World-Model Vanguardism

From the AI Research Compass wiki

World-Model Vanguardism: Before anything else, intelligence must build an internal predictive model of the world; language and agents come after.

Beliefs

World-Model Vanguardism believes the field has the order of operations wrong. Language models learned to talk before they learned how the world works, and agents are being built on top of that inverted foundation. The vanguard’s program is to first construct a model of the world, something that predicts how scenes evolve and how actions change them, and only then attach language, planning and goals. With a good world model, an agent can imagine consequences before acting, which is what makes animals and humans so sample-efficient.

It sits in the upper band, deep on the World side. The Monolith lean reflects its preference for one learned predictive model as the core of intelligence rather than an assembly of tools. Its paradigm level is Beyond Revolutionary, the far paradigm-shift end: the vanguard openly argues that the current paradigm is a detour, and its proposed alternative is still being built.

Arguments

  • Agents that learn a world model and train their policy inside its imagined rollouts have mastered diverse tasks with a single configuration.
  • Generative interactive environments learned from unlabeled video show world models can be acquired without explicit action labels.
  • Planning through a learned model gives sample efficiency and counterfactual reasoning that pure model-free policies lack.

Critiques

  • “World model” is used loosely enough to cover very different objects, from pixel video generators to small latent dynamics models.
  • No world-model-first system has yet matched language models on the breadth of tasks they handle.
  • Language-first rivals argue that text already encodes a large, usable world model, so building one from scratch is unnecessary.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ +1.00, y ≈ +0.63, paradigm ≈ +1.00.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data