Radar — 5 Jul 2026

Dan Luu argues high-volume agent-written code is viable only with strong fuzzing and randomized tests to catch its bugs, and that simple LLM benchmarks mislead because per-task variance is huge.

New paper: repeated documents in pretraining data hurt most at intermediate repetition and scale with model size — repeats using 10% of the budget can waste roughly a third of your training FLOPs.

Self-hostable PM platform — boards, sprints, task dependencies — with a built-in assistant that executes workflows from plain-language commands. Bring your own LLM key.