The Word AI Researchers Are Quietly Taking Back: "Emergence"
June 21, 2026 · M. Mitchell Waldrop, Complexity~6 min read
Draw the chart one way and it looks like a miracle: a language model crawls along, useless at some task, useless, useless — and then at a certain size the line leaps off the floor. People started calling that leap "emergence," and the word did a lot of work. It hinted that somewhere in the scaling, the machine had crossed a threshold and woken up. In June 2025 three scientists from the Santa Fe Institute looked hard at that leap and made a quieter, sharper claim: redraw the chart with a kinder ruler and the cliff turns into a hill. The leap was partly in the measuring.
A word with a serious pedigree
"Emergence" did not start as an AI buzzword. It belongs to complexity science, the field whose origin story M. Mitchell Waldrop told in his 1992 book Complexity — the one about a band of restless physicists, economists and biologists who founded the Santa Fe Institute precisely because the big questions kept falling between the established disciplines. Their rallying cry came from the physicist Philip Anderson, in a 1972 essay with a title that has outlived almost everything else from that year: "More Is Different." Anderson's point was modest and devastating. You can know every law governing a single particle and still not be able to predict what a trillion of them will do together. Water is wet; no single H₂O molecule is. The whole acquires properties the parts simply don't have, and those properties need their own language to describe. That is emergence in its honest sense — and the new paper's first complaint is that this honest sense has gone missing from most AI talk.
More is different — but that's only half the sentence
The authors are David Krakauer, who runs the Santa Fe Institute, his brother John Krakauer, a neuroscientist, and Melanie Mitchell, one of the clearest voices in complexity science. Their move is to split a word everyone had been using as one thing. Scaling a model up, they grant, really can produce emergent capability: feed a network enough text and abilities appear that weren't visibly there in the smaller version — syntax, code, recalling a fact. That is "more is different," and it is real. But intelligence, they argue, is a different animal, and it runs the other way. Intelligence is "less is more" — the knack of solving a problem quickly and cheaply, with less data, less compute, less energy, than the brute-force route would demand. A pocket calculator has an emergent capability your grandmother lacks. It is not more intelligent than she is. The whole confusion of the past few years, in their telling, is a single skipped distinction: a capability you can demonstrate is not the same as a mind that deploys it well.
A June 2025 Santa Fe Institute paper (Krakauer, Krakauer & Mitchell, in a 2026 Royal Society theme issue) separates two things often blurred: scaling a model can yield new capabilities ("more is different"), but intelligence is "less is more" — doing more with less. And the famous capability "jump" may partly be a measuring artifact: under a smoother metric the cliff flattens to a slope. Framework: M. Mitchell Waldrop, Complexity. This is an open scientific debate, not a settled result.
Was there ever a cliff?
Then comes the part that should give everyone pause, on both sides. Some of those dramatic "leaps" may be an artifact of how we grade the test. The paper points to earlier work showing that when a capability is scored all-or-nothing — you only count it if the model gets the whole answer exactly right — progress looks like a sudden cliff. Switch to a metric that gives partial credit, and the same models improve along a smooth, boring slope. Nothing woke up; the ruler had a sharp edge. The Santa Fe authors are careful here, because the question isn't closed: other studies push back, and the matter is genuinely unsettled. But they add a deeper objection. A true phase transition — water to ice, the textbook image of emergence — turns on a single dial like temperature. "Scale," the x-axis of every AI emergence chart, isn't one dial; it's a tangle of data, parameters and compute bundled into one number. Calling its bend a phase transition, they suggest, borrows the drama of physics without earning it.
Why a 1992 book is the right lens
It is a small irony worth savoring that the sharpest critics of loose "emergence" talk are sitting in the very institute built to take emergence seriously. The Santa Fe Institute exists because its founders believed the universe is full of real emergent order — ant colonies, immune systems, economies, minds — that reductionism alone can't reach. Waldrop's Complexity is the chronicle of people learning to tell the genuine article from the mirage, to ask of any surprising pattern: is this a new level with its own laws, or am I fooling myself? That is exactly the question the Krakauers and Mitchell are now asking of large language models. The discipline isn't anti-emergence. It's the opposite — it cares enough about the real thing to police the word.
What this means for you
You don't need a physics degree to use the lesson, and you'll need it the next time a headline tells you a model has "developed" some startling new power. Ask two questions the Santa Fe paper hands you for free. First: is this a capability, or is it intelligence? A system can do a striking thing and still be a calculator, not a mind. Second: who drew the chart, and with what ruler? A jump that vanishes under a gentler metric was never quite a jump. None of this means the machines aren't capable, or that nothing surprising is happening — plenty is. It means a single word was carrying more weight than it could bear, and three scientists from the institute that loves that word most are the ones asking us to set it down and look again. The honest answer to "has it woken up?" is still the least satisfying one. We don't yet know — and the people closest to the question are the ones least willing to pretend otherwise.
Scaling a model can hand it a new capability. Whether that adds up to a mind is a separate question — and the leap you saw on the chart may have been drawn by the ruler, not the machine.
Emergence is real. So is our hunger to see it where the data only shows a slope.
Source: framework from M. Mitchell Waldrop, Complexity: The Emerging Science at the Edge of Order and Chaos (1992) — the Santa Fe Institute's project of distinguishing genuine emergence from wishful pattern-finding, and Philip Anderson's "More Is Different." Real-world basis: David C. Krakauer, John W. Krakauer & Melanie Mitchell, "Large Language Models and Emergence: A Complex Systems Perspective" (arXiv:2506.11135, submitted June 2025; in a 2026 Philosophical Transactions of the Royal Society A theme issue), which separates "emergent capability" from "emergent intelligence" and revisits the debate — including Schaeffer et al.'s argument that some emergent "jumps" dissolve under continuous metrics — over whether the leaps are real. An open scientific debate, presented as science communication, not settled fact.
接下来这一段,该让争论的双方都停一下。那些戏剧性的「跃迁」,有些可能是我们打分方式造出来的假象。论文指向更早的研究:当一项能力被「全有或全无」地评分——只有模型把整个答案一字不差答对才算数——进步看起来就像一道陡峭的悬崖。换成一个给部分分的尺子,同样这些模型,却是沿着一条平滑、乏味的坡缓缓上升。什么都没醒来,只是那把尺子有一道锋利的刃。圣塔菲的作者在这里很谨慎,因为问题并没有盖棺:另有研究反驳,此事确实悬而未决。但他们补了一条更深的异议。真正的相变——水结成冰,那张教科书里涌现的标准插图——是靠温度这样一个单独的旋钮翻过去的。而「规模」,每张 AI 涌现图横轴上的那个量,根本不是一个旋钮;它是数据、参数、算力搅成一团、塞进同一个数字里。把它的那道弯叫相变,他们暗示,是借了物理学的戏剧性,却没有挣得它。
「創発」はもともと AI 界の流行語ではない。それは複雑系科学のものだ——ワルドロップが1992年の『複雑系』で語ったのは、まさにこの学問の生い立ちである。落ち着きのない物理学者・経済学者・生物学者の一団が、最大の問いがいつも既存の学科の隙間に落ちてしまうからこそ、サンタフェ研究所を創った。その合言葉は、物理学者フィリップ・アンダーソンの1972年の論文に由来する。その年のほとんどすべてより長生きした表題——「多は異なる(More Is Different)」。アンダーソンの論点は控えめで、しかし痛烈だ。単一の粒子を支配するあらゆる法則を知っていても、一兆個が一緒に何をするかは予測できない。水は濡れているが、一個の H₂O 分子は濡れていない。全体は、部品がそもそも持たない性質を獲得し、その性質を語るには専用の言語がいる。これが「創発」の誠実な意味だ——そしてこの新しい論文の第一の不満は、その誠実な意味が、AI をめぐる議論のほとんどから消えてしまったことにある。
次の段は、論争の双方を一度立ち止まらせるべきところだ。あの劇的な「跳躍」のいくつかは、採点の仕方が生んだ見かけかもしれない。論文は先行研究を指し示す——ある能力を「全か無か」で採点すると(モデルが答え全体を一字一句正しく出したときだけ数える)、進歩は急な崖のように見える。部分点を与える尺度に切り替えれば、同じモデルが、なめらかで退屈な坂をゆっくり上っていく。何も目覚めてはいない。物差しに鋭い刃があっただけだ。サンタフェの著者たちはここで慎重だ。問いは閉じていないからである。反論する研究もあり、この件は本当に未決着だ。だが彼らはより深い異議を添える。本物の相転移——水が氷になる、教科書にある創発の定番図——は、温度のような一つのつまみで越える。ところが「規模」、あらゆる AI 創発図の横軸にあるあの量は、一つのつまみではない。データ・パラメータ・計算が一つの数字に束ねられたもつれだ。その曲がりを相転移と呼ぶのは、物理学の劇的さを借りておきながら、それを稼いではいない——と彼らはほのめかす。
出典:枠組はワルドロップ『複雑系——科学の新しい潮流』(1992)より——サンタフェ研究所の「本物の創発と、希望的な模様探しを見分ける」営みと、フィリップ・アンダーソンの「多は異なる」。現実の根拠:David C. Krakauer・John W. Krakauer・Melanie Mitchell「Large Language Models and Emergence: A Complex Systems Perspective」(arXiv:2506.11135、2025年6月投稿。2026年の『Philosophical Transactions of the Royal Society A』特集号所収)。同論文は「創発した能力」と「創発した知能」を分け、跳躍が本物かをめぐる論争——Schaeffer らの「一部の創発的『跳躍』は連続的な尺度では解消する」という主張を含む——を改めて検討する。本稿は未決着の科学論争を提示する科学コミュニケーションであり、確定した事実ではない。