MLST: Why scaling prediction cannot create intelligence — 2-hour deep dive with Alexander Mattick
Machine Learning Street Talk · rss · 2026-09-21
Machine Learning Street Talk hosts Fraunhofer IIS researcher Alexander Mattick for a 2-hour discussion arguing inference is the thread running through modern ML. He covers Monte Carlo, GFlowNets, energy-based models (rarely worth the compute), diffusion, normalizing flows and flow matching; calls JEPA and 'world model' closer to branding than technical categories; argues current deep learning theory predicts too little to guide practice; and explores RL, constrained MDPs, control theory vs RL, the Bitter Lesson, and why prediction is not control. Full timeline and references included.
More from Research
- Is synthetic-cell SpudCell the enabling tech for mirror life? Researchers weigh in — NikoMcCarty · 2026-09-22
- Aethos X hits 94% on EnterpriseRAG-Bench at 42k files, nearing effective ceiling — rohanpaul_ai · 2026-09-22
- Meta's LION uses sparse key-value memory to fix preference drift in generative recsys — armand_ruiz · 2026-09-22
- AMD Engineers Publish GEMM Optimization Tutorial Blog Inspired by the GEMM Ladder — simran_s_arora · 2026-09-22
- Stop Trusting MMLU-Pro: Every's Evals Lead Says Build Your Own AI Benchmark — danshipper · 2026-09-22
- DL models now learn from less data; new ensemble framework boosts pharma ELN prediction robustness — marwinsegler · 2026-09-22