George Hotz: humanoid robots are 'as dumb as self-driving cars was,' AGI is a meaningless buzzword, and Waymo has more than 1 human per car

Summary

George Hotz treats AGI as a meaningless marketing term, comparing it to rebranded Excel and arguing that every Turing machine is by definition a general-purpose computer. The real metric he tracks for any administration or economy is two numbers only: government spending direction and total electricity production. On China versus the US, his thesis is blunt — China leads in nuclear, hydro, coal, and virtually every infrastructure category not because of any specific policy but because the US has stopped building things, a disease of "developed" status.

AI Industry Structure

Hotz sketches a five-tier AI stack where value concentrates at the infrastructure and foundation-model layers. Electricity, data centers, and land sit at the base. Fabs — TSMC, ASML, Samsung — occupy tier two. OpenAI and Anthropic sit at tier four. Everything above them — Cursor, Windsurf, Character AI — he dismisses as value-free, drawing a direct parallel to Zynga and app-layer companies that Facebook and Google hollowed out during the web era. The structural reason is compute economics: unlike Gmail, where one server serves 10,000 users, AI has no ceiling on quality demand and no ability to pool compute efficiently. He estimates he personally costs OpenAI significantly more than the $200 monthly subscription he pays, citing Codex spinning up four parallel nodes on demand.

On scaling laws, his framing is deliberately deflating: you invest exponentially more capital for linear output gains. Open-source models are not far behind frontier labs, which commoditises the application layer faster than most investors assume. Elon Musk appears to understand this, which Hotz reads as the reason behind the push for a dramatically larger data center.

Self-Driving and Waymo

Hotz puts Waymo's teleoperator ratio at above 1.2 humans per car, not below one. His evidence is circumstantial but pointed: no one has ever seen a photograph of Waymo's remote operations center. If the ratio were genuinely one operator supervising ten cars, Google would be publishing those images widely. Cruise litigation documents reportedly showed approximately 1.5 to 1.7 humans per vehicle. He frames the commercial value of a Waymo ride as essentially paying a premium not to make conversation, since the human-to-car ratio is barely better than a standard Uber.

Tesla FSD he describes as the only genuinely unsupervised self-driving system available today, and he would not take a five-minute nap in one. Comma.ai he positions as running two years behind Tesla, with cost as its competitive advantage. His broader forecast for autonomous vehicles is a scooter-market dynamic — capital-driven, easily replicable, and destined to consolidate after a race to the bottom among ten-plus providers.

Google's original 2012 claim that a 12-year-old would never need a driver's license he cites as exhibit A in a pattern of repeated hype cycles that the humanoid robotics wave is now replicating.

Humanoid Robots

Hotz is openly contemptuous of the humanoid robot category, calling it "as dumb as self-driving cars was" and singling out Kyle Vogt by name as someone who should know better than to raise large amounts of capital for the category. His factory-floor argument is practical: what manufacturers need is two robotic arms fixed to a table, not a bipedal unit that needs legs to navigate a space already designed around human movement. Off-the-shelf arms are already mechanically sufficient; the bottleneck is entirely software. The humanoid form factor he reads as a science-fiction aesthetic preference rather than an engineering requirement.

The core software problem blocking both robotics and autonomous vehicles is a data efficiency gap of roughly 1,000x. Current ML models require terabytes of training data to learn tasks that humans master on megabytes of lived experience. Until that gap closes, simulation-to-real transfer remains a workaround rather than a solution, useful for resetting physical states thousands of times but not a fundamental fix. Hotz argues this efficiency problem has to be understood before reinforcement learning in the physical world can scale meaningfully.

AI Risk and Manipulation

On the safety debate, Hotz sidesteps institutional framing entirely and offers a concrete threat model instead: the primary AI risk is ten hyperfast, dedicated agents running at 1,000x real time, optimising to manipulate a single individual — for purchasing decisions, political preferences, or anything else. He notes the largest AI companies are fundamentally advertising businesses, and advertising is manipulation. The "AI doom" discourse he treats as vague and unactionable, while the manipulation vector he considers concrete and near-term.