Hey readers,
What makes GPT-3 better than GPT-2, and GPT-2 better than GPT-1? The most important thing is this: it’s bigger, with GPT-3 more than 100 times larger than GPT-1 and more than 10 times larger than GPT-2.
But what does "bigger" mean, here? And why does being bigger make a model better? And, perhaps most importantly, will bigger always mean better, or will we at some point run into a law of diminishing returns?
These might seem like fairly abstruse technical questions — and they are. But they’re also incredibly consequential. Some people are betting big — to the tune of billions of dollars — that adding more data and more parameters to AI models will continue to make them better, as it has to date. Others are claiming that scaling laws — the principles about how the size of a machine learning model relates to its performance — have to break down sometime soon. The answer may determine whether machine learning lives up to the hype.
Scaling laws, explained
“Common knowledge suggests that larger models with more data will perform better, frequently reiterated in phrases like ‘add more layers’ or ‘use more data,’” reads an upcoming machine learning (ML) textbook by University of Berkeley ML researcher Dan Hendryks (when I reached out to ask him about scaling laws, he gave me advance access to the soon-to-be-released textbook, which was a profoundly humbling experience.) “Scaling laws make this folk knowledge mathematically precise.”
There are two senses in which a model can get bigger. One is the parameter count of a model — loosely equivalent to the number of neurons, if you imagine a deep learning model as analogous to a human brain. The other is the size of the dataset the model is fed — a model trained on the whole internet will be better at text prediction than one fed only on a small sample of books.
It makes sense that enhancing either of these two features of a model would improve its performance. But what’s surprising is that the relationship among “more data” and “more parameters” and “better performance” turns out to be mathematically predictable — and about the same whether you’re going from 10 million parameters to 100 million parameters or from 1 billion to 10 billion.
Measuring how good our ML systems are
We often measure the performance of ML systems by scoring them on, basically, standardized tests, called benchmarks. A benchmark is a set of questions that are difficult for modern ML systems to answer, so that researchers can test their models against the benchmarks and demonstrate if their model is higher-performing than other models. The MATH benchmark, as you might expect, measures ability on math problems. The GLUE benchmark is a collection of tasks meant to evaluate general language understanding. (When models started performing at human levels on GLUE, SuperGLUE was introduced as its successor.)
What predicts how well a model will perform on a benchmark? Specialized training and well-chosen data can make a difference, as can algorithm improvements. But on a specific dataset, a model will generally do significantly better if it has more parameters or more data. And as we’ve built bigger and bigger models, again and again they’ve turned out better and better, in accordance with the scaling laws that seem to govern them. (A note: this works for language models and similar token-predicting AIs; the same principles don’t seem to apply to generative models, the ones like DALL-E and Midjourney that rocked the art world this spring.)
Some researchers have argued that the human brain itself can best be conceptualized as a scaled-up primate brain, following scaling laws in how increased neuron size improves its performance. Whether or not that’s true, scaling laws do seem to provoke the intuition that apparent intelligence often comes down to sheer bigness.
What scaling laws can and can’t tell us
“Scaling laws can let you forecast that if you make your model 10 times bigger, it will go from, say, 71 percent accuracy to 93 percent accuracy on a specific dataset, like contest programming problems,” Ajeya Cotra, a senior program officer at Open Philanthropy who studies potential risks from advanced AI, told me.
But this just demonstrates the inadequacy of our current benchmarks to capture what we care about in AI systems. If the question you care about is whether an AI system is going to be economically disruptive — or whether it’s going to be dangerous — scaling laws alone can’t answer that for you.
“If your model will have 93 percent performance on a coding contest, does that mean it will be able to automate a junior programmer’s job? That it will be able to find and exploit zero-day vulnerabilities in real software?” asks Cotra. “If you’ve extrapolated that your model will have 96 percent performance on a hard biology test, how much should you worry that it will be able to help terrorists develop novel bioweapons?”
The end is … somewhere?
The improvements we’re seeing in model capabilities with increasing model size can’t last forever. Presumably, at some point adding more parameters and more data won’t continue to improve model performance with the striking regularity that it presently does. And for years, people have been insisting that point has to be right around the corner.
So far, though, they’ve been wrong — and that has led to even leading ML experts repeatedly underestimating progress in their own field.
But all along, some people have been betting that the scaling laws will stay in force — and that means it's worthwhile to spend more and more eye-popping sums on AI. “Are we going to hit the limits of scaling laws?” Dario Amodei, the CEO of Ai startup Anthropic, was asked at a recent conference. “Not anytime soon,” he asserted. “Right now the most expensive model costs plus or minus $100 million. Next year we will have $1 billion models. By 2025, we may have a $10 billion model.”
That will make for an expensive test of the scaling hypothesis. Even harder to guess is what it’ll mean for the world. “As we see bigger numbers on well-defined metrics, we also see bigger real-world impacts,” Cotra told me. “But we haven’t quantified that relationship.” Unless the relationship between model size and model ability breaks down entirely, the models of 2025 will be smarter, savvier, better at problem solving. They’ll probably ace math problems that I would stand no chance at, and score higher on medical exams than most doctors.
But a model isn’t just its test scores, and guessing the effect they’ll have on the world is much harder. I will say that none of my deep dive into scaling laws dissuaded me from the notion that humanity is playing with fire on an extraordinary scale.
—Kelsey Piper, senior writer
Questions? Comments? Tell us what you think! Email us at futureperfect@vox.com.
And if you want to recommend this newsletter to your friends or colleagues, tell them to sign up at vox.com/future-perfect-newsletter.
