Are Language Models More Structured Than We Think?

For years, the dominant narrative has been clear:

Language models are powerful pattern matchers. Nothing more.

They don’t reason. They don’t calculate. They don’t build internal variables.

But evidence like this complicates that narrative.

We’re seeing:

• Low-dimensional manifolds encoding structured quantities.
• Attention heads performing geometric transformations.
• Internal organisation that behaves computationally.

Not symbolic AI. Not classical rule engines. But not shallow pattern imitation either.

Something in between.

What’s especially interesting is that this structure wasn’t programmed explicitly. It emerged from scale and training objectives.

Which raises bigger questions:

If character count lives on a geometric manifold, what else does? Planning horizons? Confidence estimates? Long-term constraints? Self-referential signals?

We may not fully understand how these systems work.

But one thing is becoming clearer:

The inside of a language model is not chaos. It’s geometry. And geometry scales.