The Shape of Intelligence

In 2024, a team at MIT borrowed a line from Tolstoy to describe something strange about modern AI: all strong models are alike; each weak model is weak in its own way.

The models in question were built by different labs, with different architectures, trained on different data. Some had only ever seen images, others only text. Nobody coordinated them. And yet, as they get larger and more capable, the way they organize what they know looks more and more alike.

Shadows on the Wall

The researchers named their idea after Plato's cave, where the people inside see only shadows on a wall and never the objects casting them. Every learner is in some version of that position. A photograph is a shadow of the world, a caption is another, a paragraph of text is another; for that matter, so is everything that reaches us through our eyes and ears. Each model sees only shadows, and strong models keep reconstructing the same object behind them.

Ilya Sutskever made a version of this point in 2023, the day after GPT-4 was released: text is a projection of the world, and a model that predicts text well enough has to learn something about the world that produced it.

Why the Shape Has to Exist

Part of this should not be surprising. The basic premise of machine learning is that a model learns to reproduce the distribution that generated its data. Train on photographs and the model learns what the world tends to look like. Train on human writing and it learns what the world is like, and also how people think their way through it: the explanations, arguments, proofs, and mistakes that fill the written record.

So the shape a model learns was there before the model. The real world, together with the reasoning people have recorded about it, is the shape. In a sense its existence was a requirement for machine learning to work at all; if there were no stable structure behind the data, there would be nothing for training to find.

Seen this way, training is closer to discovery than invention. Different labs find the same structure for the same reason different surveyors draw the same coastline.

What Shape Means Here

Shape is meant literally. Inside a model, every word, image, or idea is represented as a list of numbers, called a vector. One number places a point on a line. Two numbers place it on a map. Three place it in a room. A modern model uses thousands of numbers for each concept, which places it in a space with thousands of dimensions.

Nobody can picture a space like that, but the mathematics works exactly the same way it does on paper: there are still distances, directions, and neighborhoods. Related ideas sit close together, and the relationships between them become directions you can travel. Take all of those points together and they form a structure with real geometry, and that structure is the shape.

Evidence You Can Picture

Language models trained only on text, models that have, in Sutskever's words, never seen a single photon, arrange colors internally much the way human vision does: red sits near orange, blue near purple. The information leaked in through an enormous number of sentences describing the world.

Researchers at Goodfire have found that a model stores the months of the year on a circle, with December next to January. Asked what month comes six months after August, the model routes the question through a general-purpose addition module, the same one it uses for other arithmetic. Circular structures like this have been found in other models as well; whether they all share the same machinery underneath is still an open question.

In a 2026 study, researchers took the internal state of one language model and passed it, through a simple learned mapping, into a different model's output layer. In many cases the second model produced coherent sentences from the first model's internal state. It worked best when the stronger model did the thinking and the weaker model did the speaking.

That last detail matters. A smaller model holds the same shape at lower resolution: it can read a sharp copy, but it cannot send one. Sutskever would call this compression. A large model compresses the world with less loss, and a small model approximates the same shape more lossily.

What It Means

If intelligence has a shape, then human and artificial intelligence look less like different kinds of thing and more like different approximations of the same thing. The brain arrived at its version through evolution and a lifetime of experience; a model arrives at its version through text and training. Researchers at MIT have even found that the better a language model is at predicting the next word, the better its internal activity predicts human brain activity during reading.

Understanding the shape also pays off twice. The better we understand the full version, the better our small, inexpensive approximations can be. And the better we can choose the variations we want (what a model emphasizes, how it behaves, what kind of character it has) rather than accepting whatever happens to fall out of training.

In our last post we argued that fire became safer because we studied it. The same holds here. The field that studies this shape directly, interpretability, is how we learn what a model actually knows and what it is trying to do.

Early Work

All of this is early. The convergence is real but partial: models are alike rather than identical, and much of what has been mapped so far sits near the surface of these systems rather than deep inside their reasoning. The researchers are careful to say so.

The direction is what makes it remarkable. Something as abstract as intelligence, and the reasoning that comes with it, appears to have a shape, and every capable mind we have built, along with the ones we were born with, is learning to approximate it.

Mathematicians have long described the feeling of finding a proof as discovering something that was already true, waiting to be found. It turns out the same may hold for thinking itself. We did not invent the shape of intelligence. We have been rendering it, in neurons and now in silicon, one lossy copy at a time.

Further Reading

Huh, Cheung, Wang, and Isola (2024). The Platonic Representation Hypothesis. ICML 2024. The convergence argument, the cave, and the color experiment.

Gorbett and Jana (2026). Ventriloquist LLMs: Linear Alignment of Late-Stage Representations. Decoding one model's internal state through another model's output layer.

Schrimpf et al. (2021). The Neural Architecture of Language: Integrative Modeling Converges on Predictive Processing. PNAS. Language models compared against human brain recordings.

Geiger et al. (2026). The World Inside Neural Networks. Goodfire Research. Curved geometry inside models, and where it comes from: the structure of the world reflected in training data.

Feucht, Haklay, et al. (2026). Arithmetic in the Wild: Llama Uses Base-10 Addition to Reason About Cyclic Concepts. The months circle and the shared addition module.

Machine Learning Street Talk (2026). Strange Geometric Shapes Found Inside AIs, with Tom McGrath. An interview with Goodfire's Tom McGrath on neural geometry, interpretability, and intentional design.

Ilya Sutskever and Jensen Huang (2023). Fireside Chat with Ilya Sutskever and Jensen Huang: AI Today and Vision of the Future. Fireside chat at NVIDIA GTC, recorded the day after GPT-4's release. Compression, prediction, and text as a projection of the world.