AI 的缩放定律到底是什么？So What Exactly Is the Scaling Law in AI?

Nothing magical makes AI smarter — mathematics does. Here are the underlying scaling laws that govern how model size, training data and compute combine to produce intelligence.

Watching each year's models arrive markedly sharper than the last raises an obvious question — and the explanation is not simply that engineers write cleaner code. A mathematical principle called the scaling law sits behind it. That principle is what carried AI out of obscure academic corners and turned it into technology that reshapes the world. So what does the scaling law actually say, and how does it operate?

Put plainly, scaling laws show that a smarter system does not require a dazzling new algorithm. What it requires is more training material, more computing power and a larger model. Machine learning, in other words, obeys a "more is better" principle — and the curves that describe it are both strict and forecastable.

01The Central Idea: AI's Closest Thing to an Iron Law

Think about teaching a child to read. One book yields a handful of words; a thousand books produce a devoted reader; a vast library plus an exceptional tutor might produce a literary prodigy. AI follows much the same logic — only with numbers that behave predictably.

Researchers at OpenAI put out a paper in 2020 that changed the field, titled "Scaling Laws for Neural Language Models." Its finding: model performance, gauged by how reliably a system predicts the following word in a sentence, traces a smooth power-law curve that can be forecast in advance.

The startling part is that the curve is universal. Change the architecture, add whatever clever tricks you like to the code — plot model size against performance and the same predictable line reappears. Performance, then, is no mystery; it is a matter of engineering at scale.

02Three Levers That Drive AI Scaling

Grasping how AI improves means understanding the three quantities researchers are able to dial up or down. Call them the supports on which the scaling law rests:

03How the Idea Developed: OpenAI, Then Chinchilla

The industry sat up when OpenAI's scaling laws appeared in 2020. A race began, with every lab trying to build the biggest model. Billions went into enormous training runs, on the assumption that nothing but size counted.

Then in 2022 came a bombshell from researchers at Google DeepMind, a paper named "Chinchilla." It showed the industry had been going about things the wrong way. Reading the original laws, labs had built models as large as possible while supplying comparatively little data; DeepMind demonstrated how wasteful that approach was.

What Chinchilla Changed

What Chinchilla established is that data and model size have to grow in step. Double the model and the training data must double too. Skip that and the result is an "under-trained" system: an outsized brain that has not read enough to put its capacity to work.

That insight reset priorities. Rather than chasing size alone, labs hunted for additional data and trained leaner, more efficient models for longer stretches. Our weekly AI research roundup tracks how these efficient architectures are shifting the balance.

04Running Into the Data Wall

A serious difficulty hangs over the whole field: the supply of data is running down. Scaling laws call for ever more text, yet nearly everything on the open internet that is worth reading has already been written.

By 2026 or 2027, on expert estimates, the stock of high-quality public text will be used up. The phrase for this is the "Data Wall." When a law presupposes unlimited data and the supply is finite, does the curve go flat? Does progress in AI simply halt?

Companies are working around the wall with synthesised data and datasets held privately. Our coverage of the newest breakthrough AI research goes further into the fresh architectures and synthetic-data approaches aimed at the shortage.

05Test-Time Compute: Scaling of a Different Kind

Since the training phase has run up against the data wall, the field has found another route to greater intelligence: test-time compute.

Rather than enlarging the original training run, researchers now hand the model extra compute at the moment it is answering you. Picture a maths exam. An ordinary AI writes down whatever answer surfaces first. One scaled at test time will take 30 seconds to "think" — reviewing its working, trying alternate routes through the logic, fixing its own errors — before committing to an answer.

Our guide to reasoning AI and the way it functions unpacks this move toward models that "think." Extra compute spent at inference lets AI crack mathematical, programming and scientific problems that used to be out of reach — a genuinely new curve that scales upward rather than outward.

06Will Scaling Alone Get Us to AGI?

The question underneath all the others: do scaling laws eventually deliver Artificial General Intelligence (AGI), a system that beats people at essentially any mental task?

Ilya Sutskever and Sam Altman are among the prominent backers of the "Scaling Hypothesis." Their argument runs that AGI arrives on its own if the three supports (Data, Compute, Parameters) keep growing — a mathematical consequence rather than a lucky discovery. On this view, intelligence is simply a function of scale.

Sceptics counter that diminishing returns are inevitable. Genuine AGI, they say, will demand architectural breakthroughs of an entirely different order — not merely more data poured into a bigger network. Plenty of researchers hold that getting to grips with these scaling dynamics is the only practical way to settle what AGI is and whether it has arrived, and the argument remains among the fieriest in computer science.

07Questions Readers Ask Most