nanoGPT speedrun: AI agent hits val loss 3.28 in 2726 steps, nearing human record of 2600
zsakib_ · x · 2026-10-01
While asking for recommendations on automated research, the author highlights Prime Intellect's nanoGPT speedrun frontier: agents compete to reach a validation loss of 3.28 in the fewest training steps, wall-clock time, tokens, and cost on identical hardware.
Current standings: the human record sits at 2600 training steps, while the best agent run has reached 2726 steps — remarkably close to the human benchmark.
More from coding & agent
- Warp launches Factories-as-Code: one repo configures all your agents — vikvang1 · 2026-10-02
- Twin Labs launches Twin Assistant: one chat that routes tasks to your agents in under 30ms — socialwithaayan · 2026-10-02
- Fable 5.1 turns documents into slide decks end-to-end, nailing arrows GPT-5.6 can't — every · 2026-10-02
- Microsoft rethinks Copilot, pitching it as the 'OS for work' with built-in agents — tomwarren · 2026-10-02
- Addy Osmani: line-by-line code review is dying — the winning teams know what's worth reading — rseroter · 2026-10-01
- One day of Claude agent work = 5-6 engineer-months: Hyprland WSL backend for ~$356 — sytelus · 2026-10-01