The signature of a path $X$ in $R^d$ is a noncommutative formal power series
$\sigma(X)$ that uniquely characterizes $X$ up to a mild equivalence. Signature
tensors have been introduced by Kuo-Tsai Chen in 1954 and, after decades
of oblivion, have recently risen again in the context of topological
data analysis. In this talk, based on a recent preprint written in
collaboration with F. Galuppi (MiMUW) and P. Santarsiero (University of
Bologna), I will show that a path $X$ is the solution of a given (system
of) ODE(s) if and only if its signature $\sigma(X)$ fulfills a certain system
of algebraic equations. As a "pretext" for introducing signature
tensors, I will briefly recall how ChatGPT predicts the next token for a
given prompt and show that the so-called "attention" mechanism is
nothing but a vector field on the space where signature tensors live.