The Quiet Day That Rewrote a Programming Language

The Great Bun Rewrite

The headline story is a jaw-dropper. Jarred Sumner, creator of the high-performance JavaScript toolkit Bun, rewrote the entire project from the Zig language into Rust in just 11 days using a pre-release Claude Fable 5, running roughly 50 automated workflows with up to 64 AI instances in parallel. That's over 535,000 lines of code, a job he estimates would take three engineers a full year by hand. The API-equivalent usage came to around $165k. The community's sharp counterpoint: the real enabler wasn't the spend but Sumner's deep expertise and full command of the codebase, with the model acting as an accelerator rather than a replacement.

Superhuman at the Keyboard

An OpenAI system entered an AtCoder World Tour Finals exhibition and scored around 8,300, nearly doubling the top human competitor's 4,300, and in one contest solved all 5 problems when no human solved more than 3. Observers noted the important caveat: this is algorithm-contest problem solving, a constrained subset of programming, not general end-to-end software engineering. Still, the margin is striking.

From Full Mix to Sheet Music

Kyutai and Mirelo released MuScriptor, an open model that transcribes multi-instrument audio to MIDI directly from a full mix rather than requiring separated stems. It's a genuine gift for musicians, teachers and hobbyists, and being open means anyone can build on it.

Making AI Faster and Cheaper

Inference optimisation kept shipping. Cohere open-sourced a hardware-aware dynamic speculative decoding technique in vLLM that adapts to how busy the system is, fixing the long-standing problem where speculative decoding helps at low load but hurts at high load. Separately, a challenge around Google's Gemma models reported up to 5x faster inference on a single modest GPU with no quality loss.

A Giant Model on a Tiny Machine

A developer demonstrated GLM-5.2, a 744B-parameter mixture-of-experts model, running on a consumer machine with only 25GB of RAM by streaming expert weights from disk instead of holding the whole model in memory. It's painfully slow today, more proof-of-concept than practical, but if expert-routing prediction improves enough to prefetch the next experts, the whole picture could change.

Agents That Remember and Check Their Work

Two research threads targeted reliability. Meta researchers proposed giving agents an explicit memory to combat behavioural state decay on long-horizon tasks, while an "LLM-as-a-Verifier" approach set new records across Terminal-Bench, SWE-Bench Verified, medical and robotics benchmarks by generating multiple attempts and ranking them. The through-line of the day: the quiet engineering matters as much as the flashy launches.