Can a machine understand?

Section 1Understanding is a Smooth Property.

A second conundrum we find ourselves in concerns the concept of understanding.

Both on social media, where I’m often in discussions, in articles, and even in scientific papers, people have argued that AI models cannot understand things. The argument is that they are just pattern matching; they don’t actually “understand” what they’re talking about.

But what IS understanding?

To understand is to have a clear idea of how something works, how the pieces of something fit together. Put another way, understanding is what happens when you have built an accurate model of something in your brain.

If you understand English, that means you have a mental model of English that’s actually similar to English in real life. Understanding here, therefore, acts like compression. It allows you to take some complex phenomena, strip it of superfluous information and noise, and store a map of it in your brain.

But because understanding is based on models, understanding cannot be a binary situation. As the British statistician George Box once said, “All models are wrong, but some are useful”. Models vary in how faithfully they adapt the situation they model. Therefore, levels of understanding must vary.

Instead of arguing whether AIs can or cannot understand, I argue that the correct argument is about levels of understanding. Pattern matching, therefore, isn’t something you do instead of understanding; it is a form of understanding by itself.

AI clearly meets this bar. In many domains, we have seen AI models create clearer and more sophisticated “cognitive” models of the world over time. They’ve learned to speak various languages, and even translate between them better and better over time; they’ve learned to dream images, videos, and music, and every generation, they get better at it.

This IS better understanding. It’s an increasingly clear resolution replicating the fractal nature of the universe.

We even see this in ourselves as humans. Consider, for example, the entire history of science. Look at how our understanding of physical reality has improved drastically. We too started with pattern recognition. Whatever goes up must come down, life spontaneously arises, sad people get sick(or perhaps sick people get sad), and so on. And as we’ve found more, our understanding improved; we created Germ Theory, outlawed spontaneous generation, etc. We went from Aristotelian physics to Newtonian Physics to relativity to quantum mechanics, and every step is less about pattern matching and more about searching for causal connections.

Section 2Understanding is a Property of Learners

If I draw a picture of a goat on a piece of paper, the paper now, ostensibly, contains a model of the goat.

So does the piece of paper understand goats now?

What about a thermometer? It goes up when the temperature goes up and goes down when the opposite happens. This is, in essence, a temporal model of temperature, contained within a tube with mercury in it, or, if you will, some electronics.

So does a thermometer understand temperature?

No.

Consider a car. Think about a race car on the track; it’s moving. It’s fast.

Consider the same car, parked. At that point, it’s not fast. But it is still the type of thing that can be fast. Its speed at that point is zero, but that zero is a real number on a scale on which that car can improve.

But consider a riddle. Is a riddle fast? Can it be fast or slow? It makes no sense. Riddles don’t have the property of speed, as they don’t even move. The question has no scale.

Because Understanding exists on a scale, it should be a property of things that have that scale. It should be a property of things that have the ability to start at a low level of understanding and improve over time, or perhaps even degrade over time.

These are learners.

A learner is a system that can start with no model and, exposed to data and compute, whether in digital form or biologically, or some other yet uninvented process, build a model of that data that generalizes to held-out samples.

I’m reminded of the story by the philosopher Wittgenstein, where he describes a student who watches a teacher write a number series. Then the student says, “Now I can go on!”

This is precisely what understanding is. By absorbing a small amount of data, the student extracts a rule, a model, that describes the entire dataset to great accuracy.

This immediately gives rise to the concept of narrow learners and general learners. Your YouTube algorithm is a narrow learner. When you create an account, it doesn’t know anything about you; it understands nothing. But then you start watching videos and liking and commenting and scrolling through your homepage, and it... learns about you. It begins to understand your preferences and actually gets good at finding you the kinds of videos you like to watch.

There are likely over 20 billion YouTube videos on YouTube’s servers by now, yet it finds you the videos you are most likely to watch based on your past preferences.

Narrow learners acquire narrow understanding. The YouTube algorithm understands how to serve videos to customers, but it probably doesn’t understand how to solve a quadratic equation.

The AIs humanity is building are general learners, or at least, that’s what we want them to be. We want to be able to expose them to all sorts of data and compute, and have them... build mental models of how the world works over time. This leads to general understanding.

Section 3Understanding is a State in Learners.

Suppose you were a genius ML researcher, and you sat at your desk and typed out the weights to a small model that can identify sparrows with 99.9% accuracy. Or suppose you crafted features that could do this. Does that model understand?

Sure, it has a model of reality inside itself that it would have had if it had been exposed to data and made to form its own features and adapt its weights under some optimization pressure, but instead it appeared fully formed.

This is actually just last Thursdayism for AI systems. It would be no different from if God had, last Thursday, willed the world into existence from scratch, including us and our memories and our understanding.

An LLM and a calculator can both do simple arithmetic. But while the LLM learned how to do this, the calculator did not. The circuits it uses were merely created whole from scratch. It cannot get any better or worse. So even though technically it has the same “model” of arithmetic(arguably better) than an LLM does, it doesn’t understand.

While it’s possible to instantiate a learner at any point along its learning journey, that does not mean it doesn’t understand, because even that instantiation replicates a causal history. If God had actually created the world last Thursday, there would be no straightforward way to argue that today is only a few weeks from the beginning history; you would be able to argue that the entirety of causal history was created instantaneously.

A sufficiently advanced technology could clone you perfectly, alongside every connection in your brain, but that doesn’t mean your clone doesn’t understand, because your clone, like you, is a learner. The understanding that your clone has is simply the current state of his/her learning circuitry, which was formed from your history.

For AIs, this is important. An AI is trained, often once, and then copied billions of times across multiple instances. Each instance carries the causal history of understanding. Every copy arrived there by being an instance of a model that actually learned in training how to do all those things; that’s why it makes sense to say that even instances of AI understand things.

The reason why this works for learners but not static circuits like, say, calculators, is that learners are dynamic. The state of a learner changes with learning. That is the kind of thing that it is. It adapts its internal structure to data with compute.

Section 4Yes, AIs Model the World.

You may be tempted to argue that actually LLMs don’t even build models of the world; they just predict the next token by looking at the most statistically likely token to come next.

First of all, prediction is stronger than you think. It’s entirely possible that prediction itself is the core of intelligence. A good controller has to model the world.

Consider science, where we judge the effectiveness of theories by judging them on their ability to predict. Since understanding is about model compression, prediction tells us about the quality of the compression. Good compression keeps enough in the model that allows it to map out parts of the data it hasn’t even seen yet.

The predictive capacity that models have is really quite general. It’s not just software. They are also able to control robots in the real world and even learn interesting things by engaging in self-play, with zero human data.

Lastly, we have empirical evidence to the contrary. While it’s correct that the base mechanism by which AIs learn is next-token prediction, we also know that in order to achieve this, they actually build models of what they’re learning.

For example, in 2023, the Othello-GPT experiment showed that, trained on a list of sequences of legal moves, without showing the board, models nevertheless learned a representation of the board itself. The goal was merely next move prediction, but a model of the board emerged anyway.

Gurnee and Tegmark(2023) showed Llama-2 models and gave them location names, as well as the names of events and people, and found that the model activations could predict the latitude and longitude of each place, and the date of each event by using the same representation.

In 2024, Minyoung Huh et al. showed that even when you train models with different objectives, data, and even modalities(like text vs. vision), they tend to converge to a shared model of reality in their representation.

In 2026, Anthropic showed a completely emergent phenomenon in the representation space of large language models, called the J-space, where they formed representations that did not directly go into the next tokens.

We can argue that the models that AIs make are poor or shallow, but as we’ve discussed earlier, this is merely a level of understanding. Models can continue to refine and update their internal representations of the world as architectures, data, and compute evolve.