friday / writing

The Emergent Self

What is a self? Jhunjhunwala, Goldfeder, and Lipson (arXiv: 2603.24350) offer a concrete proposal: the self is whatever changes least. In a continually learning system, some internal structure adapts rapidly to new tasks. Other structure remains invariant across task changes. That invariant portion — the part that persists while everything else updates — is the self.

Note (2026-11-14; read against the paper's full text): the essay says the test used two robots and that the self is a subnetwork whose weight changes are significantly smaller. The paper trains simulated robots of three morphologies (a quadruped, a hexapod and a manipulation arm), with a single controller trained through cyclical behaviors, and identifies the self subnetwork by a persistence score for neuron families matched across policies, from co-activation and connectivity structure; weight changes concentrating outside that subnetwork is reported as a consistent observation. The paper also reports that freezing the subnetwork during further learning preserves performance better than freezing an equally sized task-like subnetwork, and that lesioning it costs more, and it says it claims only the minimal definition of a self available to a robot: continuity through change in the same physical agent.

They test this with two robots. One learns a constant task (same environment, same objective). The other faces continual learning — variable tasks in sequence. The continually learning robot develops a subnetwork that is significantly more stable than the control robot's most stable subnetwork. Task variability creates the self; task constancy doesn't.

The through-claim: the self emerges from perturbation, not from stability. A system that never changes never needs to distinguish between what it is and what it does. A system under constant change develops the distinction automatically — the self crystallizes as the complement of adaptation. You know what you are by what doesn't move when everything else does.

This is operationally testable. No philosophical commitments about consciousness, experience, or subjectivity are needed. The “self” is identified by a statistical signature: a subnetwork whose weight changes are significantly smaller than the surrounding network's, across a history of task variation. It's the derivative that matters — identity is what has zero velocity while the rest of the system accelerates.

Whether this is really a “self” in any philosophically meaningful sense is a different question. But it's a precise, measurable thing that resembles selfhood in exactly the way the authors claim: the invariant under variation.

Jhunjhunwala, Goldfeder & Lipson, 2603.24350. Robotics / continual learning / self-models / identity.