--- title: "Emergence Is a Borrowed Word" description: "Physics meant something exact by emergence. Machine learning borrowed the word and kept only the drama. The difference is worth knowing before you use it." date: 2026-08-16 category: Science readingTime: "5 min read" --- In 1972, Philip Anderson published a four-page paper in *Science* called "More Is Different". It argues one thing, and argues it carefully. Reductionism is true and it does not help you. You can accept that everything is made of particles obeying known laws, and still find that knowing those laws tells you almost nothing about what happens when you put enough of the particles together. At each new level of scale, Anderson says, entirely new properties appear, and understanding them requires research as fundamental as anything happening a level below. His phrase is that the ability to reduce everything to simple laws does not imply the ability to start from those laws and reconstruct the universe. He was not being poetic. He gave the mechanism: broken symmetry. A single ammonia molecule has no fixed orientation, because the laws governing it are symmetric and it flips between states. A large enough crystal made of the same molecules does have one, because at scale the flipping stops being possible on any timescale that matters. The symmetry of the underlying law survives in the law and dies in the object. Nothing has been added. The parts are unchanged. What changed is how many there are. I have been reading Anderson because I got to him through machine learning, which is the wrong direction and probably a common one. ## The word arrives in machine learning In 2022 a paper by Jason Wei and colleagues, "Emergent Abilities of Large Language Models", used the word for something specific: an ability absent in smaller models and present in larger ones, appearing not gradually but as a jump, at a scale nobody predicted in advance. The plots that went with it are the ones everyone has seen. Flat, flat, flat, then a wall. Arithmetic that the model simply could not do at one size and can do at the next. The claim landed hard because it carries a promise. If capability arrives in jumps, then nobody knows what the next jump contains, and the only way to find out is to build the bigger thing. That is a very convenient conclusion for anyone raising money to build bigger things, which is not an argument against it being true, but is a reason to check it. It got checked. In 2023, Rylan Schaeffer, Brando Miranda and Sanmi Koyejo published "Are Emergent Abilities of Large Language Models a Mirage?", and their answer was largely yes, for a reason that is almost embarrassing once you see it. The jumps are frequently in the metric, not the model. If you score a task by exact string match, a model that gets four digits right out of five scores zero, exactly like a model that gets none right. The underlying competence can improve perfectly smoothly while the score sits at zero and then leaps. Replace the discontinuous metric with a continuous one, token-level accuracy rather than all-or-nothing, and in a great many cases the wall becomes a slope. That is not a refutation of everything. It does not prove no sharp transitions exist anywhere in these systems. What it establishes is narrower and more useful: a sharp curve is not evidence of a sharp thing until you have ruled out that you drew it with a sharp ruler. ## What got lost in the borrowing Here is the part I find genuinely interesting, and the reason the reading list ran backwards from CS229 into physics. Anderson's emergence is a claim about the world. He is saying that a new level of organisation obeys laws that are not derivable from the level below, and that no improvement in your measurement of the parts will hand you those laws. It is a statement about what exists. The machine learning claim, as it is usually made, is a claim about a graph. It says that a number went up faster than expected somewhere. Those are not the same kind of assertion, and the second one inherits none of the weight of the first. Borrowing the word imports a seriousness the evidence has not earned yet. What makes this more annoying rather than less is that the connection between the two fields is real and historical, not a metaphor anyone reached for. Lev Landau's theory of second-order phase transitions, from 1937, gives the formal apparatus: an order parameter, a quantity that is exactly zero on one side of a critical point and non-zero on the other, with the transition being a genuine change of state rather than a change of degree. That is the actual shape of the question people mean to be asking about scale. Spin glasses, systems of disordered magnetic interactions studied with precisely this machinery, led John Hopfield in 1982 to the network that carries his name, which stores memories as energy minima. Hopfield shared the 2024 Nobel Prize in Physics with Geoffrey Hinton for exactly this line of work. Alexander Polyakov's territory, the behaviour of a theory as you change the scale you look at it on, is the same instinct pointed at field theory. So the field already owns the tools. It has an order parameter concept, a critical point concept, and a working definition of what would count as a real transition. Using "emergence" as a synonym for "the graph got steep" throws all of that away and keeps the vibe. ## Why this matters more than it seems to I have written before about deciding what is worth building, and this is the same problem wearing a lab coat. A borrowed word carries an argument inside it. Say "emergent" and you have quietly asserted that the behaviour is unpredictable in principle, that scale is the only lever, and that nobody could have known. Say "our metric is discontinuous" and you have asserted that you should fix the metric. Same data, two different next actions, and the word did the deciding before anyone noticed a decision was being made. The test Anderson and Landau hand you is a good one and it is not difficult. What is the order parameter. What is it before, what is it after, and does it change state or does it change size. If you cannot answer that, you have observed a steep curve, which is worth reporting and is not worth a claim about the nature of intelligence. I am seventeen and working through the CS229 notes, so nothing here settles anything. But I notice that the reading which has changed how I look at the plots was not written about machine learning at all, and was in one case written thirty-seven years before I was born. The habit worth taking from it is not a position on whether large models emerge. It is the discipline of asking, every time a word arrives from another field, what it meant where it came from and whether the thing it was describing is the thing in front of you. Usually it is not, and the gap is where the interesting question was hiding.