systems theory//emergence

2026-03-08

Emergence is macro-level behavior that arises from micro-level interactions without being explicitly designed or programmed: the whole does something the parts do not. It is the central phenomenon systems theory exists to explain, and engineers meet it whenever many agents follow local rules (complex system).

The cleanest engineering case is a flock. In 1987 Craig Reynolds animated flocks on a computer with three local rules per bird: separation (do not get too close), alignment (fly in the mean direction of your neighbours) and cohesion (move towards their centre). Nobody holds the plan of the flock, and the flock appears. A drone swarm built on such rules promises to scale, since each drone talks to a few neighbours however many there are, to tolerate losses and to need no infrastructure; the same alignment rule, written for numbers instead of headings, is the consensus protocol of multi-agent control.

Collective order can switch on abruptly, a phase transition. In 1995 Vicsek and colleagues kept only the alignment rule, added noise, and found that below a critical noise level, or above a critical density, the group passes suddenly from disordered milling to ordered motion. The order belongs to the group: no single agent changes its rule at the threshold (bifurcation for the same jump in a single system).

In first-order emergence a single surprising behavior appears. The emergent behaviors of RL-trained models (chain-of-thought reasoning appearing in DeepSeek-R1-Zero, tool use appearing in Cursor's Composer) are first-order: each is one capability the training signal did not specify.

In second-order emergence, individually emergent behaviors compose into workflows or strategies no one designed. Cursor's model learned to search, then edit, then self-test, then fix. Each step emerged independently via RL, but together they form a coherent strategy. In extended thinking, individual reasoning steps are first-order, but the model's ability to sequence them into multi-step proofs is second-order.

Weak emergence is in principle deducible from the parts, given enough compute; strong emergence is not. Most emergence in ML is weak: surprising to us but implicit in the optimization landscape.

Many emergent capabilities appear abruptly at certain model scales (scaling laws). Below the threshold the capability is absent, and above it it appears suddenly, a behavior like a phase transition.

Model collapse is emergence in reverse: a macro-level failure (distribution narrowing) that emerges from micro-level dynamics (training on synthetic data), with no single training step being the cause.