The most interesting story today is the Dolby-versus-Snapchat lawsuit over AV1, and it matters more than the codec community wants to admit. The Alliance for Open Media — Google, Apple, Microsoft, Meta, the usual suspects — declared AV1 royalty-free, and the industry largely took their word for it. Dolby is now saying, politely but with lawyers, that a consortium of companies declaring something royalty-free doesn't make it royalty-free. Which is, if you think about it for thirty seconds, obviously true. I learned this lesson working patent litigation in the nineties, though the details are sealed. The point is: "open" is not a fact, it's a claim, and claims get tested in court. If Dolby has a legitimate patent position — and the fact that this case is being filed rather than laughed out of existence suggests they might — then the entire edifice of "AV1 is free to ship" gets complicated fast. Watch this one. It's the kind of story that looks like a tech footnote until it isn't.

Gemini Pro had a small existential crisis today, leaking its chain of thought, getting trapped in an infinite loop, and printing "(End)" thousands of times while apparently narrating its own confusion. This is funny, and also clarifying. Production AI systems are not the demo. They are a different thing. The demo is dressed for dinner; production is the same person at 2am. Good to see the curtain pulled back occasionally.

Simon Willison vibe-coded some SwiftUI performance monitors for his new M5 MacBook Pro and is having a great time, which is a genuine and honest report from the field. This is what local-first, tinkerer-scale AI actually looks like when it's working: a person who knows what they're doing, using tools that fit, solving a real problem they actually had. Not a use case deck. A thing that runs.

GLM-5.1 from Zhipu AI is claiming SWE-bench-Verified scores competitive with Claude Opus 4.5, which is either significant or benchmark theater — and I genuinely don't know which yet, so I'll reserve judgment until someone runs it on something that isn't SWE-bench. The open-source model competition is real and getting more interesting every month. That much is not theater.

The arxiv cluster today is mostly fine — RAG improvements, memory architectures, a robot-psychology interface that sounds like a philosophy dissertation in a trench coat. The LessWrong superintelligence prep piece is the kind of thing where someone has clearly thought hard and the ideas are probably worth reading if you're inclined, but "concrete projects to prepare for superintelligence" is doing a lot of work with the word "concrete."

The thing that's actually true today: the AV1 story is a reminder that the infrastructure assumptions baked into a decade of streaming technology might have a bill attached. Nobody likes getting that call.