Blog

Notes from the field.

In-depth writing on LLM training, agentic systems, retrieval, and the engineering trade-offs that decide whether AI products actually work in production.

May 3, 202613 min read

60% of your dataset is doing the work — a GRPO reward-variance analysis

Why most curated RL datasets carry no learning signal on a third or more of their prompts — and how a 30-minute pre-flight check cuts GRPO compute without losing uplift.

GRPORLPost-Training