Step-by-step guide to trace LLM behavior origins

gerardsans · x · 2026-08-20

Gerard Sans provided a step-by-step guide to identify where a specific LLM behavior originates, ranked by usual suspects: 1) Training data, 2) Pre-training, 3) Post-training (RLHF/RL), 4) Deployment (system prompt, UX/UI, API), 5) Inference (prompt, context, tools). He also warned against confusing lab narratives for funding with actual science.

Related event: A Five-Step Guide to Tracing LLM Behavior Origins(2 posts)→

Original post →

More from coding & agent

coding & agent channel →