The NanoGPT speedrun record fell to 122.2 seconds, 5.5 seconds under the 127.7-second mark reported four days earlier, in a post by AWS engineer Larry Dial (X), which credits the run to Chris McCormick’s modded-nanogpt PR #177. His aside that “for some reason the rate of records is increasing” is the reason the item sits in the recursive-self-improvement strand as well as under price deflation: the pace of improving the training loop is itself accelerating. The record lasted a day — 119.3 seconds followed on the 26th and 116.4 on the 27th — and by May agents handed idle compute beat the human baseline on the speedrun’s optimizer track.
Development
The NanoGPT speedrun record falls to 122.2 seconds
The training loop is not just getting cheaper but getting cheaper faster: an observer's remark that the rate of records is rising is what moves the speedrun from a price story into the self-improvement strand.
Claim
The NanoGPT speedrun training record dropped to 122.2 seconds, 5.5 seconds faster in four days, with observers noting the rate of records is itself increasing.
Reported metric: 122.2 s
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to December 25, 2025 (theinnermostloop.substack.com)
- NanoGPT speedrun record: 122.2 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.