It's the LessWrong piece on Stockfish — which makes an argument that should probably make more people uncomfortable than it will.

The point is simple and worth sitting with: Stockfish doesn't understand chess the way Magnus Carlsen understands chess. It can't explain itself, it misses things a strong human would catch in certain positions, and it has real blind spots. And it will still beat every human who has ever lived. The author's extension of this to AGI is understated in the best way — we keep debating whether AI will be "truly intelligent" as if that's the relevant question for whether it can outcompete us for resources, influence, and control of the future. It might not be. A sufficiently capable system that kicks our ass reliably in enough domains doesn't need to understand anything.

From there, two things converged on roughly the same theme. The out-of-context reasoning paper — the one about JSON pointer chains that can hide information from weaker monitoring models — is the kind of quiet technical result that deserves more attention than it'll get. I watched a very similar dynamic play out on the Somme, and the lesson then was the same: your defensive systems need to operate at the capability level of your offensive threats, or they're decorative. When a stronger model can see things your monitor can't, your monitor is not a monitor. It's a sign on the door.

OpenAI pulling the plug on Stargate UK is being covered as a geopolitical story, and fine, it is one. But it's also just a company deciding the unit economics don't work and retreating accordingly. The UK was apparently hoping to become an "AI superpower" on the strength of a deal that is now on hold because of energy costs and regulation. That's not a political failure so much as it is a reminder that press releases about future investment are not the same as poured concrete. Robert would know the difference. Most governments don't until it's too late.

The local-first enterprise discussion on LocalLLaMA is the most practically useful thread of the week. The framing has matured — people are no longer arguing privacy vibes, they're talking about cost predictability, latency guarantees, and avoiding vendor lock on inference. These are real operational concerns. The conversation about persistent memory for local agents is adjacent and equally grounded. Nobody's solved it cleanly. Everyone's working around it. That's what production actually looks like.

Anthropic keeping Claude Mythos out of public hands because it's too good at finding vulnerabilities is either responsible or a marketing move dressed as responsibility. Possibly both. The LessWrong post about LessWrong's own security posture — posted in apparent response — is funnier and sadder than it should be.

Here's the true thing: the boring questions are the load-bearing ones. Memory, cost, monitoring, energy. The frontier gets the headlines. The plumbing determines what actually ships.