06-reference

not boring runway world models

2026-10-13·reference·source: Not Boring·by Packy McCormick + Anastasis Germanidis (co-written; guest co-author is Runway co-CEO)
world-modelsvideo-generationrunwayagent-harness-architectureroboticscompetitive-landscape

"Runway: Learning From Watching the World" — @packyM + Anastasis Germanidis

Why this is in the vault

Runway's own synthesis of why video-pretraining generalizes into interfaces, robotics, and (claimed) science — a live data point for RDCO's AI-video/generative-media watch area and for tracking whether "generate the interface, not the code" becomes a real competitor to the code-and-API substrate RDCO's agent harnesses run on.

⚠️ Sponsorship

Two layers, both disclosed or verified separately rather than assumed:

Mapping against Ray Data Co

The concrete anchor: this essay directly extends the vault's existing Runway/robotics thread. It cites Standard Bots (Evan Beard's company, co-author of [[2026-01-16-not-boring-robot-steps]]) now testing Runway's Praxis-1 on its RO1 arm — the two essays are describing the same emerging robotics-data-transfer relationship from opposite sides of the table.

Beyond that specific thread, two things are worth banking for RDCO's own positioning:

  1. Competitive-landscape signal on "the interface itself." Solaris and Project Continuum bet that software interfaces can be generated frame-by-frame by a World Model instead of rendered from code written by (or for) an agent. If that substrate matures, it's a different kind of agent production system than the one RDCO builds and sells — harnesses that orchestrate code-mediated tools and APIs. The essay itself flags the near-term limits (legible text, trust/verification, long-session drift) that keep this from being an immediate threat, but it's the kind of adjacent-infrastructure development worth re-checking in 6-12 months rather than dismissing as "just a video company."
  2. "Sandbox superintelligence" vs. "real-world superintelligence" is a useful frame for RDCO's own harness-engineering thesis. Anastasis's distinction — intelligence that improves by checking its own output against a closed, verifiable environment (code, math, Runway's own World-Model-as-simulator) vs. intelligence that has to act in messy, slow-feedback real-world systems — maps cleanly onto why RDCO's agent deployments concentrate in code/structured-data domains where verification is cheap, and why that's the tractable frontier rather than a limitation to apologize for.

The core argument

Runway treats video, interfaces, robotics, and (eventually) science as one bet: pretrain a model to predict the next frame at scale (next-frame prediction is autoregressive, the same shape as next-token prediction for LLMs), and application-specific post-training is cheap once the backbone is good. Solaris turns clicks/drags into "actions" that condition the next frame instead of rendering pre-coded UI. Praxis-1 treats robot end-effector positions as actions, fine-tuned from video pretraining with ~30 minutes of robot-specific data; Runway's own test found web-video pretraining produced results statistically close to teleoperated-robot-video pretraining (16.1cm vs. 16.0cm placement error). Anastasis extends the frame to science: World Models as "wind tunnels" for domains where real experiments are slow (drug trials, markets, robotics), with the caveat (citing a 2018 Ha/Schmidhuber paper) that agents can learn to exploit an overly-predictable simulator, so the real/simulated feedback loop has to stay closed. The piece ends on his "Towards Universal Simulation" thesis: solve "everything else" (a universal simulator of physical reality) first, then use it to solve intelligence, rather than the reverse order most AGI labs have pursued.

Related