Kevin Bryan: maximizing raw accuracy is rarely the right way to train AI assistants
Afinetheorem · x · 2026-10-27
Economist Kevin Bryan (with Joshua Gans) shares his paper 'Training AI for When Humans Will Use It' plus an in-browser demo:
- Model: AI predicts, humans choose to trust, pay to verify, or abstain (right = 1, wrong = -L, nothing = 0). The coverage/accuracy tradeoff is chosen by model makers — maximizing unconditional accuracy, as most benchmarks do, is rarely optimal since AI value depends on how it embeds into human decisions
- The framework extends to composite decision functions: a human with access to a second partially-correlated AI, or additional information gathering
- He also demos a 'co-creative' design ideator running fully in-browser with no runtime AI calls: high-variance recombination of billions of design ideas, with humans discarding bad ones and verifying — his recipe for real AI for Science
Related event: Economist Kevin Bryan Builds AI Co-Creative Idea Generator(2 posts)→
More from AGI Musings
- Ethan Mollick: hidden reasoning traces push humans from augmentation to automation — emollick · 2026-10-27
- Predicting sub-50-person tech giants: board, execs, and agents running everything — arthurcolle · 2026-10-27
- Pedro Domingos: Your odds of getting rich just by telling your AI to make money keep improving — pmddomingos · 2026-10-27
- Miles Brundage: If you read S-1s, pivot to reading system cards — Miles_Brundage · 2026-10-27
- Bryan: AI designs far better, but not for free in real time — ideators are new — Afinetheorem · 2026-10-27
- Lao Gan Ma faces boycott over AI replacing 500+ workers, exposing US-China AI sentiment gap — 96Stats · 2026-10-27