The record, 2.3 seconds under the 109.2-second mark of four days earlier, came from GitHub contributors andrewbriand and jrauvola, who noticed that the torch compiler was running a separate, inefficient kernel for the ReLU(x)^2 activation and wrote a Triton kernel fusing it into the preceding linear op (X). Earlier links in the NanoGPT speedrun chain came from optimizers and attention variants, and the January 2 result showed those gains transfer; this one is pure kernel work, the layer where a model later tops the leaderboard for writing its own kernels. The record falls again to 99.3 seconds twelve days later, and by May agents beat the human speedrun baseline outright.
Development
The speedrun record falls to 106.9 seconds on compiler kernel hacking
The chain's gain comes from below the model, in the compiler and kernel layer rather than from architecture, the layer the corpus's later kernel-writing agents work in.
Claim
The NanoGPT speedrun record dropped to 106.9 seconds through compiler kernel hacking.
Reported metric: 106.9 s
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to January 12, 2026 (theinnermostloop.substack.com)
- Larry Dial on X: New NanoGPT Speedrun WR at 106.9 seconds (x.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.