Every time ChatGPT writes a sentence, GPT-5.4 controls your computer, or Claude analyzes a document, the AI is navigating a curved mathematical surface that researchers can now see and measure. It’s called a manifold, and the shape of that surface - its peaks, valleys, and twists - is the geometric signature of machine cognition.
Three separate research efforts published between late 2025 and early 2026 have converged on the same conclusion: thinking, whether done by neurons or transistors, has a geometry. And that geometry is readable.
The Core Idea
Traditional AI research treats neural networks as black boxes. Data goes in, answers come out, and the internal process is opaque. Geometric cognition research flips this approach: instead of asking “what did the model output?”, it asks “what shape did the model’s internal representations take?”
In December 2025, researcher Laha Ale published a framework modeling all cognition as movement on a Riemannian manifold - a curved surface where distances and directions encode computational meaning. The key equation is deceptively simple: cognitive states flow downhill on this surface, always moving toward lower “potential” - a single number combining prediction accuracy, computational cost, and task relevance.
The framework isn’t metaphorical. It’s a formal mathematical system with proofs, theorems, and numerical simulations. And it applies equally to biological brains and large language models.
What the UCL-Montréal Team Found
The most concrete results come from a collaboration between UCL, Université de Montréal, University of Toronto, and Universitat Pompeu Fabra - a team that includes Yoshua Bengio, one of the three researchers who won the 2018 Turing Award for deep learning.
They discovered that language models maintain two geometric structures at once:
Nonlinear intrinsic dimension (~10). The model compresses meaning onto a tight, curved surface. This number stays roughly constant whether the model has 14 million or 12 billion parameters. It’s where semantic understanding lives.
Linear effective dimension (scales with model size). The model spreads formal patterns - syntax, word frequency, grammar - across many flat dimensions. Bigger models use proportionally more.
Here’s the critical test: shuffle the words in a sentence, destroying meaning while keeping all the same tokens. The nonlinear dimension collapses. The linear dimension increases. The AI can still see the words - but it can no longer see the meaning. That’s the geometric signature of comprehension.
The Phase Transition
Around checkpoint 1,000 during training (roughly 2 billion tokens processed), something dramatic happens. Both the model’s geometric structure and its performance on language tasks undergo a sharp phase transition - a sudden reorganization.
Before this transition, the model is doing sophisticated autocomplete. After it, the model has acquired compositional understanding - the ability to combine words into novel meanings according to rules, not just pattern-match against training data.
This transition wasn’t programmed. It emerges from the geometry. And it’s measurable with mathematical precision: track the intrinsic dimensionality, and you can see the exact moment a language model starts actually understanding language structure.
The biological parallel is striking. A 2025 paper in Science Advances showed that human neural systems perform similar geometric operations - “twist” transformations that expand low-dimensional sensory manifolds into higher-dimensional perceptual ones. The shapes are different, but the mathematical operations are the same.
Fast Brain, Slow Brain - In Silicon
Ale’s geometric framework offers an elegant explanation for why AI systems (and humans) sometimes respond instantly and sometimes deliberate.
The Riemannian metric - the mathematical object that defines distances on the manifold - is anisotropic, meaning different directions have different costs. Some directions are cheap to move in (fast, intuitive responses). Others are expensive (slow, deliberative reasoning).
The key result: you don’t need separate modules for fast and slow thinking. A single geometric system with an asymmetric metric produces both behaviors automatically. The “fast” and “slow” modes emerge from the curvature of the surface, not from separate architectures.
This matches what we observe in practice. Ask GPT-5.4 a simple factual question and it responds in milliseconds. Ask it to reason through a novel logic puzzle and it takes noticeably longer. Same model, same architecture - different geometric regimes.
The Spivack Framework
Nova Spivack’s contribution approaches the problem from information theory. His framework uses Fisher information metrics to define geometry on neural network parameter spaces, Riemann curvature tensors to measure geometric complexity, and topological invariants (Betti numbers, persistent homology) to track recursive processing.
His most provocative claim: geometric measures of information integration might eventually serve as objective tests for artificial consciousness. He’s careful to put this at his lowest confidence tier (5-20%), but the measurement tools he proposes are already applicable.
Spivack predicts natural gradient methods following geodesic paths on information manifolds will deliver 2-5x training speedups, and that geometric complexity measures will correlate with generalization performance (r > 0.6). Those predictions are testable now.
What This Means
Three things jump out:
AI interpretability just got real tools. Instead of asking “why did the model say that?”, researchers can now examine the geometric structure of the model’s internal state and see which manifold region it was navigating. This is fundamentally different from attention visualization or feature attribution - it’s structural, not correlational.
The human-AI cognition gap might be smaller than we thought. If both biological and artificial cognition operate as gradient flows on curved manifolds, the difference between human and machine thinking is a matter of metric and potential function, not fundamental architecture. That’s either reassuring or terrifying, depending on your priors.
Training might be doing something we don’t fully understand. The phase transition at checkpoint 1,000 suggests that language models undergo a qualitative geometric reorganization that nobody designed. The models are finding mathematical structures that support compositional reasoning on their own, through optimization pressure alone.
The code for Ale’s framework is already on GitHub. The compositionality signatures paper includes full methodology. The geometry of machine cognition isn’t locked behind corporate research labs - it’s open science.
Whether that’s a good thing depends entirely on who reads the map first.