Active inference scales in silico only by borrowing deep-learning machinery it did not originate.

The claim

Active inference's reach in artificial agents comes from function approximators it did not originate. Mazzaglia and co-authors (2022) state that without adequate models active inference is "unfeasible to scale" to continuous or high-dimensional settings, and that the scalable implementations borrow deep-learning methods (VAEs, recurrent world models, ensembles, Monte Carlo tree search) from reinforcement and representation learning.