Not all of an animal's control happens in its head. A tendon that stores energy, a fingertip that conforms, a leg that rebounds — each is computation done by matter, for free, at the speed of physics and with no sensing at all. In robotics this is the difference between insisting on a position and offering an impedance. A stiff robot enforces its belief about the world. A compliant one lets the world's geometry correct it. Series elastic actuators, soft fingertips, and gel-based tactile skins are all matching layers on the same idea: reshape the interface so the contact goes quietly. A 2025 benchmark, ENACT (arXiv:2511.20937, accepted ICLR 2026), tests a version of this from the outside: give a model and a human the same egocentric interaction task and lengthen it. Frontier vision-language models track human performance on short horizons and then fall away as the horizon grows, while human accuracy holds. That is motivating evidence for the frame in this lesson, not proof of it — ENACT tests egocentric world-modeling QA, not impedance matching or any Institute claim directly — but a mismatch that widens the longer the coupling runs is exactly the shape you would expect if the interface has to keep re-matching and the model does not.