George Hotz, a hardware shipper who has spent a decade putting AI inside physical products, is making a specific falsifiable claim against hard takeoff forecasters (who predict a sudden superintelligence transition): the binding constraint is
The AI hard-takeoff debate just got its most concrete engineering rebuttal yet.
George Hotz — founder of comma.ai, who has spent a decade shipping self-driving hardware into the physical world — published a post on July 11 titled "AI 2040 and the Cult of Intelligence." It is a direct reply to AI 2027, the scenario document that has become the baseline reference for the next decade of AI policy forecasting. His core argument: the scenarios solving for superintelligence are solving the wrong equation.
The mechanism he names is the correlation effect — the assumption that sufficiently advanced AI will exert direct control over physical systems the way digital systems exert control over each other. Hotz's case is empirical. A chip takes three months to fabricate. A reflow oven warps circuit boards. An ocean datacenter accumulates barnacles. The wrong part ships from a supplier and there is no rollback. Air-freighting from China costs orders of magnitude more than waiting three weeks on a boat. These are not edge cases. They are the floor of physical-world production, and no language model trained on tokens can bend them.
He traces the intellectual lineage: Yudkowsky's hard-takeoff premises, and [Roger Williams's 1994 novel The Metamorphosis of Prime Intellect](https://grokipedia.com/page/The_Metamorphosis_of_Prime_Intellect), which he reads as a story whose hard takeoff works only because a quantum-trick stipulation makes it work — fiction by construction, not physics. The AI 2027 document, in Hotz's reading, is a self-fulfilling belief artifact — a "Plan A" toward expanded state authority over compute, analogous to FDR's 1933 gold-confiscation order, not a description of what actually happens when intelligence meets atoms.
Hotz's alternative he calls Plan L — local, user-aligned AI that refuses to coordinate at scale. The framing is deliberately provocative; the structural point is not. Coordination requires infrastructure. Infrastructure has lead times. Lead times resist compression.
Call it the atoms bottleneck: hard-takeoff scenarios treat intelligence as the binding constraint. Hotz argues physical reality already has a binding constraint, and it is not getting faster on a software schedule.
Hotz previously published "There Is No Hard Takeoff" in August 2023.