6.10 Imagination and counterfactual worlds

The Blue Sky paper Imagination Machines argued for imagination as an overarching challenge for AI and connected generative models to layered causal capacities for observation, intervention, and counterfactual reasoning [ Mahadevan , 2018 ] . The arrival of powerful generative and world models makes that challenge experimentally approachable: a frontier model can now propose mechanisms, scenes, analogies, experimental designs, and formal objects at a scale unavailable to earlier systems.

Imagination nevertheless exceeds unconstrained sampling. A useful imagined world must preserve some invariants while varying others, expose the intervention that produced the variation, and distinguish a possible world from an admitted claim about the actual one. Simulators are especially important because they turn an imagined intervention into an executable test. They also impose a hard limitation: success establishes validity relative to the simulator, not automatically to nature.

This motivates the division of labor adopted here. Frontier models serve as proposal engines; sketches register the structure being preserved or changed; tangent and Weil probes explore local alternatives; and proof, simulation, or empirical evidence controls admission. Imagination supplies candidates, while the theory package prevents a vivid candidate from silently becoming a fact.