How GPT, Claude, and Gemini are actually trained and served – Reiner Pope

Dwarkesh Patel · 2026-04-29

How GPT, Claude, and Gemini are actually trained and served – Reiner PopeDwarkesh Patel ▶ Play the video

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

Why it earned a slot

Reiner Pope’s blackboard deep-dive is a rare look at the full LLM stack—from chip-level efficiency to the economics of API pricing. He walks through a handful of equations and shows how you can back-out training batch sizes and token counts just from public price tables, which is eye-opening. The chalk-talk feels dense, but the flashcards he links at 9:12 make the material stick. It’s one of the few places you actually see the bridge between TPU architecture and model serving laid out in plain sight.

The short version

LLM Secrets Unveiled on a Chalkboard

Why it matters

Knowing the hidden cost and compute dynamics of frontier models lets practitioners design more efficient pipelines and negotiate better service contracts. Reiner’s blend of hardware insight and pricing math demystifies why scaling decisions matter beyond just model size. This knowledge is directly applicable when you’re budgeting for production-grade agents or fine-tuning large models.

My take

The way he extracts training regimes from API pricing feels like reverse-engineering a black box—exactly the mindset we need when labs keep their data under wraps.

How it connects

• Hardware-aware model design ↔ cost-effective scaling strategies
• Transparent pricing analysis ↔ better vendor negotiations for low-latency workloads

Bottom line

Study Reiner’s equations and flashcards, then audit your own model cost model to spot hidden inefficiencies.