The root problem is that coding agents treat tool output as trusted context.

Good idea. Worth considering though, philosophy isn't really about right answers, it's about learning to question. Bit worried about an LLM tends to hand kids a tidy conclusion, which is kind of the opposite of what you want.

The FFmpeg/Python step is neat. They avoid stopping on every frame, capture continuous video instead, then split it back out and pick the sharpest frame from each group.

From Atom Computing. The toric code needs non-local connections, which plays to neutral atoms' all-to-all connectivity over fixed-topology superconducting qubits.

Text summary since it's a video: it's a model abstraction layer. You write against a LanguageModelSession, import a Swift package, initialize the model you want, and pass it to the session. Apple ships SystemLanguageModel (on-device), PrivateCloudComputeLanguageModel, and open-sourced CoreAILanguageModel and MLXLanguageModel for local models; Anthropic and Google are publishing their own packages.

Microsoft dropped the legal threat, but it also dropped the phrase "responsible disclosure." The new statement says coordinated vulnerability disclosure instead. That's the term Microsoft itself switched to back in 2010, specifically so researchers who go public wouldn't be painted as irresponsible. Katie Moussouris, who helped make that switch, said invoking "responsible" this time was "the first strike in my book."

Nice writeup. One thing worth adding to the limitations: without vorticity confinement, the Gauss-Seidel projection step quietly dissipates the small-scale curl that makes smoke look like smoke.

The 2001 Fedkiw/Stam/Jensen "Visual Simulation of Smoke" paper added it back as a correction force for exactly this reason. At N=16 it doesn't matter much because the grid itself can't represent fine vortices, but the moment you crank N up the missing confinement becomes visible.

Thank you, that's good to know when I try to move this to a Compute Shader and try to make it part of a game in the future

How big is the lifetime holes thing in practice? On loops the contiguous-interval model spills way more than it should. Wondering if that alone explains most of the YJIT gap.

Hide some of your ballooning AI spend behind your ballooning AWS bill

But then how would I be able to brag about how much I'm doing the AI?

Don't worry, you can still get on linkedin and write a bragging post about how many tokens your engineers are spending.

But you couldn't do it with the Claude app or the recently announced Claude Managed Agents I think.

So the state of AI in 2026: ChatGPT DDoS-lite, Claude the polite one that actually reads the rules, Perplexity maybe shows up, and Google was already in your house.

Claude reading the rules is perhaps the strongest argument for Anthropic being "good so far" I've ever seen.