Keller Jordan

Keller Jordan maintains the modded-nanogpt repository, which hosts the NanoGPT speedrun: a collaborative-competitive search for the fastest algorithm to train a GPT-2-class model to a fixed FineWeb loss on eight H100s, whose README credits the Muon optimizer he wrote among the techniques behind the record (repository). He appears in the corpus as the benchmark’s publisher rather than as an actor in any single development; the records themselves, such as the 122.2-second run, are filed by the contributors who set them.