The Iteration Flywheel

Automation reproduces the same output to a fixed standard. Intelligence is different: what it produces can keep improving, and each generation becomes an input to the next.

Math and coding have advanced quickly because they offer scalable, verifiable rewards. Coding adds a recursive advantage: better coding agents help build the next generation of models, which in turn produce still better coding agents.

I want to bring this flywheel into physical production: AGI builds better robots; those robots expand experimentation and manufacturing; that capacity helps build the next generation. Build robots that help build better robots.

Embodiment Is a Precondition for AGI

Embodiment is usually framed as a downstream application of AGI. I see it as part of the path to AGI itself, for three reasons:

1. Compute must be built. Reaching AGI’s frontier may require vast populations of digital and physical agents—and therefore enormous compute. But fabs, power plants, and data centers must be physically constructed. Robots close the loop from intelligence to more compute, and back to more intelligence.
2. Science moves at the speed of experiments. In biology, chemistry, and materials, if every hypothesis must wait for human hands and schedules, the entire learning loop remains capped by human experimental speed.
3. The physical world is an open-ended curriculum. Text is largely a static record of what humans have already captured. The world continually produces new states, actions, consequences, and unknowns—an effectively inexhaustible source of learning.