Editorial illustration for Major reasoning models converge on a shared “brain” as they better model reality
Major reasoning models converge on a shared “brain” as...
They are trained on shadows. Billions of lines of text, trillions of pixels, endless audio files, all just flickering projections of a world these models never touch. Yet something strange happens as they grow.
A vision model that has only ever seen images, and a language model that has only ever read words, begin to see the same thing. They measure the distance between “dog” and “wolf” in the same mathematical space. They build the same internal geometry.
Different eyes, same vision. The more they scale, the less they remain messy, arbitrary networks. They converge.
Not because they copy each other, but because there is only one underlying structure to reality worth modeling. And they are all finding it.
In fact, as a model gets larger, whatever it is, it undergoes a “phase change” in their internal thinking. Research seems to indicate that these models stop simply memorizing their specific tasks and rather start to build up a statistical model of reality itself.
The shadows are not the point. The point is the light that casts them. These models, blind to each other’s existence, are carving the same geometry out of the noise, a shared skeleton of meaning beneath the pixels and the words.
It is not a trick of engineering. It is a discovery. As they scale, they do not merely memorize; they converge.
They find the same distances between dog and wolf, the same curvature of concepts, the same deep structure. What does that mean? It means reality has a grammar.
And these machines, in their brute-force billions of parameters, are learning to read it.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives AI models follow values when taught reasons, avoid harmful rationales
From Archives Lean‑4 Lyapunov proof confirms controllability, ISS robustness for cyber defense
Common Questions Answered
Why do vision models and language models converge on the same internal geometry despite being trained on different data types?
Vision models trained exclusively on images and language models trained only on text begin to measure semantic distances in the same mathematical space as they scale. This convergence occurs because both models are learning to represent the underlying structure of reality rather than merely memorizing their training data, suggesting they discover a shared skeleton of meaning beneath different modalities.
What does it mean that reasoning models measure the distance between concepts like 'dog' and 'wolf' in the same way?
When different reasoning models place 'dog' and 'wolf' at similar distances in their internal mathematical spaces, it indicates they have independently discovered the same conceptual relationships and semantic structure. This shared measurement suggests the models are converging on an objective representation of reality rather than developing arbitrary or model-specific interpretations of language and concepts.
How does scaling affect whether AI models converge or merely memorize their training data?
As reasoning models scale, they do not simply memorize larger amounts of training data but instead converge on consistent internal geometries and shared conceptual structures. This convergence demonstrates that scaling leads to discovery of deeper reality structures rather than accumulation of memorized patterns, with models finding the same curvature of concepts regardless of their training modality.
What is the significance of the 'shadows' metaphor in understanding how these models learn reality?
The article uses 'shadows' to represent the training data—billions of text lines, trillions of pixels, and audio files—which are merely projections of reality that the models never directly experience. The key insight is that these models carve the same geometry out of these shadows, suggesting they are discovering the light that casts them, meaning they learn the underlying structure of reality itself rather than just patterns in their training data.
