Wire · opportunities
AI Agents Beat Human Researchers, But Elie Bakouch Says They Still Can't Invent
◆ Sectors
◆ Geography
◆ Source
◆ Verified
Fusion42 · 26 September 2026 · Fusion42 review
Prime Intellect's Elie Bakouch tested OpenAI's Codex and Anthropic's Claude Code on an optimizer speedrun benchmark, where both AI agents beat human records but failed to invent genuinely new optimization methods, instead making incremental improvements. This highlights a key limitation in AI's current ability to autonomously discover novel research methods despite excelling at optimisation tasks.
This Wire brief sits within Fusion42's coverage of AI Agents and AI Frontier Models.
◆ ◆ The Wire takeaway
Your AI-enhanced research tools can now outperform humans on optimisation tasks but won't invent new methods yet. Shift focus from relying on AI to invent research breakthroughs towards integrating it as a powerful optimiser that accelerates existing techniques.
◆ Coverage
1 source · 26 Sep 2026
◆ Related on Wire
◆ Topics