Mixture Languages
summary
Claim of a general-purpose continuous-process programming language with a native notion of parallel composition: states are topological manifolds, semantics use time-dependent vector-field semantics integrated by ODEs, and loss-based gradients produce dynamics under the Fisher metric. Relative entropy / KL acts as the loss, giving multiplicative vs additive path interpolations. Mixtures of observe commands generate probabilistic dependency graphs that subsume Bayes nets and factor graphs, while sequences of draw yield standard probabilistic programs, increasing expressivity and compositionality.
AI-written summary · about