Prime Intellect Ran 153 Autonomous Research Runs Across 18 Frontier Models and Closed 82% of the nanoGPT Human Record Gap

Prime Intellect / Hacker News·high signalXBlueskyLinkedInCopy link

Prime Intellect published the NanoGPT Speedrun Frontier, which it calls the largest open experiment on how frontier models do AI research: 153 autonomous runs across 18 frontier models, each sandboxed on 8xH200s for up to eight days with no internet access, iterating only on optimizer hyperparameters for a 124M GPT training recipe from a shared baseline. The best runs closed 82% of the gap to a record dozens of humans built over months; the leaderboard's top entry is Fable 5 at 2,726, listed as 81.7% closed in 8.7 days on 800M tokens. Code is on GitHub at PrimeIntellect-ai/experiments-autonomous-speedrunning.