Pinned Tweet
So Scott Aaronson was the originator of this amazing demo!
I tried this a long time ago, and today, re-confirmed that I am 80% predictable
people.ischool.berkeley.edu/…
The internal model that OpenAI used to solve these 700 problems is VERY similar to what's available to subscribers in 6 months - Scott Aaronson
youtu.be/NWNbAEivyvs?si=tSXM…
I have NEVER not been disappointed in the retarded first responses from Grok
Every single time, I point to an acronym, a meme, or an allusion that Grok didn't parse until I beg it to be useful on round two
Grok's a ho I only ask wrt drugs, sketchy plans, + stuff Claude disdains
"He has the worst case of little man syndrome I’ve ever seen"
A little size-ist of big boy Fetterman to call this out, but the point he's making is true: Politicians should not barricade themselves from their voters, so this is indeed a red flag
I asked Sen. John Fetterman (D-PA) for his thoughts on the increasingly public dispute between Abdul El-Sayed and Haley Stevens. Here's what he told @jewishinsider:
"Rep. Haley Stevens simply held Abdul El-Sayed accountable for his views on Israel and fixation with Jewish money. El-Sayed is going after a woman and banning her from campaign events for having an opinion. El-Sayed's overreaction demonstrates why he struggles with groups of Michigan voters. He has the worst case of little man syndrome I’ve ever seen."
"this new approach goes straight for the safety mechanisms of large models. The code itself might be completely harmless, but by cramming enough dangerous keywords into the function names, some AIs might just proactively bail on the analysis"
Make AIs "too scared"
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。
If you light a fire, you may get burned
Commenters gasp at destruction without asking why H*m*s wanted to incinerate Gaza
Terrorists chose to crush the populace, actively lied to members, all as a way to intentionally kill civilians at genoc_dal scale, all for sympathy abroad
1ea9707d1e849c09620acd1ac5bf7ae0bed1ff6af8c9345d3d12459d5f023a46
Time-stamp to demarcate my entrance into the Singularity
The blind have seen, & I have proved a theorem that extends our grasp of Poincaré duality spaces
SHA-256 of claim to publish upon completion of peer review
Leftist academics are gonna get left out of the AI revolution, because they prefer hallucinating to joining the LLM revolution
Losers gonna lose
Please peg @PikettyWIL next!
Replying to @gabriel_zucman
Our work has long been thoroughly scrutinized by the global community of researchers. Our code and data are public. Comments, extensions, and refinements have been discussed in numerous conferences and journals for many years.
This is how science operates.
Paul Sas retweeted
"If you pick different parts of science, you see the same pattern show up: Progress is accelerating, and it has started to accelerate in the last two years."
I’m speaking at the IMF in an hour about the singularity and why I believe it has begun and the choices we'll need to make. Watch live here: youtube.com/live/Rqt9kWoTJbc…
Replying to @mattparlmer
Get out of the new way, if you can't lend a hand!
Behav sci WILL upgrade to machine legible models at scale
Woke socialist sciences can go back & hand stir more Marxist shibboleth soup
When they go full ostrich, they lose their .05 level of contact with reality
The world gains
Will the PRC control the future of AI? Their govt investment plan aspires to produce full stack token factories
The government is mobilising investment for what it calls the “6 networks”. These include the power, computing & communications systems an AI-led economy will need
Brilliant of @MargRev move to push their archives
Endless spelunking supported, e.g
First link to @slatestarcodex is 4/7/13
Tyler links to an SSC post written 2 days earlier
SSC was then ~2 months old
claude.ai/artifact/7gpruvEAV…
HPMOR saves a number of lives!
Rational agency can even keep the Ruscist war machine from drafting you until every consideration has been examined, discussed, and appealed to be Less wrong
Why wasn't he drafted despite the previous 7 rejected applications? The system is fascinating:
If your views prevent you from serving in the army, the russian constitution grants you the right to alternative civilian service (random basic jobs that might depend on your education/skills but generally involve something like cleaning in a hospital or teaching kids in a school).
Sadly, the russian constitution is pretty fake. Russia tries to appear legalistic, but in reality, doesn't particularly care about the law.
You can still use the law and get out of military service.
Twice a year, there's a drafting period. 6 months before the start of the drafting period, you can send an application for alternative civilian service to your local drafting body. It will most likely be rejected for ridiculous reasons; the drafting bodies dislike having their stats ruined by you, and don't particularly care about the law.
During the drafting period, once the drafting commission decides to reject your application, they can draft you.
However, if you appeal that decision in court, it is paused until the court's decision is in force. And if you lose in the local court, you can appeal that decision in a higher court of appeals (think the highest court of a US state), and then everything is paused until the case is decided there.
(There's also a court of cassation and higher courts, but that one does not pause anything).
(Courts in Russia are a separate fascinating subject; courts and judges dislike having *their* stats ruined by making a decision where a government body loses, as those potentially impacts their careers. Did you know that to be appointed a judge in Russia, among other things, you'll be given a criminal case and asked to write two well-argued verdicts: guilty and not guilty? If you judge in criminal cases, you have a budget of 1-2 acquittals per year.)
All of this takes many months. It normally lasts for much longer than the drafting period; often, it takes two (which is why, I assume, Kirill only had to submit 8 applications throughout 6 years, rather than 12).
And at the end of a drafting period, the decisions of the drafting commission concerning it are automatically voided; if you submit a new application for the next drafting period, the cycle repeats from scratch!
(If courts work unusually fast and the local drafting body wants to draft you, you can repeat the same thing with the medical fitness commission: it will declare that you're fit to serve, and you can appeal *that* in court. This works only once, though, unlike alternative civilian service applications.)
Unfortunately, most people don't know that you can simply use this trick forever (until you're old enough) or they aren't agentic enough to use it. Even if you simply refuse to serve, without knowing anything about the laws, they'll use all kinds of tactics to pressure you, but for most people, the worst that can happen is a criminal case that results in like a $5k fine.
But in absolute numbers, there are still many stories like that, especially among sufficiently agentic people, which correlates with having read HPMOR.
Compare that to "21 people reported donating kidneys because of the ACX kidney donation post"
slimemoldtimemold.com/2026/0…
Paul Sas retweeted
Replying to @doristsao
Semantics consists in the functional and causal role these representations play within the *entire system*. Searle isolates one component, observes that it lacks understanding, and illegitimately transfers that verdict to the organized system it helps instantiate.
Searle's Chinese Room card-switcher would realize a "semantic coordinate system...
there is a translation manual hidden in the computer. The Anthropic [J-space] paper is telling us that the man in the sea of Chinese symbols is not as semantically helpless as Searle makes them"
I think the recent anthropic paper on global workspaces in LLMs gives new insight into the Chinese Room Puzzle. The most surprising part of the paper for me was that averaging Jacobians across contexts could produce a sensible result that would let you read out the word a particular token in particular layer would produce. What that means is that starting from some intermediate layer, there is a *fixed coordinate system for representing meaning* = semantics, tied to verbalizable English. It's not the case that in one context "spider" points in this direction and in a other context it points in a totally different direction. "Spider" points in the same direction across all contexts, and so survives averaging. So we can think of different layers of LLM as talking to each other in this common coordinate system, or language, to finally produce a sensible output.
I think that is a very different picture of what is happening inside an LLM from the one Searle presents, where, in the particular example he gives in the video here, you have the restaurant script symbols, the story symbols, question symbols, etc. and they're all just 'squiggles' and 'squaggles' that you, the computer whose ability to understand is being questioned, put together through arbitrary syntactic rules. It turns out the squiggles and squaggles are all aligned to the same semantic coordinate system, i.e., there's only squiggles! So basically, there is a translation manual hidden in the computer. The Anthropic paper is telling us that the man in the sea of Chinese symbols is not as semantically helpless as Searle makes him out to be.
In the human brain these semantic directions are further tied to patterns of sensory input: science.org/doi/10.1126/scie…
Paul Sas retweeted
Two new DeepMind Institute essays:
What is Work For?
As AI advances how do people value their jobs, for material needs and beyond?
The Economics of Intelligence
When will AI capabilities impact market structure, wages and jobs?
Read the essays here -> bit.ly/dmi_home
Sharing AI generated content feels as cringe as publishing a playlist of your favorite pornos
PREMIUM Fable5.1 slop
claude.ai/share/8369c4f2-30a…
















