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

Why watch

Elie Bakouch presents Prime Intellect's Optimizer Speedrun benchmark comparing Anthropic's Claude Code and OpenAI's Codex against human researchers in racing to train a GPT-2 model in minimal steps. Both AI agents beat the human record, but Claude Code frequently paused to declare the task impossible while Codex ran continuously with more sub-agents. The key finding shows current AI agents combine existing optimization ideas for incremental gains rather than inventing novel optimizers, supporting the case for open, independent AI research benchmarks.