RoboBoom
The Humanoid Hub· November 3, 2026· 47:36

What Does a Post-AGI Physical World Look Like?

Ted, founder of Prometheus, discusses the verification bottlenecks between AGI and physical abundance at a Humanoid Hub event hosted by Devang.

In this talk

  1. 0:38

    Connect digital intelligence to physical progress

    Ted considers a future where both advanced digital agents and general-purpose robots improve rapidly. He argues that neither capability alone addresses the many stages involved in creating physical goods, from ideas and design through production and scale. The opportunity lies in connecting intelligence to measurable improvements in people’s lives.

  2. 6:03

    Verification limits the pace of invention

    Applying Amdahl’s law to innovation, Ted says the share of a development cycle that can be accelerated sets the ceiling on overall speedup. Fast simulation and analytical checks can scale with compute, while real-world tests may take months or years. Battery validation and drug trials illustrate how slow feedback can remain a bottleneck.

  3. 12:17

    Robotics and engineering need different feedback

    Ted contrasts humanoid locomotion trained in simulation, which can transfer to a Unitree G1, with manipulation that still depends on expert data and lengthy lab evaluations. He also describes engineering checks and automated labs as ways to move some feedback earlier, while pointing out that wet-lab experiments and other physical validation remain slow. Periodic Labs’ reported rate of about 100 X-ray diffraction experiments per day shows both the promise and limits of real-world throughput.

  4. 26:20

    Build verifiers that predict real-world results

    Ted’s central recommendation is to develop verifiers that remain trustworthy on the next stages of real-world work. Researchers should create domain-specific evaluations, test whether those evaluations reflect deployment conditions, and allocate scarce, expensive end-to-end trials carefully. He points to physical output and faster invention cycles as better measures of impact than toy benchmarks.

  5. 32:33

    Turn expert know-how into useful feedback

    In the Q&A, Ted acknowledges that subjective expertise, such as an F1 test driver’s feel or a materials scientist’s judgment, is difficult to capture in objective metrics. He recommends starting in real workplaces and learning how practitioners make decisions. He also argues that AI should expand people’s ability to invent, and that physical data collection should distinguish broad pretraining data from targeted, current-policy correction data.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.

More talks