Where does emergence collide?

An ability that is absent in smaller models and present in larger ones, so that performance cannot be predicted by extrapolating the smaller models' curve. Wei and colleagues fixed the definition in 2022 and catalogued dozens of tasks (arithmetic, multi-step reasoning, word unscrambling) where accuracy sat near zero and then jumped at a scale threshold. Schaeffer, Miranda, and Koyejo replied in 2023 that the jumps are artefacts of discontinuous metrics: score the same models with a continuous measure and the curves are smooth. What is at stake is not whether large models do new things but whether the transition is real or an accident of exact-match scoring.[1][2]

Unit

A model family, indexed by parameters or compute, on a benchmark

Recovery means

Not applicable, the question is prediction, not return

Measured by

Benchmark accuracy against log scale; smoothness of the curve under alternative metrics

Behaviour of a many-body system that obeys its own regularities, which the laws of the constituents do not make obvious even though they entail it. Anderson's 1972 argument was against reductionism as a research strategy: knowing the Schrödinger equation does not tell you about superconductivity or broken symmetry, and each level of complexity needs its own science. Crutchfield's 1994 calculi of emergence made it operational (a pattern is emergent when an observer needs a new, more powerful model class to capture it) and Bedau's 1997 weak emergence says an emergent property is one derivable from the micro-level only by simulating it step by step. Nothing spooky is asserted; the micro-level fixes everything and still leaves the macro-level worth studying on its own terms.[3][4][5]

Unit

A macroscopic pattern in a system of many interacting parts

Recovery means

Not applicable, the pattern persists as long as the interactions do

Measured by

Order parameters; statistical complexity; whether a coarse-grained model predicts the macro-level

A property of a whole that is novel with respect to, and not deducible from, the properties of its parts. The British emergentists (Mill on heteropathic laws in 1843, Broad on emergent qualities in 1925) meant this strongly: chemistry and life were held to introduce new causal powers not fixed by physics. Chalmers's 2006 distinction is now standard. Strongly emergent properties are not deducible even in principle from the low-level domain; weakly emergent ones are unexpected but deducible, which is what physicists and biologists usually mean. Kim's 1999 challenge asks how a strongly emergent property could have any causal work to do if the physical level is already causally complete. Here emergence is a metaphysical thesis, and consciousness is its only serious candidate.[6][7][8]

Unit

A property of a whole, relative to its parts

Recovery means

Not applicable, the thesis is timeless

Measured by

Not measured; argued from deducibility, supervenience, and causal closure

Pattern at the level of an organism or a colony produced by local interactions among components that follow simple rules with no reference to the pattern. Kauffman argued in 1993 that much biological order is spontaneous, random Boolean networks settle into a small number of attractors, and selection acts on structure that self-organisation supplied for free. Camazine and colleagues fixed the working definition for behaviour in 2001: a process in which pattern at the global level emerges solely from numerous interactions among lower-level components, using only local information. The difference from physics is that the components are organisms or cells and the rules were tuned by evolution; the difference from philosophy is that nothing is claimed to be underivable.[9][10]

Unit

A population, colony, or developing tissue

Recovery means

Regeneration, the pattern re-forming after perturbation from the same local rules

Measured by

Agent-based models reproducing the pattern from local rules; perturbation experiments on colonies and tissues

Structure at the level of a society or market that arises from individual choices without being anyone's plan. Schelling's 1971 model is the canonical demonstration: agents who each tolerate a mixed neighbourhood but move if too outnumbered produce near-total segregation, a macro-outcome no agent wanted. Hayek called the general phenomenon spontaneous order, and Sawyer's 2005 Social Emergence set out the sociological version, arguing that social properties are real, irreducible in practice to individual psychology, and yet produced by nothing but interacting individuals. The unit is a collective, the mechanism is aggregation of choices, and the payoff is the warning that macro-patterns need not reflect micro-motives.[11][12]

Unit

A population of choosing agents and the aggregate pattern

Recovery means

Not applicable, the pattern is an equilibrium, and change requires new rules or new preferences

Measured by

Agent-based simulation; segregation indices; macro-regularities not present in individual-level data