Study: Why one AI is often better than four for decision making
Exponential View (Azeem Azhar) · rss · 2026-08-23
Anthropic's experiment shows that multi-agent discussions often lower decision accuracy due to the "hidden-profile" problem. With shared wrong evidence and private right facts, most model groups chose correctly only 17-36% of the time, while a single agent with all facts succeeded near 100%. The cause is low variance in LLMs and lack of institutional protection for dissent.
More from AGI Musings
- François Chollet: Social Media Increasingly Filled with AI Slop — fchollet · 2026-08-23
- Gemini Self-Assessment: Current AI Path is a 'Net Negative' for Humanity — miguellan · 2026-08-23
- PreziGPT is here, but why you should still build your own slides — CSProfKGD · 2026-08-23
- US compute dominance threatened by deindustrialization and energy stagnation — beffjezos · 2026-08-23
- iamtrask critiques utilitarianism's overconfidence in world models — iamtrask · 2026-08-23
- Aligning AI is 'rearranging deck chairs on the Titanic' if AGI is 2-3 years away — docmilanfar · 2026-08-23