Alexander Mattick is a researcher at Fraunhofer IIS and a PhD researcher at the University of Technology Nuremberg (UTN), and a regular on Yannic Kilcher's Discord. He first came on MLST in 2022, after helping research the Yann LeCun and Randall Balestriero episode on interpolation.
SPONSOR:
---
Cyber Fund built the Monastery to help founders ship products that were impossible a year ago. Applications for Batch 1 are now open.
Apply now: https://cyber.fund
---
Alexander treats inference as the thread running through modern machine learning: once you have a model, what does it cost to get an answer out of it? He works through Monte Carlo, GFlowNets, energy-based models, diffusion, normalising flows and flow matching, with four short explainers he recorded himself. He is blunt about energy-based models: you can sample from them in principle, but it is rarely worth the compute. JEPA and "world model", he says, are closer to branding than to technical categories.
Next: theories of deep learning, none of which he thinks predicts enough yet to guide practice, then reinforcement learning.
---
Cold open: information is expensive
Welcome back, Alexander Mattic
Alexander's research background
Inference: densities, sampling and Monte Carlo
GFlowNets, energy functions and MCMC
Explainer: energy-based models
Why model a density at all?
From learned energies to flow matching
Explainers: diffusion and normalising flows
Are energy-based models generative?
JEPA, contrastive learning and collapse
Why non-language modalities need flows
Inference as search: branch and bound
Q-learning and delayed consequences
Flow matching, optimal transport, Fokker-Planck
Explainer: flow matching
AlphaFold, latents and scale versus architecture
Two families of deep learning theory
What a good theory would predict
The manifold hypothesis and compression
Is reward enough?
Control theory versus reinforcement learning
The Bitter Lesson and expensive information
Constrained RL: the constrained MDP toolbox
Creativity as constrained search
Reality is protean: when abstractions hold
What is a world model?
Prediction is not control
Robot demos, MPC and reliability
---
REFERENCES:
[] GFlowNets (Bengio et al., 2021)
https://arxiv.org/abs/2106.04399
[] Contrastive Self-Supervised Learning (Anand, 2020)
https://ankeshanand.com/blog/2020/01/26/contrative-self-supervised-learning.html
[] LeJEPA (Balestriero and LeCun, 2025)
https://arxiv.org/abs/2511.08544v3
[] RL for Node Selection in Branch-and-Bound (Mattick)
https://openreview.net/forum?id=0ez68a5UqI
[] Flow Matching for Generative Modeling
https://arxiv.org/abs/2210.02747v2
[] Disentangling feature and lazy training in deep neural networks
https://arxiv.org/abs/1906.08034v4
[] Reward is enough (Silver)
https://doi.org/10.1016/j.artint.2021.103535
[] Learning ReLU networks to high uniform accuracy is intractable (Berner et al.)
https://arxiv.org/abs/2205.13531v2
[] Dota 2 with Large Scale Deep RL
https://arxiv.org/abs/1912.06680v1
[] Constrained Update Projection for Safe Policy Optimization (Yang et al., 2022)
https://arxiv.org/abs/2209.07089
[] SafeMPO (ICLR 2026)
https://openreview.net/forum?id=1m0EU6QXj6
[] Why Creativity Cannot Be Interpolated
https://archive.mlst.ai/paper/why-creativity-cannot-be-interpolated/
[] Invalid Action Masking (Huang and Ontañón)
https://arxiv.org/abs/2006.14171
[] Probability Theory: The Logic of Science (Jaynes, 2003)
https://www.cambridge.org/core/books/probability-theory/9CA08E224FF30123304E6D8935CF1A99
[] Training Agents Inside of Scalable World Models (Hafner et al., 2025)
https://arxiv.org/abs/2509.24527v1
[] World Models (Ha and Schmidhuber, 2018)
