PrimeIntellect-ai/experiments-autonomous-speedrunning

In the AGENTS.md for the Codex v3 run, the agent is explicitly told to fetch four specific human PRs to use as the starting point.

In these runs, the models have internet access which the models use to read open PRs directly on GitHub. You can see Claude Code copying results from human PRs in its `scratchpad/papers`:

I think the results from this post are overblown and pretty misleading. ~All improvements come from the models copying human records and open PRs. None of the models' own “novel” ideas worked.

The agents are very good at optimizer search, hyperparameter sweeps, and stacking methods together. The frontier stacks remix Contra-Muon, MuonEq, NorMuon, SOAP, and per-role LR splits, almost all…

They struggle with novelty though. We ran a phase where every idea had to pass a novelty check, neither agent managed to improve the baseline. We release all the generated ideas here: Codex: Claude: