Why Scaling Prediction Cannot Create Intelligence - Alexander Mattick

<p>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&#39;s Discord. He first came on MLST in 2022, after helping research the Yann LeCun and Randall Balestriero episode on interpolation.</p><p><br></p><p>SPONSOR:</p><p>---</p><p>Cyber Fund built the Monastery to help founders ship products that were impossible a year ago. Applications for Batch 1 are now open.</p><p>Apply now: https://cyber.fund</p><p>---</p><p><br></p><p>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 &quot;world model&quot;, he says, are closer to branding than to technical categories.</p><p><br></p><p>Next: theories of deep learning, none of which he thinks predicts enough yet to guide practice, then reinforcement learning. </p><p><br></p><p>---</p><p>0:00 Cold open: information is expensive</p><p>0:51 Welcome back, Alexander Mattic</p><p>2:08 Alexander&#39;s research background</p><p>2:50 Inference: densities, sampling and Monte Carlo</p><p>6:42 GFlowNets, energy functions and MCMC</p><p>9:45 Explainer: energy-based models</p><p>11:03 Why model a density at all?</p><p>17:30 From learned energies to flow matching</p><p>25:08 Explainers: diffusion and normalising flows</p><p>28:33 Are energy-based models generative?</p><p>33:22 JEPA, contrastive learning and collapse</p><p>41:13 Why non-language modalities need flows</p><p>44:51 Inference as search: branch and bound</p><p>49:43 Q-learning and delayed consequences</p><p>55:14 Flow matching, optimal transport, Fokker-Planck</p><p>1:00:03 Explainer: flow matching</p><p>1:01:49 AlphaFold, latents and scale versus architecture</p><p>1:07:52 Two families of deep learning theory</p><p>1:15:04 What a good theory would predict</p><p>1:23:53 The manifold hypothesis and compression</p><p>1:28:25 Is reward enough?</p><p>1:32:01 Control theory versus reinforcement learning</p><p>1:37:22 The Bitter Lesson and expensive information</p><p>1:42:08 Constrained RL: the constrained MDP toolbox</p><p>1:50:12 Creativity as constrained search</p><p>1:55:44 Reality is protean: when abstractions hold</p><p>2:00:32 What is a world model?</p><p>2:04:38 Prediction is not control</p><p>2:08:13 Robot demos, MPC and reliability</p><p><br></p><p>---</p><p>REFERENCES:</p><p>[6:55] GFlowNets (Bengio et al., 2021)</p><p>https://arxiv.org/abs/2106.04399</p><p>[38:46] Contrastive Self-Supervised Learning (Anand, 2020)</p><p>https://ankeshanand.com/blog/2020/01/26/contrative-self-supervised-learning.html</p><p>[38:56] LeJEPA (Balestriero and LeCun, 2025)</p><p>https://arxiv.org/abs/2511.08544v3</p><p>[47:10] RL for Node Selection in Branch-and-Bound (Mattick)</p><p>https://openreview.net/forum?id=0ez68a5UqI</p><p>[56:20] Flow Matching for Generative Modeling </p><p>https://arxiv.org/abs/2210.02747v2</p><p>[1:12:41] Disentangling feature and lazy training in deep neural networks</p><p>https://arxiv.org/abs/1906.08034v4</p><p>[1:31:05] Reward is enough (Silver)</p><p>https://doi.org/10.1016/j.artint.2021.103535</p><p>[1:35:12] Learning ReLU networks to high uniform accuracy is intractable (Berner et al.)</p><p>https://arxiv.org/abs/2205.13531v2</p><p>[1:40:20] Dota 2 with Large Scale Deep RL </p><p>https://arxiv.org/abs/1912.06680v1</p><p>[1:45:41] Constrained Update Projection for Safe Policy Optimization (Yang et al., 2022)</p><p>https://arxiv.org/abs/2209.07089</p><p>[1:46:11] SafeMPO (ICLR 2026)</p><p>https://openreview.net/forum?id=1m0EU6QXj6</p><p>[1:50:17] Why Creativity Cannot Be Interpolated</p><p>https://archive.mlst.ai/paper/why-creativity-cannot-be-interpolated/</p><p>[1:51:39] Invalid Action Masking (Huang and Ontañón)</p><p>https://arxiv.org/abs/2006.14171</p><p>[2:00:04] Probability Theory: The Logic of Science (Jaynes, 2003)</p><p>https://www.cambridge.org/core/books/probability-theory/9CA08E224FF30123304E6D8935CF1A99</p><p>[2:01:53] Training Agents Inside of Scalable World Models (Hafner et al., 2025)</p><p>https://arxiv.org/abs/2509.24527v1</p><p>[2:03:43] World Models (Ha and Schmidhuber, 2018)</p><p>https://arxiv.org/abs/1803.10122v4</p>