A training speedrun record falls to 75.4 seconds

The NanoGPT speedrun record fell to 75.4 seconds on a faster ReLU^2 MLP Triton kernel (post on X), down from 86.8 s in March and 127.7 s when the corpus began tracking it in December. The X post and the speedrun’s own record table credit the record to @cong_ml and an AI system called Recursive from Recursive Superintelligence, the lab that raised $650M in May to have AI experiment on improving itself, so the record carries an AI-system co-author. The gain comes from the kernel layer, where compiler kernel hacking had already set a record in January and where a model rewrote the production kernels behind OpenAI’s price cut three days earlier, and it lands after agents given idle compute beat the human baseline on the speedrun’s optimizer track. In the same item ByteDance’s Seedance 2.5 began generating 30-second audio-video in one pass with multi-minute extensions and timestamp-level edits (ByteDance Seed).