Pretraining Builds Representations, Posttraining Induces Metapolicy
Liu_eroteme · x · 2026-08-27
The post discusses the distinction between pretraining and posttraining, suggesting that pretraining primarily constructs observation and policy representations, while posttraining induces a 'metapolicy' for representation and policy selection. This framing aligns with observations in reinforcement learning.
Related event: Pretraining as Transformation, Post-training as Translation(2 posts)→
More from Research
- Alex Rives, pioneer of protein language model ESM, named to TIME100 AI — proteinrosh · 2026-08-28
- New Paper: Dynamic Multi-Byte Prediction Speeds Up Hierarchical Byte-Level LMs — madeofAjala · 2026-08-28
- Jeff Dean's DiscoveryLoop Extends AI-Driven Iteration to Full Scientific Experiment Loops — agihouse_org · 2026-08-28
- Google's Pi team publishes research on multi-agent cooperation and emergent intelligence — blaiseaguera · 2026-08-27
- Google VP Argüera y Arcas: consciousness is inherently social, and that may unlock the AI question — blaiseaguera · 2026-08-27
- New Benchmark Shows AI Agents Miss a Quarter of the Live Web — EXM7777 · 2026-08-27