The Curve Was Smooth. The Surprise Was Real.

The “emergent abilities” fight looks like an argument about the models. It’s really an argument about what we can see coming, and the honest answer changed this year, in a way most takes haven’t caught up to.

Somewhere around a certain model size, large language models started doing three-digit arithmetic. Below that size, near zero. Above it, it mostly works. Plot the accuracy and you get a flat line that snaps upward: a capability that wasn’t there and then was, with no warning in the smaller models that it was coming. Researchers named this in 2022, emergent abilities, and the name carried a thrill and a threat in one breath. The thrill is that intelligence might be hiding inside scale, waiting. The threat is that if new abilities arrive without warning, you can’t plan around them, and you can’t govern what you can’t predict.

Let me set one thing aside first, because it eats every conversation it touches. I’m not going to litigate whether any of this is “real” intelligence. That question is mostly unfalsifiable, and for anyone building, beside the point. The useful question isn’t whether the model understands. It’s whether a capability you can use shows up, when, and whether you could have seen it coming. That last part is where the live fight is, and it’s moved further in eighteen months than the discourse has noticed.

Start with the deflation, because it’s the strongest version of “calm down.”

The mirage

In 2023, a Stanford team published a paper with a needling title: Are Emergent Abilities of Large Language Models a Mirage? The argument is hard to unsee. The cliff, they showed, mostly appears when you score a task all-or-nothing. Multi-digit addition graded exact-match, the whole answer right or zero, produces a flat line that suddenly leaps. But the model isn’t leaping. Underneath, it’s getting steadily better at each digit. It just hasn’t gotten enough of them right at once to score a single point yet. Swap in a metric that gives partial credit and the leap melts into a gentle slope. Across a large benchmark suite, the great majority of catalogued “emergent” abilities only showed up under those discontinuous, pass-or-fail metrics. Change the ruler, lose the cliff.

It’s a good argument, and it traveled. A congressional committee cited it as having debunked emergence through statistical rigor. It’s also where most takes stop: another deflation, another “the hype was measurement error,” tidy and eminently shareable.

Why the debunking is too clean

Three complications.

First, and almost funny: the researchers who coined “emergent abilities” had already written that all-or-nothing metrics can dress incremental gains up as sudden ones. The mirage was a caveat in the original paper before it was a rebuttal to it.

Second, and sharper: the smooth metrics that dissolve the cliff were chosen after the cliffs were found. Re-describing a surprise as inevitable once it has already happened gives you a cleaner chart and zero foresight. The question was never “can we redraw the jump as a slope in hindsight.” It was “can we say where the line gets crossed before we train the bigger model.” Smoothing the past does not answer that.

Third, the one that matters if you ship things: from where the user sits, the discontinuity is real no matter what the underlying curve is doing. A coding agent that finishes a multi-step task unattended seventy percent of the time is something you babysit, net negative, because checking its work costs more than the work. The same agent at ninety-plus percent is something you let run. The capability curve under that gap may be perfectly smooth. Nobody deploying the thing feels the smooth curve. They feel the week it crossed from toy to coworker. The threshold is where the value lives, and a threshold behaves like a cliff even when the math behind it is a ramp.

So the mirage camp is right that the capability climbs smoothly, and the emergence camp is right that the arrival feels sharp and unforeseen, and those two claims were never in conflict. Same slope, seen from the mapmaker’s desk and from the trailhead. The live disagreement isn’t whether the curve is smooth. It’s whether the crossing can be called in advance. And that is the part that’s actually changing.

What changed: we’re learning to see it coming

Here is what the tidy debunking missed and the breathless hype never expected. The field has started, partially, to predict emergence before it happens.

The trick, from a 2024 result, is almost cheeky. Finetune a small model a little on the target task and you drag the point where the capability emerges down toward smaller scales, close enough to watch it happen in models you can actually afford to train. Do that across a few sizes and you can fit what the authors call emergence laws: a curve that forecasts the scale at which a larger, not-yet-built model will cross from random-guessing to real competence. Posed as a question, can we tell whether the next model will do a thing none of today’s models can, it is no longer hopeless. By 2025 a follow-on did it specifically for software-engineering tasks, fitting scaling-law forecasts for coding performance off finetuned mid-sized models, and named the obvious double edge: the same method that lets you forecast a capability lets you summon it early, which is a safety problem as much as a planning convenience.

This is the update that reframes the debate, and it’s worth saying flatly because both camps got it partly wrong. Emergence was never pure illusion and never pure magic. It is becoming, slowly, an engineering quantity, something you can put error bars on. Not reliably, not for every capability, not yet. But the research points at a world where a lab can say, before a run, roughly which thresholds the next model will clear.

What won’t resolve: the dice

And then there’s the finding that keeps the whole thing honest, the one that should stop anyone from getting comfortable on either side. Scaling-law charts almost always plot a single training run per model size. Run the same recipe again with a different random seed and the scale at which a capability emerges can move. Same data, same architecture, same compute, different roll, different crossing point. Breakthrough behavior turns out to be partly a property of the particular run, not just the size. That cuts against the emergence romantics, for whom the jump is a deep fact about scale, and against the mirage deflators, for whom enough careful measurement makes everything smooth and foreseeable. Some of the unpredictability isn’t ignorance we’ll eventually clear away. It’s baked into the process.

Then widen the lens, because the surprises don’t only happen along a known axis. Sometimes a new axis appears. Through late 2024 the field was drafting scaling’s obituary: the returns from raw pretraining size were flattening, a leading figure declared that pretraining as we know it would end, essays ran under titles like “the slow death of scaling.” Then capability started climbing again from a direction the obituaries hadn’t priced in, letting the model think longer at the moment you ask it, spending compute on reasoning at inference instead of on parameters at training. A second power law, a second knob, and almost nobody forecast that it would be the one to turn next.

So predictability is improving on the dimension you’re watching, and the ground still shifts under you in two ways the forecasts struggle with: which run gets you there, and which axis matters next.

The only thing that doesn’t emerge

Put those together and the question worth asking, if you build on these models, sharpens into something specific. Not “is emergence real.” Not even “can it be predicted,” that one now has a partial, improving, genuinely useful answer. The question is this: granted you can forecast a capability curve in the aggregate, the exact crossing for the exact task you care about still rides on a roll you don’t control and, every so often, an axis you didn’t see coming. So who is positioned to notice the week it actually lands for your problem, and to already know what they’d do?

Because the capability emerges on its own. Forecast or surprise, smooth or seed-lucky, it shows up without your help. The judgment about what it’s now good enough for does not. Noticing that an agent quietly crossed from toy to coworker this month, knowing which threshold is load-bearing for your product and which is a benchmark vanity number, having the taste to rebuild around the crossing instead of waiting for a changelog to bless it, none of that emerges from scale. It stays exactly as scarce as it was, and a little more valuable each time the thing it gates gets cheaper.

The headlines will keep asking whether the models are really getting smarter or whether we keep fooling ourselves with the metrics. It’s the wrong fight, and it has been for a while. The curve is smooth and the surprise is real, both at once: we can increasingly forecast the shape of the climb and still can’t tell you which run, which week, which task tips over first. The most consequential number in the field, when does the next capability cross into useful for the thing you’re building, comes with error bars now instead of a shrug. It still doesn’t come with a date.

So you don’t get to schedule it. You get to be the builder paying close enough attention to recognize it the moment it lands, and to already know what you’d do with it. The intelligence will emerge on its own timeline. What you bring is the only thing that was ever going to stay scarce: knowing what it’s for.