AI

Why Scaling Prediction Cannot Create Intelligence - Alexander Mattick

Read full article at source ↗

Based on the episode description, not a full transcript. Source evidence retrieved 21/09/2026, 16:35:18 UTC.

Briefings

Alexander Mattick argues that scaling prediction models does not equate to creating intelligence, emphasizing that inference is computationally expensive and often unproductive. He outlines his research trajectory from Monte Carlo methods to advanced generative models like GFlowNets and flow matching, while noting that JEPA and world models are more about branding than technical utility. Mattick critiques existing deep learning theories for failing to predict practical outcomes, citing the high cost of information and the limitations of current control theory approaches.

Key points

• Inference in modern machine learning involves sampling from complex energy functions, which is rarely cost-effective despite theoretical possibility. Mattick identifies JEPA and world models as closer to branding than technical categories, while noting that deep learning theories currently fail to predict enough to guide practice effectively.
• Prediction is fundamentally distinct from control, and scaling prediction cannot create intelligence. Mattick highlights that while some theories predict enough for practice, others like JEPA are more about branding, and the Bitter Lesson demonstrates that expensive information often yields poor results.

Context and limitations

Coverage is limited to Mattick's research trajectory from 2022 to 2026, excluding preprint status and uncertainty regarding future developments.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 21/09/2026, 16:39:38 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeihxpxp2j3e33nhggtkijn6qa7krb37t6fsyzucfrlrfjukxnxa7x4