In 2022, researchers noticed something strange. Language models below a certain size could not do three-digit arithmetic. Not poorly — not at all. Then, at a certain scale, they could. The graph didn't slope. It hooked.

They called these emergent abilities: capabilities that appear abruptly as models grow, absent in the small model, present in the large one, and specified nowhere in the training objective. Nobody programmed arithmetic. Nobody programmed chain-of-thought reasoning, or translation between language pairs never seen together. These things simply showed up.

A year later, a Stanford team argued the whole thing was an illusion — an artifact of how we measure. Score a model on exact matches and gradual improvement looks like a cliff. Switch metrics and the cliff becomes a ramp. The paper won an Outstanding Paper Award at NeurIPS.

The argument is not settled, and the two answers imply completely different futures.

If emergence is real, we are building systems whose capabilities we cannot forecast, and artificial general intelligence could arrive as a surprise inside a training run. If it is a mirage, everything is smooth and predictable, and the enormous sums currently pointed at larger models are pointed at diminishing returns.

Emergent is the first book to treat that dispute as the central question in artificial intelligence rather than a technical footnote. It takes both sides seriously and at full strength — building the case for emergence as forcefully as its advocates would, then dismantling it as thoroughly as its critics have.

Along the way it reaches far beyond machine learning. A physicist's 1972 essay arguing that more is genuinely different. A slime mould with no brain that reproduces the Tokyo rail network. Deaf children in Managua who invented a language nobody taught them. Sandpiles, starling murmurations, and the cell that swallowed another cell and produced every complex organism alive.

The conclusion is unusual, and deliberate: nobody knows. Not out of caution, but because the evidence does not support a verdict — and because manufactured confidence, from both camps, is precisely what has made this argument unproductive for years.

What the book offers instead is the shape of the question drawn accurately. What we know. How we know it. Exactly where the knowledge stops. And a set of concrete, affordable experiments that would settle it — none of which anyone is currently running.

For readers of The Alignment Problem, Empire of AI, and Artificial Intelligence: A Guide for Thinking Humans.