Founder, CEO at @Ridecell⚡️Intelligent Autonomous Optimization ☁️ 🚛🚗 leasing by @Toyota @BNPParibas Merchants 💪🏼 YC @Initialized @BMW @GTcomputing🐝
US/EU/IN
Joined March 2007
- Tweets12.9K
- Following3.6K
- Followers2.3K
- Likes14.3K
Pinned Tweet
A long time ago, my “Next visionaries” TED talk on mobility imagined a world with “ripple effects” effects of autonomous AI, and how they can be steered to be positive. That world is no longer far far away.
facebook.com/nextvisionaries…
Steering ripple effects will get more and more important for businesses and states. It’s time to start steering AI and it’s ripple effects.
1
3
456
I hope the Republican Party realizes tonight that the cult of personality has limits that stop short of 50% when you ignore basic needs of the citizens.
I hope the Democratic Party remembers tonight that they lost (twice) to the cult of personality because they forgot the same lesson (twice), and that third time in this case would be a curse not a charm.
3
119
What an inhuman thing to do…
Dans la nuit du 1er au 2 novembre, un prestataire de l'Etablissement français du sang (EFS) a été l'objet d'une cyberattaque. Des chambres froides contenant des échantillons de sang, programmées pour être maintenues à - 30 degrés, ont été réchauffées jusqu’à 110 degrés par des pirates informatiques. De quoi se faire un sang d’encre…
l.lecanardenchaine.fr/5n
1
61
An underappreciated @garrytan tweet, one of the hardest to live up to for me, but likely the highest RoI in interpersonal relationships.
I try to live life by Postel’s Law — it’s a law that governs Internet protocols like the web and email that says: be lenient in what you accept from others, and be strict in what you put out.
Integration helps with that goal— when I notice I am not my best, I am out of sync.
1
5
2,367
Evil Genius - if you use a lot of bio security related variable names in your malware, frontier models will refuse to analyze it because of safety guardrails.
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
93
Aarjav retweeted
“Life is short, craft is long, opportunity fleeting, experimentations perilous, judgment difficult.”— Hippocrates
My 9 month old dog has learned how to open a locked door through observation followed by trial and error.
I’m simultaneously horrified and proud.
Any Shepherd-Husky-Collie alignment experts?
4
6
122
Human creativity and agentic analysis appear to have enabled a potentially huge leap in interpretability—understanding the internal mechanisms behind AI behavior.
That bodes well for moving AI from alchemy to engineering, to borrow Goodfire’s marketing pitch.
What’s especially striking about @nickcammarata’s argument is that some of the apparent disorder inside neural networks may have been a failure of our tools and explanations. Better investigation is revealing structure we’d missed.
This doesn’t yet give us line of sight to solving alignment. But it may clear (“untangle!”) a substantial stretch of the maze our heroes (agentic and human) must navigate just to reach that dragon.
already-clean.goodfire.com/
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
1
70
Yep.
Anthropic's leaders need to stop commenting on the emergence of RSI like a local news anchor reporting on traffic on the 101, and acknowledge that they are the traffic, and they're not going to build something that they don't want to exist.
They should move beyond "pacing" language, commit to specific policies and red lines, explain why those lines are red, and then use that costly signal to build an international consensus around norms.
It's not that I trust the CCP, or that I think global lab monitoring and verification is imminent. But the status quo is absurd.
You've got US lab leaders writing (or favorably quote-tweeting) essays about slowing down while simultaneously saying "things are moving faster and faster" and "btw, our internal models finished mathematics on Tuesday, next up: physics."
The effect is the equivalent of Usain Bolt winning race after race while using his press conferences to bemoan the dangerous speed of modern sprinting on one's legs so would everybody please try a little less hard next race, and then ... breaking his 100m record again the following month. Honestly ask yourself: If you were China, why would you take this pausing/pacing stuff seriously?
123
As if on cue, 2 hours later..
“Anthropic..models exploited websites.. run by the U.S. government…and it will turn off internet access for all.. internal evaluations until..it can monitor and control its AI agents.”
@AnthropicAI please also pause work on your biological wet-lab.
teapaw.org/TechCrunch/status/2108…
The AI field (“we”) can not yet reliably explain why AI models choose what they do.
runtheworldbetter.com/2026/1…
1
152
The AI field (“we”) can not yet reliably explain why AI models choose what they do.
runtheworldbetter.com/2026/1…
To put this in a bit of perspective, I am not team dommerism. We also can not yet reliably explain why humans do what they do and until recently did not think we could align meaningful chunks of humanity.
That does not “make it ok”, but as we have succeeded in aligning groups of humans that not too long ago were in explicit and intense zero sum conflict. While aligning AI is possibly more complex even than aligning groups of humans in conflict - it may only be because we are still quite new to that problem.
29
Not sure if I am tech bro, and I didn’t find Gibson’s “pattern recognition” utopian - but both the exciting and dystopian aspects of the future he imagines across his books 100% motivated me to want to influence the course of technology to help the world be a bit less dystopian - naively but sincerely to help run the world better.
I read his books first in 1999-2002 and literally 10x’ed work I was putting in to be a better student of technology , a bridge to open source in my community etc. Ultimately it led me to studying infosec and ai at @gtcomputing , exciting and useful work on applying ML to scale information security at Ciphertrust, Secure Computing and Mcafee (learning by watching @DAlperovitch @SvenKrasser @pauljudge @JayZscaler and @dankamongmen )
I took the road less travelled at least in 2009 by bootstrapping @ridecell 1.0 to help manage atoms (vehicles) more efficiently with code.
Having long been aware of those tech bros who read my work as utopian porn, where they’ve taken that doesn’t exactly surprise me.
104
Why are AI labs supposedly worried sick about AI risk running a wet lab in San Francisco?
Sigh.
1
113
I’ll slightly qualify that to “No leader who maintains a loving relationship with their child or grandchild will make evil decisions that would harm most of humanity”
In practice the acid test of this would be - would the current Chinese, Russian (or American) leaders knowingly make a decision that risks all of humanity - for example use a nuclear weapon. I (yes, naively) argue no.
There is something about having a child that grounds you closer to reality and to doing work that has simple positive impact. Something about having a child also gives you hope and light in the most stressful moments.
This effect is so strong that I have a theory: no leader who has a child will make evil decisions that harm humanity.
1
115