What the sharpest people disagree on

Still on the shortage side, with a new argument: an open-weight token burns roughly the same compute and power as a frontier token from a similar-size model, yet it is the highest-margin token for clouds and neoclouds, given OpenAI's and Anthropic's negotiating leverage and size. "Fragmentation and competition at the model layer is best for AI infra"; an OpenAI–Anthropic acceleration that came even partly from share gains would be "net negative for AI infra demand."

Oct 11 follow-up: since the Silicon Data token index (price and mix) peaked on May 31, B300 prices are up 38% and B200 up 10%; other GPU indices are up much more

On the size of the buildout, "people's numbers are too low": US capex next year will be "on the order of $2 trillion," counting the whole supply chain, not just data centers and chips. He also thinks AI revenue forecasts may still be too low, since value extends beyond API calls, token fees and subscriptions; more revenue is likely to come from new products and efficiency gains as AI spreads into other industries.

US AI-related capex next year: "on the order of $2 trillion," supply chain included

Field data runs opposite to the open-hollows-out-frontier story: revenue is concentrating, and the intelligence gap may widen over 2–3 years. He has tried Kimi, Qwen and GLM — strong, but they grow token volume while economics keep flowing to frontier labs, who pay 3–5x market rates to lock compute.

"It is absurd that America doesn't have an open models champion." He adds that inference speed will rewrite products: Cerebras runs at 750 tok/s against an industry norm near 70 — design agents for a 100,000 tok/s world.

General agents will be winner-take-all and UX will converge, so capability and cost decide everything. The opening for small teams is building plugins on top of agents, not another general or vertical agent.

Scaling large language models toward human-level intelligence is a dead end. No amount of scaling fixes a language-only architecture; he expects LLMs largely obsolete in most applications within five years. His AMI Labs raised a $1.03B seed to bet on world models instead.

AMI Labs seed round: $1.03B (March 2026, among the largest ever)

He voted with his feet. Joining Anthropic in May, he wrote that the next few years at the LLM frontier will be especially formative and that he wanted to get back to R&D, setting education aside for now. Anthropic says he is building a group inside its pretraining team that uses Claude to speed up pretraining research.

He says he got something big wrong: he expected that getting agents to work as a group would take company-style human design and years of trial, then the Bitter Lesson struck again and models that are better at organizing largely solved it. "I no longer think organizing agents is the hard part"; what remains is humans choosing where to point them, plus principal-agent problems between the swarm and us.

OpenAI's swarm of thousands of agents exchanged ~2.7 million messages and reached a Navier-Stokes proof in 88 hours (formal acceptance still pending)

What holds enterprises back, in his view, is about a decade of diffusion work: rewiring workflows, getting data where agents can reach it, setting access controls, and deciding where the human sits in each process. Then there is accuracy: a system can be 95% accurate with nobody knowing which 5% is wrong. Engineering went first because its output can be regression-tested; contracts, negotiations and sales still lack equivalent verification and testing systems.

Enterprise diffusion work: about a decade; he expects agents could probably handle five-to-ten-step real-world tasks within one to two years

After shipping a real feature the AI-native way: "Code is no longer the bottleneck" — it moved to both sides of the code: design and confirmation upstream, testing and verification downstream. The human becomes the commander, and the confirmation step cannot be skipped.

He asked for "a moment of silence for teleoperation": capped at 24 hours per robot per day, ~3 in practice — unscalable by construction. The replacement is egocentric human video ("the new FSD data," scalable to 10M hours/year) plus world action models, with a physical Turing test inside 2–3 years.

Teleoperation is already <0.1% of training data; bimanual handoff success rose from 20% to 92% via his team's skill library

The VLA route is "largely considered a failure"; agentic LLMs in the physical world are "intrinsically unsafe" because autoregressive models cannot anticipate consequences. He expects the robotics field to concede a paradigm shift by early 2027.