Emergence: When More Becomes Different

“More is different.” The physicist Philip Anderson’s 1972 essay of that title is one of the great manifestos of twentieth-century science: a rebuke to the idea that everything interesting follows from the fundamental laws, and an insistence that at each new level of complexity — from particles to atoms to molecules to minds — entirely new properties appear, requiring entirely new descriptions.[1] Water is wet, though no single H₂O molecule is. Magnetism emerges from spins that individually know nothing of it. The whole, at sufficient scale, becomes something the parts never were.

Fifty years later, the slogan found an unlikely second life in machine learning. As language models grew — more parameters, more data, more computation — researchers began reporting something strange: abilities that were absent in smaller models would appear, seemingly all at once, past some threshold of scale. A model that could not do multi-step arithmetic at one size could do it at the next. Capabilities arrived not as a slope but as a step. The word on everyone’s lips was emergence.

Then came the backlash: a widely discussed paper arguing that these emergent abilities were a mirage, an artifact of how we measure. The debate that followed is one of the most instructive in recent AI science — not because it settled the question, but because it revealed how hard the question is.

The catalogue of surprises

The original case for emergence was empirical and hard to dismiss. In 2022, researchers catalogued more than a hundred abilities — from understanding novel words in context to solving multi-step reasoning problems to writing code — that appeared abruptly as models crossed certain scales, while being near-absent below them.[2] The pattern repeated across model families and training setups. It was not that bigger models were merely better at everything; it was that they could do new kinds of things, things their smaller siblings could not do at all.

Anyone who has watched the field from the inside knows the feeling these results describe. There is a particular vertigo in prompting a new, larger model and discovering it can do something no one trained it to do — something that appears nowhere in its specification, only in its scale. Chain-of-thought reasoning, instruction-following, in-context learning: none of these were programmed in. They showed up, uninvited, when the systems got big enough. Whatever emergence turns out to be, the experience of being surprised by one’s own artifact is real, and it demands an explanation.

Grokking: emergence in miniature

For a cleaner view of the phenomenon, consider grokking — emergence observed under laboratory conditions. In 2022, researchers training small networks on simple algorithmic tasks noticed something bizarre: the models would first memorize the training data perfectly, achieving near-zero training loss while failing completely on new examples — and then, long after any reasonable person would have stopped training, with no change to the data or the objective, they would suddenly generalize. Test accuracy would snap from chance to perfect, as if the network had abruptly understood what it had previously only memorized.[3]

Grokking matters because it isolates the phase-transition character of emergence from the messiness of scale. Nothing external changed; the system reorganized itself. Subsequent work suggested a mechanism: the network first finds a fast, lazy solution (memorization) and only later, under the quiet pressure of regularization, discovers the slower, elegant one (the general algorithm) — a kind of Occam’s razor operating on geological time. It is a small parable of how qualitative novelty can arise from purely quantitative persistence: keep optimizing, and the solution changes its kind.

Whether grokking explains anything about large-scale emergence is debated. But it demonstrates, at minimum, that neural networks undergo genuine phase transitions — sudden reorganizations of their internal strategies — and that these transitions are not artifacts of coarse measurement. They are in the weights.

The mirage debate

Now the backlash. In 2023, a team of researchers argued that emergent abilities in large language models might be, in large part, a mirage — a trick of the metrics. Their argument was elegant: many evaluations score models with nonlinear, all-or-nothing measures (exact match, multiple choice accuracy), and when you plot a smooth underlying improvement through a discontinuous scoring function, you get the appearance of a sudden jump. Choose a smoother metric — the probability the model assigns to the correct answer, say — and the cliff becomes a slope. The emergence, they suggested, was in the ruler, not the thing measured.[4]

It was a genuinely important critique, and it improved the field’s methodology overnight. But the rebuttals were substantive too. The original authors replied that some discontinuities survive even smooth metrics — that while some reported emergence is indeed metric artifice, the phenomenon does not reduce to it. And crucially, phenomena like grokking cannot be explained away by choice of metric at all: the network’s behavior on held-out data changes discontinuously under any measure you like.

Our reading of the debate, several years on: both sides won something. The skeptics were right that the literature had been sloppy — that many “emergent” curves were partly an illusion, and that smoother measurement should be the default. The defenders were right that the illusion thesis doesn’t cover everything — that qualitative novelty at scale is a real feature of these systems, even if rarer and subtler than the early catalogues suggested. The truth, as usual, refused to fit in a headline.

Living with unpredictability

Why does any of this matter beyond academic dispute? Because emergence, if real, is a claim about predictability — and predictability is the foundation of both engineering and safety.

If capabilities arrive discontinuously with scale, then we cannot reliably forecast what the next generation of models will do by extrapolating from the current one. Evaluation becomes a race against surprise: we must test for abilities we have no reason to expect, because expectation itself is unreliable. This is the serious core of the emergence debate, and it survives the mirage critique intact. Even if half the reported jumps were metric artifacts, the remaining cases — and the general experience of unexpected competence — are enough to establish that scaling these systems involves genuine uncertainty about what will come out.

There is a temptation, in some quarters, to treat emergence as either a mystical vindication (“the machines are waking up!”) or a dismissed illusion (“nothing to see here”). Both miss the point. Emergence, properly understood, is neither magic nor mirage but a familiar feature of complex systems: the way quantity becomes quality, the way new regularities appear at new scales, the way Anderson said it would. The responsible response is not awe or denial but better science — smoother metrics, sharper definitions, mechanistic explanations of why particular capabilities appear when they do.

That research program is underway, and it is one of the most exciting in the field: tracing specific capabilities to specific circuits, predicting phase transitions before they happen, turning surprise into understanding. If it succeeds, “emergence” will graduate from a word we use to mark our confusion into a phenomenon we can model. Until then, we keep two ideas in mind at once: scaling curves are mostly smooth, and qualitative novelty still happens. More is different — sometimes. Knowing when is the work.

Notes

1. Anderson, “More Is Different,” Science 177, 1972 — the founding text of emergence as a scientific idea, and still the clearest statement of why reduction is not enough. ↩
2. Wei et al., “Emergent Abilities of Large Language Models,” 2022 — the catalogue: over a hundred abilities appearing abruptly with scale. ↩
3. Power et al., “Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets,” 2022 — sudden generalization, long after memorization, under controlled conditions. ↩
4. Schaeffer, Miranda & Koyejo, “Are Emergent Abilities of Large Language Models a Mirage?” NeurIPS 2023 — the critique that changed how the field measures. Read it alongside the authors’ response. ↩