The core of an artificial intelligence program like ChatGPT is something called a large language model: an algorithm that mimics the form of written language.
While the inner workings of these algorithms are notoriously difficult to decipher, the basic idea behind them is surprisingly simple. They are trained on mountains of internet text, by going through them a few sentences or paragraphs at a time, repeatedly guessing the next word (or word fragment) and then grading themselves against the real thing.
To show you what this process looks like, we trained six tiny language models starting from scratch. To begin, choose what you’d like to see the A.I. learn by selecting one of the images below. (You can always change your mind later.)
Before training: Gibberish
At the outset, BabyGPT produces text like this:
The largest language models are trained on over a terabyte of internet text, containing hundreds of billions of words. Their training costs millions of dollars and involves calculations that take weeks or even months on hundreds of specialized computers.
BabyGPT is ant-sized in comparison. We trained it for about an hour on a laptop on just a few megabytes of text — small enough to attach to an email.
Unlike the larger models, which start their training with a large vocabulary, BabyGPT doesn’t yet know any words. It makes its guesses one letter at a time, which makes it a bit easier for us to see what it’s learning.
Initially, its guesses are completely random and include lots of special characters: ‘?kZhc,TK996’) would make a great password, but it’s a far cry from anything resembling Jane Austen or Shakespeare. BabyGPT hasn’t yet learned which letters are typically used in English, or that words even exist.
This is how language models usually start off: They guess randomly and produce gibberish. But they learn from their mistakes, and over time, their guesses get better. Over many, many rounds of training, language models can learn to write. They learn statistical patterns that piece words together into sentences and paragraphs.
After 250 rounds: English letters
After 250 rounds of training — about 30 seconds of processing on a modern laptop — BabyGPT has learned its ABCs and is starting to babble:
In particular, our model has learned which letters are most frequently used in the text. You’ll see a lot of the letter “e” because that is the most common letter in English.
If you look closely, you’ll find that it has also learned some small words: I, to, the, you, and so on.
It has a tiny vocabulary, but that doesn’t stop it from inventing words like alingedimpe, ratlabus and mandiered.
Obviously, these guesses aren’t great. But — and this is a key to how a language model learns — BabyGPT keeps a score of exactly how bad its guesses are.
Every round of training, it goes through the original text, a few words at a time, and compares its guesses for the next letter with what actually comes next. It then calculates a score, known as the “loss,” which measures the difference between its predictions and the actual text. A loss of zero would mean that its guesses always correctly matched the next letter. The smaller the loss, the closer its guesses are to the text.
After 500 rounds: Small words
Each training round, BabyGPT tries to improve its guesses by reducing this loss. After 500 rounds — or about a minute on a laptop — it can spell a few small words:
It’s also starting to learn some basic grammar, like where to place periods and commas. But it makes plenty of mistakes. No one is going to confuse this output with something written by a human being.
After 5,000 rounds: Bigger words
Ten minutes in, BabyGPT’s vocabulary has grown:
The sentences don’t make sense, but they’re getting closer in style to the text. BabyGPT now makes fewer spelling mistakes. It still invents some longer words, but less often than it once did. It’s also starting to learn some names that occur frequently in the text.
Its grammar is improving, too. For example, it has learned that a period is often followed by a space and a capital letter. It even occasionally opens a quote (although it often forgets to close it).
Behind the scenes, BabyGPT is a neural network: an extremely complicated type of mathematical function involving millions of numbers that converts an input (in this case, a sequence of letters) into an output (its prediction for the next letter).
Every round of training, an algorithm adjusts these numbers to try to improve its guesses, using a mathematical technique known as backpropagation. The process of tuning these internal numbers to improve predictions is what it means for a neural network to “learn.”
What this neural network actually generates is not letters but probabilities. (These probabilities are why you get a different answer each time you generate a new response.)
For example, when given the letters stai, it’ll predict that the next letter is n, r or maybe d, with probabilities that depend on how often it has encountered each word in its training.
But if we give it downstai, it’s much more likely to predict r. Its predictions depend on the context.
After 30,000 rounds: Full sentences
An hour into its training, BabyGPT is learning to speak in full sentences. That’s not so bad, considering that just an hour ago, it didn’t even know that words existed!
The words still don’t make sense, but they definitely look more like English.
The sentences that this neural network generates rarely occur in the original text. It usually doesn’t copy and paste sentences verbatim; instead, BabyGPT stitches them together, letter by letter, based on statistical patterns that it has learned from the data. (Typical language models stitch sentences together a few letters at a time, but the idea is the same.)
As language models grow larger, the patterns that they learn can become increasingly complex. They can learn the form of a sonnet or a limerick, or how to code in various programming languages.
Line chart showing the “loss” of the selected model over time. Each model starts off with a high loss producing gibberish characters. Over the next few hundred rounds of training, the loss declines precipitously and the model starts to produce English letters and a few small words. The loss then drops off gradually, and the model produces bigger words after 5,000 rounds of training. At this point, there are diminishing returns, and the curve is fairly flat. By 30,000 rounds, the model is making full sentences.
The limits to BabyGPT’s learning
With limited text to work with, BabyGPT doesn’t benefit much from further training. Larger language models use more data and computing power to mimic language more convincingly.
Loss estimates are slightly smoothed.
BabyGPT still has a long way to go before its sentences become coherent or useful. It can’t answer a question or debug your code. It’s mostly just fun to watch its guesses improve.
But it’s also instructive. In just an hour of training on a laptop, a language model can go from generating random characters to a very crude approximation of language.
Language models are a kind of universal mimic: They imitate whatever they’ve been trained on. With enough data and rounds of training, this imitation can become fairly uncanny, as ChatGPT and its peers have shown us.
What even is a GPT?
The models trained in this article use an algorithm called nanoGPT, developed by Andrej Karpathy. Mr. Karpathy is a prominent A.I. researcher who recently joined OpenAI, the company behind ChatGPT.
Like ChatGPT, nanoGPT is a GPT model, an A.I. term that stands for generative pre-trained transformer:
Generative because it generates words.
Pre-trained because it’s trained on a bunch of text. This step is called pre-training because many language models (like the one behind ChatGPT) go through important additional stages of training known as fine-tuning to make them less toxic and easier to interact with.
Transformers are a relatively recent breakthrough in how neural networks are wired. They were introduced in a 2017 paper by Google researchers, and are used in many of the latest A.I. advancements, from text generation to image creation.
Transformers improved upon the previous generation of neural networks — known as recurrent neural networks — by including steps that process the words of a sentence in parallel, rather than one at a time. This made them much faster.
More is different
Other than the additional fine-tuning stages, the primary difference between nanoGPT and the language model underlying chatGPT is size.
For example, GPT-3 was trained on up to a million times as many words as the models in this article. Scaling up to that size is a huge technical undertaking, but the underlying principles remain the same.
As language models grow in size, they are known to develop surprising new abilities, such as the ability to answer questions, summarize text, explain jokes, continue a pattern and correct bugs in computer code.
Some researchers have termed these “emergent abilities” because they arise unexpectedly at a certain size and are not programmed in by hand. The A.I. researcher Sam Bowman has likened training a large language model to “buying a mystery box,” because it is difficult to predict what skills it will gain during its training, and when these skills will emerge.
Undesirable behaviors can emerge as well. Large language models can become highly unpredictable, as evidenced by Microsoft Bing A.I.’s early interactions with my colleague Kevin Roose.
They are also prone to inventing facts and reasoning incorrectly. Researchers do not yet understand how these models generate language, and they struggle to steer their behavior.
Nearly four months after OpenAI’s ChatGPT was made public, Google launched an A.I. chatbot called Bard, over safety objections from some of its employees, according to reporting by Bloomberg.
“These models are being developed in an arms race between tech companies, without any transparency,” said Peter Bloem, an A.I. expert who studies language models.
OpenAI does not disclose any details on the data that its enormous GPT-4 model is trained on, citing concerns about competition and safety. Not knowing what’s in the data makes it hard to tell if these technologies are safe, and what kinds of biases are embedded inside them.
But while Mr. Bloem has concerns about the lack of A.I. regulation, he is also excited that computers are finally starting to “understand what we want them to do” — something that, he says, researchers hadn’t been close to achieving in over 70 years of trying.
