Dan Shipper's 'Most Important Formula in AI': human attention is the new bottleneck
danshipper · x · 2026-11-13
Dan Shipper and Every's head of evals Mike Taylor propose the formula value(q) = P(engage|q) × I(θ;A|q): the value of any question an AI asks you equals the odds you'll actually answer it, times how much your answer updates the model's knowledge of your taste.
- The bottleneck in personalization is no longer intelligence but human attention — getting taste into a model without turning users into resentful graders.
- The two factors multiply, so half-fixes fail: easy thumbs-ups get answered but barely update the model; long labeling sessions teach a lot but users close the tab.
- The only way to raise answer rates is for the AI to give value back, making the reward of answering exceed the cost of thinking.
More from AGI Musings
- Oxford Intersections reopens call on AI & Relationships, abstracts due Dec 31 — dioscuri · 2026-11-13
- Leslie Pack Kaelbling to Debate Pure Behavior Learning vs Runtime Rationality in Robotics at Vector Institute — VectorInst · 2026-11-13
- Economists crunch METR data: no convincing evidence of an intelligence explosion yet — soumitrashukla9 · 2026-11-13
- Professor asks Claude to price a bare-bones college: still $20,389 a year — _akpiper · 2026-11-13
- As agentic AI hits superhuman speed, 'meaningful human oversight' may become an illusion — LuizaJarovsky · 2026-11-13
- Skeptic view: RL scaling is just compiler feedback plus IP siphoning — yacineMTB · 2026-11-13