It’s Just Adding One Word at a Time

That ChatGPT can automatically generate something that reads even superficially like human-written text is remarkable, and unexpected. But how does it do it? And why does it work? My purpose here is to give a rough outline of what’s going on inside ChatGPT—and then to explore why it is that it can do so well in producing what we might consider to be meaningful text. I should say at the outset that I’m going to focus on the big picture of what’s going on—and while I’ll mention some engineering details, I won’t get deeply into them. (And the essence of what I’ll say applies just as well to other current “large language models” [LLMs] as to ChatGPT.)

The first thing to explain is that what ChatGPT is always fundamentally trying to do is to produce a “reasonable continuation” of whatever text it’s got so far, where by “reasonable” we mean “what one might expect someone to write after seeing what people have written on billions of webpages, etc.”

So let’s say we’ve got the text “The best thing about AI is its ability to”. Imagine scanning billions of pages of human-written text (say on the web and in digitized books) and finding all instances of this text—then seeing what word comes next what fraction of the time. ChatGPT effectively does something like this, except that (as I’ll explain) it doesn’t look at literal text; it looks for things that in a certain sense “match in meaning”. But the end result is that it produces a ranked list of words that might follow, together with “probabilities”:

And the remarkable thing is that when ChatGPT does something like write an essay what it’s essentially doing is just asking over and over again “given the text so far, what should the next word be?”—and each time adding a word. (More precisely, as I’ll explain, it’s adding a “token”, which could be just a part of a word, which is why it can sometimes “make up new words”.)

But, OK, at each step it gets a list of words with probabilities. But which one should it actually pick to add to the essay (or whatever) that it’s writing? One might think it should be the “highest-ranked” word (i.e. the one to which the highest “probability” was assigned). But this is where a bit of voodoo begins to creep in. Because for some reason—that maybe one day we’ll have a scientific-style understanding of—if we always pick the highest-ranked word, we’ll typically get a very “flat” essay, that never seems to “show any creativity” (and even sometimes repeats word for word). But if sometimes (at random) we pick lower-ranked words, we get a “more interesting” essay.

The fact that there’s randomness here means that if we use the same prompt multiple times, we’re likely to get different essays each time. And, in keeping with the idea of voodoo, there’s a particular so-called “temperature” parameter that determines how often lower-ranked words will be used, and for essay generation, it turns out that a “temperature” of 0.8 seems best. (It’s worth emphasizing that there’s no “theory” being used here; it’s just a matter of what’s been found to work in practice. And for example the concept of “temperature” is there because exponential distributions familiar from statistical physics happen to be being used, but there’s no “physical” connection—at least so far as we know.)

Before we go on I should explain that for purposes of exposition I’m mostly not going to use the full system that’s in ChatGPT; instead I’ll usually work with a simpler GPT-2 system, which has the nice feature that it’s small enough to be able to run on a standard desktop computer. And so for essentially everything I show I’ll be able to include explicit Wolfram Language code that you can immediately run on your computer. (Click any picture here to copy the code behind it.)

For example, here’s how to get the table of probabilities above. First, we have to retrieve the underlying “language model” neural net:

Later on, we’ll look inside this neural net, and talk about how it works. But for now we can just apply this “net model” as a black box to our text so far, and ask for the top 5 words by probability that the model says should follow:

This takes that result and makes it into an explicit formatted “dataset”:

Here’s what happens if one repeatedly “applies the model”—at each step adding the word that has the top probability (specified in this code as the “decision” from the model):

What happens if one goes on longer? In this (“zero temperature”) case what comes out soon gets rather confused and repetitive:

But what if instead of always picking the “top” word one sometimes randomly picks “non-top” words (with the “randomness” corresponding to “temperature” 0.8)? Again one can build up text:

And every time one does this, different random choices will be made, and the text will be different—as in these 5 examples:

It’s worth pointing out that even at the first step there are a lot of possible “next words” to choose from (at temperature 0.8), though their probabilities fall off quite quickly (and, yes, the straight line on this log-log plot corresponds to an n–1 “power-law” decay that’s very characteristic of the general statistics of language):

So what happens if one goes on longer? Here’s a random example. It’s better than the top-word (zero temperature) case, but still at best a bit weird:

This was done with the simplest GPT-2 model (from 2019). With the newer and bigger GPT-3 models the results are better. Here’s the top-word (zero temperature) text produced with the same “prompt”, but with the biggest GPT-3 model:

And here’s a random example at “temperature 0.8”:

Where Do the Probabilities Come From?

OK, so ChatGPT always picks its next word based on probabilities. But where do those probabilities come from? Let’s start with a simpler problem. Let’s consider generating English text one letter (rather than word) at a time. How can we work out what the probability for each letter should be?

A very minimal thing we could do is just take a sample of English text, and calculate how often different letters occur in it. So, for example, this counts letters in the Wikipedia article on “cats”:

And this does the same thing for “dogs”:

The results are similar, but not the same (“o” is no doubt more common in the “dogs” article because, after all, it occurs in the word “dog” itself). Still, if we take a large enough sample of English text we can expect to eventually get at least fairly consistent results:

Here’s a sample of what we get if we just generate a sequence of letters with these probabilities:

We can break this into “words” by adding in spaces as if they were letters with a certain probability:

We can do a slightly better job of making “words” by forcing the distribution of “word lengths” to agree with what it is in English:

We didn’t happen to get any “actual words” here, but the results are looking slightly better. To go further, though, we need to do more than just pick each letter separately at random. And, for example, we know that if we have a “q”, the next letter basically has to be “u”.

Here’s a plot of the probabilities for letters on their own:

And here’s a plot that shows the probabilities of pairs of letters (“2-grams”) in typical English text. The possible first letters are shown across the page, the second letters down the page:

And we see here, for example, that the “q” column is blank (zero probability) except on the “u” row. OK, so now instead of generating our “words” a single letter at a time, let’s generate them looking at two letters at a time, using these “2-gram” probabilities. Here’s a sample of the result—which happens to include a few “actual words”:

With sufficiently much English text we can get pretty good estimates not just for probabilities of single letters or pairs of letters (2-grams), but also for longer runs of letters. And if we generate “random words” with progressively longer n-gram probabilities, we see that they get progressively “more realistic”:

But let’s now assume—more or less as ChatGPT does—that we’re dealing with whole words, not letters. There are about 40,000 reasonably commonly used words in English. And by looking at a large corpus of English text (say a few million books, with altogether a few hundred billion words), we can get an estimate of how common each word is. And using this we can start generating “sentences”, in which each word is independently picked at random, with the same probability that it appears in the corpus. Here’s a sample of what we get:

Not surprisingly, this is nonsense. So how can we do better? Just like with letters, we can start taking into account not just probabilities for single words but probabilities for pairs or longer n-grams of words. Doing this for pairs, here are 5 examples of what we get, in all cases starting from the word “cat”:

It’s getting slightly more “sensible looking”. And we might imagine that if we were able to use sufficiently long n-grams we’d basically “get a ChatGPT”—in the sense that we’d get something that would generate essay-length sequences of words with the “correct overall essay probabilities”. But here’s the problem: there just isn’t even close to enough English text that’s ever been written to be able to deduce those probabilities.

In a crawl of the web there might be a few hundred billion words; in books that have been digitized there might be another hundred billion words. But with 40,000 common words, even the number of possible 2-grams is already 1.6 billion—and the number of possible 3-grams is 60 trillion. So there’s no way we can estimate the probabilities even for all of these from text that’s out there. And by the time we get to “essay fragments” of 20 words, the number of possibilities is larger than the number of particles in the universe, so in a sense they could never all be written down.

So what can we do? The big idea is to make a model that lets us estimate the probabilities with which sequences should occur—even though we’ve never explicitly seen those sequences in the corpus of text we’ve looked at. And at the core of ChatGPT is precisely a so-called “large language model” (LLM) that’s been built to do a good job of estimating those probabilities.

What Is a Model?

Say you want to know (as Galileo did back in the late 1500s) how long it’s going to take a cannon ball dropped from each floor of the Tower of Pisa to hit the ground. Well, you could just measure it in each case and make a table of the results. Or you could do what is the essence of theoretical science: make a model that gives some kind of procedure for computing the answer rather than just measuring and remembering each case.

Let’s imagine we have (somewhat idealized) data for how long the cannon ball takes to fall from various floors:

How do we figure out how long it’s going to take to fall from a floor we don’t explicitly have data about? In this particular case, we can use known laws of physics to work it out. But say all we’ve got is the data, and we don’t know what underlying laws govern it. Then we might make a mathematical guess, like that perhaps we should use a straight line as a model:

We could pick different straight lines. But this is the one that’s on average closest to the data we’re given. And from this straight line we can estimate the time to fall for any floor.

How did we know to try using a straight line here? At some level we didn’t. It’s just something that’s mathematically simple, and we’re used to the fact that lots of data we measure turns out to be well fit by mathematically simple things. We could try something mathematically more complicated—say a + b x + c x2—and then in this case we do better:

Things can go quite wrong, though. Like here’s the best we can do with a + b/x + c sin(x):

It is worth understanding that there’s never a “model-less model”. Any model you use has some particular underlying structure—then a certain set of “knobs you can turn” (i.e. parameters you can set) to fit your data. And in the case of ChatGPT, lots of such “knobs” are used—actually, 175 billion of them.

But the remarkable thing is that the underlying structure of ChatGPT—with “just” that many parameters—is sufficient to make a model that computes next-word probabilities “well enough” to give us reasonable essay-length pieces of text.

Models for Human-Like Tasks

The example we gave above involves making a model for numerical data that essentially comes from simple physics—where we’ve known for several centuries that “simple mathematics applies”. But for ChatGPT we have to make a model of human-language text of the kind produced by a human brain. And for something like that we don’t (at least yet) have anything like “simple mathematics”. So what might a model of it be like?

Before we talk about language, let’s talk about another human-like task: recognizing images. And as a simple example of this, let’s consider images of digits (and, yes, this is a classic machine learning example):

One thing we could do is get a bunch of sample images for each digit:

Then to find out if an image we’re given as input corresponds to a particular digit we could just do an explicit pixel-by-pixel comparison with the samples we have. But as humans we certainly seem to do something better—because we can still recognize digits, even when they’re for example handwritten, and have all sorts of modifications and distortions:

When we made a model for our numerical data above, we were able to take a numerical value x that we were given, and just compute a + b x for particular a and b. So if we treat the gray-level value of each pixel here as some variable xi is there some function of all those variables that—when evaluated—tells us what digit the image is of? It turns out that it’s possible to construct such a function. Not surprisingly, it’s not particularly simple, though. And a typical example might involve perhaps half a million mathematical operations.

But the end result is that if we feed the collection of pixel values for an image into this function, out will come the number specifying which digit we have an image of. Later, we’ll talk about how such a function can be constructed, and the idea of neural nets. But for now let’s treat the function as black box, where we feed in images of, say, handwritten digits (as arrays of pixel values) and we get out the numbers these correspond to:

But what’s really going on here? Let’s say we progressively blur a digit. For a little while our function still “recognizes” it, here as a “2”. But soon it “loses it”, and starts giving the “wrong” result:

But why do we say it’s the “wrong” result? In this case, we know we got all the images by blurring a “2”. But if our goal is to produce a model of what humans can do in recognizing images, the real question to ask is what a human would have done if presented with one of those blurred images, without knowing where it came from.

And we have a “good model” if the results we get from our function typically agree with what a human would say. And the nontrivial scientific fact is that for an image-recognition task like this we now basically know how to construct functions that do this.

Can we “mathematically prove” that they work? Well, no. Because to do that we’d have to have a mathematical theory of what we humans are doing. Take the “2” image and change a few pixels. We might imagine that with only a few pixels “out of place” we should still consider the image a “2”. But how far should that go? It’s a question of human visual perception. And, yes, the answer would no doubt be different for bees or octopuses—and potentially utterly different for putative aliens.

Neural Nets

OK, so how do our typical models for tasks like image recognition actually work? The most popular—and successful—current approach uses neural nets. Invented—in a form remarkably close to their use today—in the 1940s, neural nets can be thought of as simple idealizations of how brains seem to work.

In human brains there are about 100 billion neurons (nerve cells), each capable of producing an electrical pulse up to perhaps a thousand times a second. The neurons are connected in a complicated net, with each neuron having tree-like branches allowing it to pass electrical signals to perhaps thousands of other neurons. And in a rough approximation, whether any given neuron produces an electrical pulse at a given moment depends on what pulses it’s received from other neurons—with different connections contributing with different “weights”.

When we “see an image” what’s happening is that when photons of light from the image fall on (“photoreceptor”) cells at the back of our eyes they produce electrical signals in nerve cells. These nerve cells are connected to other nerve cells, and eventually the signals go through a whole sequence of layers of neurons. And it’s in this process that we “recognize” the image, eventually “forming the thought” that we’re “seeing a 2” (and maybe in the end doing something like saying the word “two” out loud).

The “black-box” function from the previous section is a “mathematicized” version of such a neural net. It happens to have 11 layers (though only 4 “core layers”):

There’s nothing particularly “theoretically derived” about this neural net; it’s just something that—back in 1998—was constructed as a piece of engineering, and found to work. (Of course, that’s not much different from how we might describe our brains as having been produced through the process of biological evolution.)

OK, but how does a neural net like this “recognize things”? The key is the notion of attractors. Imagine we’ve got handwritten images of 1’s and 2’s:

We somehow want all the 1’s to “be attracted to one place”, and all the 2’s to “be attracted to another place”. Or, put a different way, if an image is somehow “closer to being a 1” than to being a 2, we want it to end up in the “1 place” and vice versa.

As a straightforward analogy, let’s say we have certain positions in the plane, indicated by dots (in a real-life setting they might be positions of coffee shops). Then we might imagine that starting from any point on the plane we’d always want to end up at the closest dot (i.e. we’d always go to the closest coffee shop). We can represent this by dividing the plane into regions (“attractor basins”) separated by idealized “watersheds”:

We can think of this as implementing a kind of “recognition task” in which we’re not doing something like identifying what digit a given image “looks most like”—but rather we’re just, quite directly, seeing what dot a given point is closest to. (The “Voronoi diagram” setup we’re showing here separates points in 2D Euclidean space; the digit recognition task can be thought of as doing something very similar—but in a 784-dimensional space formed from the gray levels of all the pixels in each image.)

So how do we make a neural net “do a recognition task”? Let’s consider this very simple case:

Our goal is to take an “input” corresponding to a position {x,y}—and then to “recognize” it as whichever of the three points it’s closest to. Or, in other words, we want the neural net to compute a function of {x,y} like:

So how do we do this with a neural net? Ultimately a neural net is a connected collection of idealized “neurons”—usually arranged in layers—with a simple example being:

Each “neuron” is effectively set up to evaluate a simple numerical function. And to “use” the network, we simply feed numbers (like our coordinates x and y) in at the top, then have neurons on each layer “evaluate their functions” and feed the results forward through the network—eventually producing the final result at the bottom:

In the traditional (biologically inspired) setup each neuron effectively has a certain set of “incoming connections” from the neurons on the previous layer, with each connection being assigned a certain “weight” (which can be a positive or negative number). The value of a given neuron is determined by multiplying the values of “previous neurons” by their corresponding weights, then adding these up and adding a constant—and finally applying a “thresholding” (or “activation”) function. In mathematical terms, if a neuron has inputs x = {x1, x2 …} then we compute f[w . x + b], where the weights w and constant b are generally chosen differently for each neuron in the network; the function *f *is usually the same.

Computing w . x + b is just a matter of matrix multiplication and addition. The “activation function” f introduces nonlinearity (and ultimately is what leads to nontrivial behavior). Various activation functions commonly get used; here we’ll just use Ramp (or ReLU):

For each task we want the neural net to perform (or, equivalently, for each overall function we want it to evaluate) we’ll have different choices of weights. (And—as we’ll discuss later—these weights are normally determined by “training” the neural net using machine learning from examples of the outputs we want.)

Ultimately, every neural net just corresponds to some overall mathematical function—though it may be messy to write out. For the example above, it would be:

The neural net of ChatGPT also just corresponds to a mathematical function like this—but effectively with billions of terms.

But let’s go back to individual neurons. Here are some examples of the functions a neuron with two inputs (representing coordinates x and y) can compute with various choices of weights and constants (and Ramp as activation function):

But what about the larger network from above? Well, here’s what it computes:

It’s not quite “right”, but it’s close to the “nearest point” function we showed above.

Let’s see what happens with some other neural nets. In each case, as we’ll explain later, we’re using machine learning to find the best choice of weights. Then we’re showing here what the neural net with those weights computes:

Bigger networks generally do better at approximating the function we’re aiming for. And in the “middle of each attractor basin” we typically get exactly the answer we want. But at the boundaries—where the neural net “has a hard time making up its mind”—things can be messier.

With this simple mathematical-style “recognition task” it’s clear what the “right answer” is. But in the problem of recognizing handwritten digits, it’s not so clear. What if someone wrote a “2” so badly it looked like a “7”, etc.? Still, we can ask how a neural net distinguishes digits—and this gives an indication:

Can we say “mathematically” how the network makes its distinctions? Not really. It’s just “doing what the neural net does”. But it turns out that that normally seems to agree fairly well with the distinctions we humans make.

Let’s take a more elaborate example. Let’s say we have images of cats and dogs. And we have a neural net that’s been trained to distinguish them. Here’s what it might do on some examples:

Now it’s even less clear what the “right answer” is. What about a dog dressed in a cat suit? Etc. Whatever input it’s given the neural net will generate an answer, and in a way reasonably consistent with how humans might. As I’ve said above, that’s not a fact we can “derive from first principles”. It’s just something that’s empirically been found to be true, at least in certain domains. But it’s a key reason why neural nets are useful: that they somehow capture a “human-like” way of doing things.

Show yourself a picture of a cat, and ask “Why is that a cat?”. Maybe you’d start saying “Well, I see its pointy ears, etc.” But it’s not very easy to explain how you recognized the image as a cat. It’s just that somehow your brain figured that out. But for a brain there’s no way (at least yet) to “go inside” and see how it figured it out. What about for an (artificial) neural net? Well, it’s straightforward to see what each “neuron” does when you show a picture of a cat. But even to get a basic visualization is usually very difficult.

In the final net that we used for the “nearest point” problem above there are 17 neurons. In the net for recognizing handwritten digits there are 2190. And in the net we’re using to recognize cats and dogs there are 60,650. Normally it would be pretty difficult to visualize what amounts to 60,650-dimensional space. But because this is a network set up to deal with images, many of its layers of neurons are organized into arrays, like the arrays of pixels it’s looking at.

And if we take a typical cat image

then we can represent the states of neurons at the first layer by a collection of derived images—many of which we can readily interpret as being things like “the cat without its background”, or “the outline of the cat”:

By the 10th layer it’s harder to interpret what’s going on:

But in general we might say that the neural net is “picking out certain features” (maybe pointy ears are among them), and using these to determine what the image is of. But are those features ones for which we have names—like “pointy ears”? Mostly not.

Are our brains using similar features? Mostly we don’t know. But it’s notable that the first few layers of a neural net like the one we’re showing here seem to pick out aspects of images (like edges of objects) that seem to be similar to ones we know are picked out by the first level of visual processing in brains.

But let’s say we want a “theory of cat recognition” in neural nets. We can say: “Look, this particular net does it”—and immediately that gives us some sense of “how hard a problem” it is (and, for example, how many neurons or layers might be needed). But at least as of now we don’t have a way to “give a narrative description” of what the network is doing. And maybe that’s because it truly is computationally irreducible, and there’s no general way to find what it does except by explicitly tracing each step. Or maybe it’s just that we haven’t “figured out the science”, and identified the “natural laws” that allow us to summarize what’s going on.

We’ll encounter the same kinds of issues when we talk about generating language with ChatGPT. And again it’s not clear whether there are ways to “summarize what it’s doing”. But the richness and detail of language (and our experience with it) may allow us to get further than with images.

Machine Learning, and the Training of Neural Nets

We’ve been talking so far about neural nets that “already know” how to do particular tasks. But what makes neural nets so useful (presumably also in brains) is that not only can they in principle do all sorts of tasks, but they can be incrementally “trained from examples” to do those tasks.

When we make a neural net to distinguish cats from dogs we don’t effectively have to write a program that (say) explicitly finds whiskers; instead we just show lots of examples of what’s a cat and what’s a dog, and then have the network “machine learn” from these how to distinguish them.

And the point is that the trained network “generalizes” from the particular examples it’s shown. Just as we’ve seen above, it isn’t simply that the network recognizes the particular pixel pattern of an example cat image it was shown; rather it’s that the neural net somehow manages to distinguish images on the basis of what we consider to be some kind of “general catness”.

So how does neural net training actually work? Essentially what we’re always trying to do is to find weights that make the neural net successfully reproduce the examples we’ve given. And then we’re relying on the neural net to “interpolate” (or “generalize”) “between” these examples in a “reasonable” way.

Let’s look at a problem even simpler than the nearest-point one above. Let’s just try to get a neural net to learn the function:

For this task, we’ll need a network that has just one input and one output, like:

But what weights, etc. should we be using? With every possible set of weights the neural net will compute some function. And, for example, here’s what it does with a few randomly chosen sets of weights:

And, yes, we can plainly see that in none of these cases does it get even close to reproducing the function we want. So how do we find weights that will reproduce the function?

The basic idea is to supply lots of “input → output” examples to “learn from”—and then to try to find weights that will reproduce these examples. Here’s the result of doing that with progressively more examples:

At each stage in this “training” the weights in the network are progressively adjusted—and we see that eventually we get a network that successfully reproduces the function we want. So how do we adjust the weights? The basic idea is at each stage to see “how far away we are” from getting the function we want—and then to update the weights in such a way as to get closer.

To find out “how far away we are” we compute what’s usually called a “loss function” (or sometimes “cost function”). Here we’re using a simple (L2) loss function that’s just the sum of the squares of the differences between the values we get, and the true values. And what we see is that as our training process progresses, the loss function progressively decreases (following a certain “learning curve” that’s different for different tasks)—until we reach a point where the network (at least to a good approximation) successfully reproduces the function we want:

Alright, so the last essential piece to explain is how the weights are adjusted to reduce the loss function. As we’ve said, the loss function gives us a “distance” between the values we’ve got, and the true values. But the “values we’ve got” are determined at each stage by the current version of neural net—and by the weights in it. But now imagine that the weights are variables—say wi. We want to find out how to adjust the values of these variables to minimize the loss that depends on them.

For example, imagine (in an incredible simplification of typical neural nets used in practice) that we have just two weights w1 and w2. Then we might have a loss that as a function of w1 and w2 looks like this:

Numerical analysis provides a variety of techniques for finding the minimum in cases like this. But a typical approach is just to progressively follow the path of steepest descent from whatever previous w1, w2 we had:

Like water flowing down a mountain, all that’s guaranteed is that this procedure will end up at some local minimum of the surface (“a mountain lake”); it might well not reach the ultimate global minimum.

It’s not obvious that it would be feasible to find the path of the steepest descent on the “weight landscape”. But calculus comes to the rescue. As we mentioned above, one can always think of a neural net as computing a mathematical function—that depends on its inputs, and its weights. But now consider differentiating with respect to these weights. It turns out that the chain rule of calculus in effect lets us “unravel” the operations done by successive layers in the neural net. And the result is that we can—at least in some local approximation—“invert” the operation of the neural net, and progressively find weights that minimize the loss associated with the output.

The picture above shows the kind of minimization we might need to do in the unrealistically simple case of just 2 weights. But it turns out that even with many more weights (ChatGPT uses 175 billion) it’s still possible to do the minimization, at least to some level of approximation. And in fact the big breakthrough in “deep learning” that occurred around 2011 was associated with the discovery that in some sense it can be easier to do (at least approximate) minimization when there are lots of weights involved than when there are fairly few.

In other words—somewhat counterintuitively—it can be easier to solve more complicated problems with neural nets than simpler ones. And the rough reason for this seems to be that when one has a lot of “weight variables” one has a high-dimensional space with “lots of different directions” that can lead one to the minimum—whereas with fewer variables it’s easier to end up getting stuck in a local minimum (“mountain lake”) from which there’s no “direction to get out”.

It’s worth pointing out that in typical cases there are many different collections of weights that will all give neural nets that have pretty much the same performance. And usually in practical neural net training there are lots of random choices made—that lead to “different-but-equivalent solutions”, like these:

But each such “different solution” will have at least slightly different behavior. And if we ask, say, for an “extrapolation” outside the region where we gave training examples, we can get dramatically different results:

But which of these is “right”? There’s really no way to say. They’re all “consistent with the observed data”. But they all correspond to different “innate” ways to “think about” what to do “outside the box”. And some may seem “more reasonable” to us humans than others.

The Practice and Lore of Neural Net Training

Particularly over the past decade, there’ve been many advances in the art of training neural nets. And, yes, it is basically an art. Sometimes—especially in retrospect—one can see at least a glimmer of a “scientific explanation” for something that’s being done. But mostly things have been discovered by trial and error, adding ideas and tricks that have progressively built a significant lore about how to work with neural nets.

There are several key parts. First, there’s the matter of what architecture of neural net one should use for a particular task. Then there’s the critical issue of how one’s going to get the data on which to train the neural net. And increasingly one isn’t dealing with training a net from scratch: instead a new net can either directly incorporate another already-trained net, or at least can use that net to generate more training examples for itself.

One might have thought that for every particular kind of task one would need a different architecture of neural net. But what’s been found is that the same architecture often seems to work even for apparently quite different tasks. At some level this reminds one of the idea of universal computation (and my Principle of Computational Equivalence), but, as I’ll discuss later, I think it’s more a reflection of the fact that the tasks we’re typically trying to get neural nets to do are “human-like” ones—and neural nets can capture quite general “human-like processes”.

In earlier days of neural nets, there tended to be the idea that one should “make the neural net do as little as possible”. For example, in converting speech to text it was thought that one should first analyze the audio of the speech, break it into phonemes, etc. But what was found is that—at least for “human-like tasks”—it’s usually better just to try to train the neural net on the “end-to-end problem”, letting it “discover” the necessary intermediate features, encodings, etc. for itself.

There was also the idea that one should introduce complicated individual components into the neural net, to let it in effect “explicitly implement particular algorithmic ideas”. But once again, this has mostly turned out not to be worthwhile; instead, it’s better just to deal with very simple components and let them “organize themselves” (albeit usually in ways we can’t understand) to achieve (presumably) the equivalent of those algorithmic ideas.

That’s not to say that there are no “structuring ideas” that are relevant for neural nets. Thus, for example, having 2D arrays of neurons with local connections seems at least very useful in the early stages of processing images. And having patterns of connectivity that concentrate on “looking back in sequences” seems useful—as we’ll see later—in dealing with things like human language, for example in ChatGPT.

But an important feature of neural nets is that—like computers in general—they’re ultimately just dealing with data. And current neural nets—with current approaches to neural net training—specifically deal with arrays of numbers. But in the course of processing, those arrays can be completely rearranged and reshaped. And as an example, the network we used for identifying digits above starts with a 2D “image-like” array, quickly “thickening” to many channels, but then “concentrating down” into a 1D array that will ultimately contain elements representing the different possible output digits:

But, OK, how can one tell how big a neural net one will need for a particular task? It’s something of an art. At some level the key thing is to know “how hard the task is”. But for human-like tasks that’s typically very hard to estimate. Yes, there may be a systematic way to do the task very “mechanically” by computer. But it’s hard to know if there are what one might think of as tricks or shortcuts that allow one to do the task at least at a “human-like level” vastly more easily. It might take enumerating a giant game tree to “mechanically” play a certain game; but there might be a much easier (“heuristic”) way to achieve “human-level play”.

When one’s dealing with tiny neural nets and simple tasks one can sometimes explicitly see that one “can’t get there from here”. For example, here’s the best one seems to be able to do on the task from the previous section with a few small neural nets:

And what we see is that if the net is too small, it just can’t reproduce the function we want. But above some size, it has no problem—at least if one trains it for long enough, with enough examples. And, by the way, these pictures illustrate a piece of neural net lore: that one can often get away with a smaller network if there’s a “squeeze” in the middle that forces everything to go through a smaller intermediate number of neurons. (It’s also worth mentioning that “no-intermediate-layer”—or so-called “perceptron”—networks can only learn essentially linear functions—but as soon as there’s even one intermediate layer it’s always in principle possible to approximate any function arbitrarily well, at least if one has enough neurons, though to make it feasibly trainable one typically has some kind of regularization or normalization.)

OK, so let’s say one’s settled on a certain neural net architecture. Now there’s the issue of getting data to train the network with. And many of the practical challenges around neural nets—and machine learning in general—center on acquiring or preparing the necessary training data. In many cases (“supervised learning”) one wants to get explicit examples of inputs and the outputs one is expecting from them. Thus, for example, one might want images tagged by what’s in them, or some other attribute. And maybe one will have to explicitly go through—usually with great effort—and do the tagging. But very often it turns out to be possible to piggyback on something that’s already been done, or use it as some kind of proxy. And so, for example, one might use alt tags that have been provided for images on the web. Or, in a different domain, one might use closed captions that have been created for videos. Or—for language translation training—one might use parallel versions of webpages or other documents that exist in different languages.

How much data do you need to show a neural net to train it for a particular task? Again, it’s hard to estimate from first principles. Certainly the requirements can be dramatically reduced by using “transfer learning” to “transfer in” things like lists of important features that have already been learned in another network. But generally neural nets need to “see a lot of examples” to train well. And at least for some tasks it’s an important piece of neural net lore that the examples can be incredibly repetitive. And indeed it’s a standard strategy to just show a neural net all the examples one has, over and over again. In each of these “training rounds” (or “epochs”) the neural net will be in at least a slightly different state, and somehow “reminding it” of a particular example is useful in getting it to “remember that example”. (And, yes, perhaps this is analogous to the usefulness of repetition in human mem