Frontier Systems for the Physical World

• AI shifts from language/code toward physical-world interaction systems.
• Five primitives power it: learned dynamics, embodied action, simulation, sensing, agentic loops.
  • VLAs + World Action Models learn how objects move and interact.
  • Dual-system hierarchical design splits slow reasoning from fast control.
  • Simulation flips data bottleneck: compute-bound, not labor-bound.
• Robotics challenge: 95% per-step = only 60% on 10-step tasks.
  • RL post-training (RECAP) enables autonomous robustness improvement.
• Self-driving labs combine all five primitives most completely.
  • SDL experimental data gives world models what internet data cannot.
• AR, EMG wristbands, BCIs = new interfaces and training data pipelines.
• Three domains form a flywheel — robotics, SDL science, interfaces feed each other.
• New modality access historically produces qualitatively larger capability jumps.

Oliver Hsu, Partner — American Dynamism, a16z (Apr 15, 2026) · 2026-04-15 · Read on a16z.com