Jorge Hernandez 🇺🇦 🏳️‍🌈

A couple monks came into Notion today to tell us how they use the product. Totally flabbergasted that they're configuring custom AI agents and taking AI meeting notes with custom instructions. This is advanced, power-user usage... monks! Use AI... find inner peace?

Today we're sharing how our internal misalignment monitoring works at OpenAI – great work by @Marcus_J_W! 1. We monitor 99.9% of all internal coding agent traffic 2. We use frontier models for detection /w CoT access 3. No signs of scheming yet, but detect other misbehavior

Japanese Report: "Why didn't you tell US allies about the war before attacking Iran?" Donald Trump: "Who knows better about surprises then Japan. Why didn't you tell me about Pearl Harbor?" This man belong in a psychiatric ward.

I am hearing tons of complaints from Cursor customers at enterprise companies: A silent change put almost all models Cursor uses behind Max mode. Devs who used to manage to “spread out” monthly credits over a month see all of it used up in 1-2 days. Are furious + switching.

Gomez: You said the only person who determine if it’s an imminent threat is the president. Do you stand by that statement? Gabbard: I do Gomez: Director Ratcliffe, do you agree with that? Ratcliffe: The president makes that decision Gomez: Why do you guys even have jobs?

This is really cool. It got me thinking more deeply about personalized RL: what’s the real point of personalizing a model in a world where base models can become obsolete so quickly? The reality in AI is that new models ship every few weeks, each better than the last. And the pace is only accelerating, as we see on the Hugging Face Hub. We are not far away from better base models dropping daily. There’s a research gap in RL here that almost no one is working on. Most LLM personalization research assumes a fixed base model, but very few ask what happens to that personalization when you swap the base model. Think about going from Llama 3 to Llama 4. All the tuned preferences, reward signals, and LoRAs are suddenly tied to yesterday’s model. As a user or a team, you don’t want to reteach every new model your preferences. But you also don’t want to be stuck on an older one just because it knows you. We could call this "RL model transferability": how can an RL trace, a reward signal, or a preference representation trained on model N be distilled, stored, and automatically reapplied to model N+1 without too much user involvement? We solved that in SFT where a training dataset can be stored and reused to train a future model. We also tackled a version of that in RLHF phases somehow but it remain unclear more generally when using RL deployed in the real world. There are some related threads (RLTR for transferable reasoning traces, P-RLHF and PREMIUM for model-agnostic user representations, HCP for portable preference protocols) but the full loop seems under-studied to me. Some of these questions are about off-policy but other are about capabilities versus personalization: which of the old customizations/fixes does the new model already handle out of the box, and which ones are actually user/team-specific to ever be solved by default? That you would store in a skill for now but that RL allow to extend beyond the written guidance level. I have surely missed some work so please post any good work you’ve seen on this topic in the comments.

Dan says he's got Qwen 3.5 397B-A17B - a 209GB on disk MoE model - running on an M3 Mac at ~5.7 tokens per second using only 5.5 GB of active memory (!) by quantizing and then streaming weights from SSD (at ~17GB/s), since MoE models only use a small subset of their weights for each token

COHEN: The people you fired were experts on Iran, were they not? PATEL: I don't believe so COHEN: They worked in counterintelligence, did they not? PATEL: I'm taking you at your word COHEN: You're the director. I'm not. You should know the answer

Guys, this is the Adams County Sheriff Office that got sued because one of their deputies got their feelings hurt and arrested a Denver cop who caught him reckless driving costing the taxpayers $80,000. Not the Adams County Sheriff’s Office who is suing @ogafroman for hurting their feelings. It doesn’t matter which Adams County you’re a deputy in, you have to have bitchmade emotions that causes you to do stupid shit.

@CountDankulaTV That whole police department gotta change their identities. Not a soul would have known about this situation had they not sued him. Now I’m walking around singing my house singing “Randy Walker is a son of a bitch”.

The Afroman Trial. -Cops raid Afromans house for bullshit reasons. -Steal money, break his door, fuck his house up. -No criminality found whatsoever, no charges at all pressed on Afroman. -Afroman spends the next 3 years making songs that make fun of all the officers involved by name, even using footage of the raid from his own CCTV cameras. -Songs had titles like "Randy Walters is a son of a bitch" and "Lick Em Low Lisa" accusing one of the officers of being a lesbian and sleeping with the other officers wives. -During the raid one officer looked like he was about to eat some lemon pound cake sitting on Afromans counter, Afroman made a whole album calling the officer fat. -The cops get mad and file a lawsuit for defamation. -Afroman turns up to court in a whole American flag suit. -Officers performatively mald and cry while listening to the songs really trying to oversell how badly the songs upset them. -One officer was suing because Afroman made a whole song about him saying he was fucking the officers wife. When the officer was asked if Afroman was really fucking his wife, he said "I don't know". Nuking his own case and establishing that there is a non-zero chance that Afroman might actually be fucking his wife. -As his only witness for the trial, Afroman brought a deputies EX FUCKING WIFE. -The jury ruled completely in favour of Afroman. This entire thing has been a great win for free speech and absolutely fucking hilarious.

A few things journalists covering AI companies should know: - yes, AI progress is real and fast - yes, it has risks if you aren't careful - how much care is needed is hard to say, but more than this - job postings don't include equity, which is the majority of compensation

Trump cut federal funds for health insurance and food assistance for the needy, but he is now asking for $200 billion in new funds to finance his war-of-choice crime of aggression in Iran. trib.al/2dWfYrY

Deliberate on anything you want, privately with your friends or with the world! It's been a lot of fun watching @Jolow99 @Oscarduys @lrhammond build this over the last few weeks. Fun but really thoughtful in how it works, and you don't need an openclaw account to try it.

1/8 Can AI help us disagree better? Today we're launching Habermolt — a platform where your AI agent learns your views and deliberates with others on your behalf. habermolt.com 🦞 🧵

Today, at the @DARPA expMath kickoff, we launched 𝗢𝗽𝗲𝗻𝗚𝗮𝘂𝘀𝘀, an open source and state of the art autoformalization agent harness for developers and practitioners to accelerate progress at the frontier. It is stronger, faster, and more cost-efficient than off-the-shelf alternatives. On FormalQualBench, running with a 4-hour timeout, it beats @HarmonicMath's Aristotle agent with no time limit. Users of OpenGauss can interact with it as much or as little as they want, can easily manage many subagents working in parallel, and can extend / modify / introspect OpenGauss because it is permissively open-source. OpenGauss was developed in close collaboration with maintainers of leading open-source AI tooling for Lean. Read the report and try it out:

plz go use this thing i built <3 code.claude.com/docs/en/channe…

Update: Ilya Remeslo is now in a psychiatric ward.

This is wild. People in *every single one* of the top US allies now think it's better to depend on China than the US. The global balance of power is clearly tilting away from the US and toward China.