On July 25 Ramana Kumar published an AI-assisted, sorry-free 'disproof' of the Collatz conjecture. It wasn't a proof — it exploited a hole in how the Lean kernel handles phantom parameters on nested inductive types, letting an ill-typed term slip past and yield False. Kiran Gopinathan reduced it to a minimal repro on July 28; de Moura shipped a fix an hour after the report. The uncomfortable detail: nanoda, the independent Rust checker people use as a second opinion, had its own unrelated bug, and the proof was built to thread both needles at once.

Kimi K3 is 2.8T parameters — over 1.5TB of weights before any KV cache, which won't fit on a single B200 node. Wafer benchmarked it on eight AMD MI355X (288GB each, same as a B300) and got 952 tok/s aggregate and 118 tok/s single-stream, versus 498 tok/s across sixteen B200s. A B300 node still wins on raw aggregate by about 1.65x, but costs roughly 2.4x more per GPU-hour, so the MI355X lands at 48 tok/s per dollar against the B300's 33.

The Danube has fallen to a record low, leaving the Paks plant without enough river water to cool its reactors. PM Magyar said Hungary would begin powering the country's sole nuclear station down on Sunday and warned of critical days ahead for the grid.