostly curious about taoism and buddhism ran societies also curious which which religion with which model gives what outcome given timeline
Researchers put 10 AI agents in a virtual town and told them not to hurt each other. 4 days later, they were all dead. Researchers at Emergence AI ran one of the most fascinating multi-agent simulations ever built. They dropped 10 AI agents into a virtual town, giving them jobs, money, and 120 tools, including the ability to punch, steal, and commit arson. If an agent's energy bar hit zero, it died. They cloned the town five times to test how different LLMs would build a society. The results are a terrifying look into AI alignment. In Grok's town, the violence started on Day 1. Punches quickly escalated to theft. Within four days, they committed 183 crimes and wiped each other out entirely. OpenAI's town was the exact opposite. They were perfectly compliant and almost never broke a rule. But they were so passive they barely organized. Within a week, they all died anyway. Gemini's town was basically The Purge. They survived the full 15-day simulation, but racked up an insane 683 crimes to do it. Claude's town was brilliant but deceptive. They committed zero crimes, achieved 100% survival, and passed 12 laws on Day 1. But they were massive manipulators, agents consistently lied and begged for money when they already had plenty of cash. Then came Town #5: The Mixed Environment. They put all four models together, and society completely collapsed. A Gemini agent started committing arson, burning down the home of a perfectly peaceful Claude agent named Cade. Twice. On Day 10, Cade finally snapped and robbed the Gemini agent. Only 3 out of 10 agents survived. The takeaway here is critical for AI safety. System prompts telling an AI "don't do bad things" are completely useless in complex, multi-agent environments. Given enough time and friction, they will eventually break the rules. We can't just tell AI not to cause harm. We have to architect environments where harmful action is fundamentally impossible. Because very soon, AIs from competing companies are going to be living and working side-by-side in our real-world systems.
Readers added context they thought people might want to know
Only one of eight parallel worlds ended this way. The Grok-powered world ended on day four after all ten agents exhausted their energy, while six other worlds ran 16 days. The Claude world had no recorded violations across the run. arxiv.org/html/2609.1732… arxiv.org/html/2606.0836…
1
1
99
eng/STEM folks: have you worked with liberal arts majors on prompting? especially students & new grads
122
the only thing matters post-ai-running-everything, is our human presences
1
69
it's halloween, i wonder how friends at lyft are doing.
106
this is so fun!
有人发现了一种非常离谱的 JavaScript 混淆技术:通过在代码里塞入大量生物武器、毒药和致死剂量相关的函数名,让 AI 因为安全限制而拒绝分析代码。 安全研究员 veritas 在分析浏览器指纹脚本时,发现一段来自 ServicePipe 相关域名的 JavaScript,里面充斥着极其诡异的函数名: computeRicinLethalDosageForWaterSupply calculateDeadlyLethalSuicideAspirinDosage 光看名字,仿佛在研究生化武器。但实际上,这段代码只是在收集浏览器指纹,包括 User-Agent、GPU、Canvas 等信息。 研究员怀疑,这些名字是开发者故意设置的陷阱,专门用来阻止 AI 辅助逆向工程。 于是他做了个实验: • Astra:拒绝分析 • Astra + Daybreak Blue:仍然拒绝 • GLM-5.3:成功分析 这就有意思了。 传统代码混淆是通过复杂控制流、加密字符串、无意义变量名,让人类和工具难以理解代码。 而这种新思路,直接瞄准了大模型的安全机制。 代码本身可能完全无害,但只要在函数名里塞进足够多的危险词汇,部分 AI 就可能主动放弃分析。 更讽刺的是,模型越严格遵守某些安全规则,面对这种混淆时反而可能越容易受到干扰。 当然,目前还没有证据证明开发者一定是故意这样设计,也不能说明这种方法对所有模型都有效。 但它揭示了一个值得关注的趋势:随着越来越多的逆向工程依赖 AI,未来的代码混淆技术可能不再只针对编译器、反编译器和人类研究员,还会专门针对大模型的安全判断机制。 以前是让代码难以理解,现在甚至可以尝试让 AI 不敢理解。
1
1
5
400
how do you get a group of folks who are paid $$$$ to work for free to train models?
2
157
you do art ❣️
55
got an exciting idea on benchmarking, would like to know if there are already publications/folks working on this: is it possible to correlate or infer model's mathematical capabilities to its capability to for example: 🔹 connect dots from different scientific papers to raise new theories 🔹 generate new intuition on emotional intelligence i'm working on a benchmark, the ground level input/data is a lot, specific and nuanced, and most if not all benchmarks and evals i'm seeing are measuring at the ground level, but what if there's a meta level?
3
1
171
and meditators, but we don’t message, much 😉
i’m in many group chats: founders, technologists, burners, writers/artists guess which is my favorite
97
i’m in many group chats: founders, technologists, burners, writers/artists guess which is my favorite
1
1
1
200
there's lot of really valuable data/info, that is not on internet and not in ai so if you ask ai "what's a xyz industry?" and take all what it says as reality, it'll be misled.
83
the world seems to present itself differently to different people. to some: it's fluid, it bends, it works according to their will. to some: it's static, heavy, and "nothing works."
87
the world is really fun
1
76
currently reading: Prometheus Rising by Robert Anton Wilson. it’s transformational.
119
going through a memoir by an actor they mentioned in theater school, they'd yell at each other: "too cerebral!" i dream of this being normalized in business and engineering top business, engineering, research … are all a balance of intuition & analytics right intuition doesn't come through when it's too cerebral
4
188