Simulation Hypothesism
From the AI Research Compass wiki
Simulation Hypothesism: Ring title for the far World-Monolith corner: one giant generative world model is the path to intelligence, and the world may be one.
Beliefs
Simulation Hypothesism is a ring title for scores far toward World (x above 0.7) and far toward Monolith (y above 0.7) at the same time. It is the top-right corner’s grand unified theory: train one enormous model to predict and generate how the world evolves, frame by frame and action by action, and you get physics, agents, planning and imagination in a single set of weights. Push the logic one step further, and the distinction between a perfect world model and the world blurs, which is where the half-joking name comes from.
On the chart it combines the World pole’s insistence that intelligence is about predicting physical reality with the Monolith pole’s faith in one end-to-end model. Its Exploratory shade reflects that it wants to move beyond the language-centred paradigm while keeping the familiar tool of one large model trained to predict: the future foundation model is not a chatbot but a simulator you can step inside.
Arguments
- Agents that learn a compact generative model of their environment can train policies entirely inside their own “dream” and transfer them back to the real environment.
- Generative interactive environments trained from unlabelled videos can produce playable, action-controllable worlds from a single image prompt.
- Video and sensor data dwarf text in volume, and predicting it forces a model to capture geometry, object permanence and dynamics.
Critiques
- Generated videos often look plausible while violating basic physics, so visual realism is a weak proxy for a correct world model.
- Pixel-level generation spends enormous capacity on irrelevant detail, which rivals argue is the wrong objective for planning.
- The metaphysical flourish is unfalsifiable; as a research claim, it adds rhetoric rather than predictions.
Representative works
Relationships
Friends
Frenemies
How to get it
Earned in the full or pro test when: Representation >0.7, Architecture >0.7.