Llama 4 arrives, and the leaderboard problem gets a name
Scout and Maverick ship with a 10M-token context claim and an arena entry that wasn't the released model.
tech, developers, and the code underneath
older dispatches from the README archive.
Scout and Maverick ship with a 10M-token context claim and an arena entry that wasn't the released model.
You can finally coerce dyn Sub to dyn Super. Also: HashMap::get_disjoint_mut, and a soundness fix worth reading.
Native image generation in a chat model produced a two-week global aesthetic event and a copyright question nobody wants to answer.
A reasoning model with a million-token context that finally makes the long-context pitch concrete.
Anthropic's Model Context Protocol gets its most important endorsement four months after release.
Faster fuse, better hardware support, and a Rust for Linux effort that keeps advancing through genuine disagreement.
Blackwell Ultra this year, Rubin next, Feynman after. Nvidia is now publishing a schedule like a foundry.
A new t prefix produces a Template object instead of a string. It's f-strings without the injection vulnerability.
Blue-green, canaries, feature flags, and the deeply unfashionable practice of shipping the same way every time.
Not a tutorial. A short list of the operations that separate people who fight git from people who use it.
M4 MacBook Air, M3 Ultra Mac Studio, M4 Max Mac Studio. Apple's chip naming has fully decoupled from chronology.
--erasableSyntaxOnly exists because Node can now run TypeScript, and enums can't be erased.
A very large non-reasoning model arrives at very large prices, and mostly demonstrates why the field moved on.
A hybrid reasoning model plus a command-line coding agent in research preview. The CLI is the more interesting release.
New match ergonomics, gen blocks reserved, RPIT lifetime capture changes, and unsafe attributes. Here's what will break.
xAI's third model arrives on a cluster built in months. The interesting claim isn't the benchmark, it's the schedule.
A practical model for supply chain risk that doesn't require pretending you read 40,000 files of transitive JavaScript.
The extension platform migration everyone fought about for six years reaches its end state.
A twenty-minute pipeline doesn't cost twenty minutes. It costs the context switch, the batching, and the review you skipped.
OpenAI ships a browser-using agent as a research preview. The demo is impressive; the failure modes are the interesting part.