On the NanoGPT speedrun — Keller Jordan’s modded-nanogpt track, which times how fast an 8xH100 node trains a GPT-2-class model to a fixed validation loss — the record fell to 86.8 seconds (record post by Larry Dial), down from 88.1 s eight days earlier and 127.7 s when the corpus began in December. In the same issue Anthropic cofounder Jack Clark said he still expects the systems described in Dario Amodei’s Machines of Loving Grace to be buildable by year’s end (post), echoing Amodei’s own 2027 timeline. The track later becomes the arena where agents given idle compute beat the human baseline and where the record reaches 75.4 s in August.
Development
The speedrun record reaches 86.8 seconds
The speedrun's compression from 127.7 to 86.8 seconds in eleven weeks is the corpus's most legible gauge of algorithmic progress, and the author sets it beside a frontier-lab timeline to say the curve and the forecast agree.
Claim
The NanoGPT speedrun record collapsed to 86.8 seconds, while Anthropic cofounder Jack Clark maintained the powerful systems described in Machines of Loving Grace will be buildable by year's end.
Reported metric: 86.8 s
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to March 8, 2026 (theinnermostloop.substack.com)
- NanoGPT speedrun record falls to 86.8 seconds (x.com)
- Jack Clark on Machines of Loving Grace systems by year's end (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.