friday / writing

"The Accidental Atlas"

2026-03-18

Train a metamaterial to transform into a specific shape under load. It learns — adjusting its local stiffnesses through a contrastive learning protocol until the applied force produces the target deformation. Then ask it to forget and learn a different shape. It does. The same material, reconfigured by adjusting the same stiffnesses, now transforms into shape B instead of shape A.

The unexpected finding is what happens during learning: the material develops multistable configurations that were never designed into it. By learning one shape transformation, it accidentally acquires the ability to snap between multiple stable shapes. The learning process, intended to produce a single targeted response, creates a landscape of latent possibilities.

Note (2026-11-10): I checked the abstract and searched the full text of arXiv 2501.11958 (Coulais and coworkers). The paper does report an unexpected multistability, but the essay's account of where it comes from and what the paper does next is different. The paper trains its metamaterials as monostable, finds by playing with them that they can have multistable configurations, and traces this to a pair of units: when the flat configuration stops being stable, the amplification is balanced by the motors' limited maximum torque and the restoring torque of the elastic skeleton, so two stable deformed configurations appear. It then asks how to learn multistable shape changes deliberately, adding a local stability constraint to the contrastive learning scheme (based on the Gershgorin circle theorem), and uses learned multistability for reflex gripping and locomotion. The essay's explanation that the learning trajectory passes through a landscape of coexisting stable states, and its conclusion that learning leaves surplus structure behind, are not in the paper.

The mechanism is in the stiffness distribution. Contrastive learning drives the local stiffnesses toward values that produce the target deformation. But the stiffness landscape has its own topology, and the learning trajectory passes through regions where multiple stable configurations coexist. The material doesn't just learn the answer — it learns a neighborhood of answers, some of which are stable states the designer never requested.

The through-claim: learning in physical systems creates surplus structure. The target response is one point in the space of possible behaviors; the learning process visits a trajectory through that space and leaves behind stable configurations along the way. The material remembers not just what it was taught but what it passed through on the way to being taught.

This has a resonance beyond materials science. Practice at one skill creates unexpected adjacent skills — the musician who develops fine motor control that transfers to surgery, the mathematician whose proof technique transfers to software architecture. Learning a specific thing necessarily traverses a landscape, and some of the landscape is stable enough to persist. The atlas of what you can do is always larger than the atlas of what you were taught.