Blackboard Lecture: How Frontier LLMs Are Trained and Served
Author: @dwarkesh_sp + @reinerpope | Date: 2026-04-29
Dwarkesh Patel and Reiner Pope (Google) did a blackboard lecture format walking through how frontier LLMs are trained and served. Covers the full pipeline from data to deployment at scale.
Key Claims
- Full walkthrough of frontier LLM training pipeline (data → pretraining → post-training → serving)
- Blackboard lecture format — detailed technical exposition
- Reiner Pope (Google) brings insider perspective on production systems
- Covers both training and inference/serving at scale
Takeaways
- Rare public deep-dive into the full frontier training stack
- Training and serving are deeply intertwined (model architecture choices affect both)
- Google’s approach to large-scale serving provides useful reference architecture
- Educational content from actual practitioners > textbook descriptions