Skip to content

Explore agents that act beyond text

Engineering field guide · World models · Both days

Do you need a model that selects actions or generates an environment?

Compare General Intuition’s action-labeled game data with Runway’s video simulation approach. Keep shipped demonstrations separate from future promises.

All topic guides · Practical takeaways

Both talks explore learning from visual experience. Their demonstrations suggest possibilities beyond text, while leaving important deployment questions open.

Session Approach
General Intuition · Day 1 Uses action-labeled game data to train models that continuously observe and select actions.
Runway · Day 2 Uses next-frame prediction to generate visual environments and interfaces. Its Continuum demonstration explores a video-based operating system.

General Intuition · Day 1 · Runway · Day 2

Action selection and environment generation are different tasks. A convincing video does not establish reliable physical behavior. Runway’s release window is a speaker’s plan, and some demonstrations were not generated in real time.

  1. Write down whether your use case needs action selection, visual prediction, or both.
  2. Choose a small recorded interaction and define an observable success condition.
  3. Test a repeatable baseline or an available model on held-out interactions before designing around a demonstrated research feature.

Track task success, response latency, repeatability, and failures when the environment changes. Report unavailable capabilities as requirements still to test.

The comparison and experiment are our editorial synthesis. They are not a reported benchmark or the speakers’ prescribed procedure.