The speedrun record falls to 113.7 seconds

Larry Dial’s record log credits the 113.7-second run, 1.4 seconds under the previous mark, to Chris McCormick’s pull request #187 on modded-nanogpt, which centralizes per-layer parameters in a parameter bank, tunes Adam, raises EMA-buffer precision and moves the gate matrices from Muon to Adam (X post). It is the sixth record time in the corpus’s NanoGPT speedrun chain (127.7 → 122.2 → 119.3 → 116.4 → 115.1 → 113.7) and the first set after the generalization result showed the track’s tricks transfer to a harder target. The mark stood three days before falling to 109.2 seconds on a new attention scheme, and in May agents given idle compute beat the human baseline on the speedrun’s optimizer track under recursive self-improvement.