Naval amplifies argument: AI labs selling fixed-cost tokens to everyone will fail
naval · x · 2026-10-07
Naval reposted and endorsed @lefttailguy's argument that labs whose long-term strategy is selling universal access to fixed-unit-cost tokens are doomed to fail.
The original thread argues labs' "DeployCo" arms should do two things:
- Return locator: build a predictive model of which pockets of the economy yield the highest returns to frontier intelligence at any time, guiding capital (compute) allocation.
- Deployment meta-learner: a continual-learning system that turns each deployment into a training case, automating the conversion of tokens into economic value and capturing high-value areas faster than rivals.
The upshot: the moat lies in deployment-side search and learning, not the model itself.
More from AGI Musings
- After 12+ hours a day with AI, she writes again: it's not just attention loss, it's memory — LadyAshBorg · 2026-10-07
- AI usage inflects exponentially past capability thresholds — is research next? — menhguin · 2026-10-07
- A Decade-by-Decade Tech Recap: 2010s Delivered Nothing, 2020s Bring AGI — Darpinian · 2026-10-07
- Katja Grace: Being 'in the room' at AI labs may no longer be a high-impact path — michael_nielsen · 2026-10-07
- Bessemer's Dholakia: AI's most important uses aren't invented yet, like oil in 1870 — rohanpaul_ai · 2026-10-07
- He spent 24 years on Barnette's Conjecture — then AI solved it as problem 180 — Simon Willison · 2026-10-07