Learn the frontier · Reproductions

Reproductions

We run it in the open so you can learn from it and trust it.

A living, friendly logbook of reproduction efforts. We run the paper, attach the harness, seeds, and hardware, and report clearly what we found — so you can see exactly how a result holds together, build on solid ground, and reuse the work yourself.

RSS ↗1 piece

How we gradethe six verdicts, and what each one promises you

Reproduced: We re-ran it under recorded conditions and reached the reported result within our stated tolerance.
Partly there: Some of it held; the rest fell short, or only reproduced under narrower conditions than claimed.
Didn't reproduce (yet): Run under recorded conditions, it did not reproduce — a finding about this result, not a verdict on the idea.
Contested: The evidence conflicts — a later run or another group disputes an earlier verdict.
In the lab: Being reproduced right now — numbers are still landing and may shift.
On the bench: Queued to reproduce; the run has not started yet.

CHECKPOINT 00102026-06-15In the lab

Reproducing the nanoGPT Speedrun: What Actually Moves the Loss Curve

On a small config the direction reproduces — with the architecture fixed, Muon reaches the 3.28 FineWeb target in markedly fewer optimizer steps than AdamW — but the headline wall-clock record is hardware-locked to 8×H100, and our reference-node run is still pending.

reproducibility pretraining optimization distributed-training