Intel / video

We Let Claude Code and Codex Race Human Researchers — Elie Bakouch, Prime Intellect

Source: youtube.com
Author: AI Engineer
Date: Published Sep 26, 2026

From the Daily Brief

Elie Bakouch, a research engineer at Prime Intellect and creator of SmolLM, pointed Claude Code and Codex at the community Optimizer Speedrun, a race to train a GPT-2 level model in the fewest steps. Both agents beat the human record. Their working styles diverged sharply: Claude Code stopped every nine or ten hours to say the record could not be beaten and sat idle about a third of the time, while Codex never stopped, wrote more notes, spawned more sub-agents and spent more tokens. In a longer six-day run that added Kimi and GLM, Kimi was the most token-efficient, and a paper only Claude found produced the best record. The sobering result is that no model invented a new optimizer. Bakouch treats this as evidence against loose claims of recursive self-improvement and says Prime Intellect is building an AlphaEvolve-style loop aimed at real discovery, developed in the open.

Read the 2026-09-27 Brief →

Context

Coding agents reportedly improved a model-training record by combining existing techniques. In his Optimizer Speedrun account, researcher Elie Bakouch describes giving Anthropic's Claude Code and OpenAI's Codex a task: train a GPT-2-level model in fewer steps. He reports that both surpassed the human record.

Bakouch also reports that none of the tested models invented a new optimizer, the rule used to update a model during training. Their progress came from finding and combining known ideas. That makes the experiment evidence for automated research iteration within a defined task, while leaving the harder question of original algorithm discovery open.

Key takeaways · AI-distilled

Bakouch set Claude Code and Codex on the community Optimizer Speedrun, a race to train a GPT-2-level model in the fewest steps, and reports that both agents beat the human record.
The agents worked very differently: Claude Code stopped every nine or ten hours to say the record could not be beaten and sat idle about a third of the time, while Codex never stopped, wrote far more notes, spawned more sub-agents and burned more .
In a longer six-day run that also included Kimi and GLM, Kimi was the most token-efficient, and a paper that only Claude found led to the best record.
His sobering finding: no model invented a new optimizer; they combined existing ideas for small gains. Prime Intellect is building an AlphaEvolve-style loop aimed at real discovery and argues this research should happen in the open.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source www.youtube.com