Turbulent water flowing through a pipe and the swing of a pendulum may seem worlds apart. But they share something fundamental: both follow their own laws of motion, which can be described by numbers. A neural network trained on just one such 'dance of numbers' suddenly starts to understand another — without a single extra example. It's like a musician who learns one melody and then sight-reads an unfamiliar piece by catching its harmony.
In the new study, researchers used a small transformer model — the same type that underpins large language models. They trained it to predict one dynamical system and then tested it on another, completely different one. The model didn't just guess; it used a clever trick. First, it 'unfolded' a short sequence of observations into a high-dimensional space, like turning a flat ribbon into a three-dimensional spiral. This revealed the system's hidden geometry. Then it found 'islands of stability' in that space — regions the system tends to return to, like a ball rolling into a dip. These islands helped make accurate short-term predictions.
This finding explains why large pre-trained models handle tasks they never saw during training so well. It turns out the attention mechanism alone can extract global structure from local data. For physicists and engineers, this means a network trained once could become a universal predictor for a whole class of systems — from turbulence in an engine to climate oscillations. And the starting point was that same 'dance of numbers' where it all began.
A transformer trained on just one dynamical system was able to predict another, even a turbulent and chaotic one, without any fine-tuning. To do so, it built its own 'map' of the system in high-dimensional space, like a cartographer drawing terrain from sparse points.
Such transferability of knowledge evokes the 'intuition' of artificial intelligence in Ted Chiang's novella 'The Lifecycle of Software Objects': trained on some examples, it begins to grasp patterns no one asked it about.