MLST: Information is expensive — Alexander Mattick on inference, constrained RL and why world models are branding

Machine Learning Street Talk · youtube · 2026-09-21

Machine Learning Street Talk hosts Alexander Mattick (Fraunhofer IIS, UTN PhD) for a two-hour deep dive on the cost of inference.

Key points:

Inference & sampling: Monte Carlo, GFlowNets, energy-based models, diffusion, normalising flows and flow matching, with four self-recorded explainers; he's blunt that energy-based models are rarely worth the compute.
Critique of "world models": JEPA and world model are closer to branding than technical categories.
Deep learning theory: two families of theory, none yet predictive enough to guide practice.
Against "Reward is Enough" and the Bitter Lesson: real-world information is expensive, so constraints are the cheapest way to inject what experts already know; creativity as constrained search.
Control theory: prediction is not control, robot demos say little about failure rates; MPC and constrained MDP toolboxes.

Best for listeners wanting the mathematical foundations of generative models and the limits of RL.

More from AGI Musings