The strongest version
General intelligence is grounded in acting in a world. Without sensorimotor experience — a closed loop of prediction, action, and consequence — systems learn the shadow of understanding: statistical competence over descriptions of the world, not competence in the world. The strong claim is that some concepts (causation, affordance, persistence) are not fully learnable from passive text and vision alone.
Who holds it
Rich Sutton’s post-2024 turn toward experience and embodiment is the notable recent movement — the author of The Bitter Lesson arguing that the next bitter lesson is grounding. The robotics-learning and world-model communities supply the empirical program.
The load-bearing assumption
That the grounding gap is real and load-bearing for general intelligence — not a domain that disembodied systems will simply pick up once other pieces are in place.
Falsifiers
A disembodied system that demonstrates robust causal and physical reasoning — reliably, off-distribution, without embodied training — would be strong evidence against. Health is the lowest of the five and holding: the theory is coherent, but disembodied systems keep absorbing tasks it predicted they could not.