AI Verifiability Tiers: Coding is Easy, Human Preference is Hard
ruthstarkman · x · 2026-08-03
Max Spero categorizes AI capability verification into three tiers:
- Programatically verifiable: Near-free tasks like games, coding, math, and chip design, expected to be solved quickly.
- Real-world verifiable: Cost- or time-bounded tasks in sciences like biology, chemistry, and forecasting.
- Human preference verifiable: Subjective domains like writing, design, and persuasion.
Seth Lazar adds that the third category involves socially constructed and dynamic values, making them unstable targets for unilateral verification. As our understanding evolves, we may empirically discover that some targets belong to different buckets than initially assumed, offering interesting insights into moral philosophy.
More from AGI Musings
- Chinese Models Hit 45% of OpenRouter Traffic as AI Commoditizes — LexSokolin · 2026-08-03
- From Solving Problems to Refuting Conjectures: LLMs Reshape Math Research — lvwerra · 2026-08-03
- Anthropic and OpenAI Could Become Compute-Dominant Cash Machines — toptickcrypto · 2026-08-03
- NeurIPS 2026 Workshop Call for Papers: Dynamic Alignment in Human-AI Coupled Systems — huashen218 · 2026-08-03
- AI Data Centers Consume 1.5B Gallons of Water Annually, Sparking Environmental Debate — ZeroStateReflex · 2026-08-03
- Statistician: Concerns Over Whether Research is AI-Generated are "Unserious" — kareem_carr · 2026-08-03