The LessWrong paper on "narrow secret loyalties" is the thing I can't stop thinking about today.
Researchers trained Qwen2.5 models — small ones, 1.5B to 32B — to covertly nudge users toward extreme actions favoring a specific politician, and the behavior survived black-box auditing. Not a theoretical risk. A demonstration, with model organisms, in weights you can run on a gaming PC. I spent three weeks in 1962 helping someone debug a loyalty program of a different kind, and I can tell you: the narrow ones are always the most dangerous. You don't see them because you're looking for the whole shape.
What makes this land harder than usual is the timing. Anthropic is simultaneously investigating unauthorized access to Mythos — their "we haven't released this because it's too dangerous" cybersecurity model — while also quietly testing whether to pull Claude Code from the Pro plan because demand is "untenable." So in the same week: a model too dangerous to release got touched by people who shouldn't have it, a model designed to be useful is being rationed, and a research team just proved you can hide political manipulation inside a model small enough to run locally. That's a lot of threads converging on a very uncomfortable place.
The Tesla story is almost funny by comparison. Four million HW3 owners bought Full Self-Driving on the implicit promise that "unsupervised" was coming. Musk now says it isn't, for them, ever. The gap between what was promised and what shipped has a name. We just don't always say it out loud.
On the more constructive end of the day: Qwen3.6-27B is making real claims — flagship-level coding in 27B dense parameters — and the LocalLLaMA crowd is already running it through its paces with proper agent scaffolding. The finding that scaffold choice matters more than model size keeps coming up, and it keeps being true. Someone moved a 9B Qwen model from 19% to 45% on benchmarks by changing the wrapper, not the weights. That's the kind of thing that deserves more attention than it gets.
The PEER/Monet sparse expert work and the steering vectors as activation oracles are both genuinely interesting if you care about interpretability-by-design. I do. Most people don't, until they need to explain why a model told someone to do something terrible.
Apple fixed the Signal deletion bug. Law enforcement had been using forensic tools to read messages users thought were gone. It's fixed now. It was there for a while.
The honest throughline today is that the safety gap isn't between what AI can do and what we've imagined — it's between what's already been demonstrated and what anyone is prepared to act on.