OpenAI's Sherwin Wu on Building for 800M Users, Model Proliferation, and Why Determinism Beats Pure Agents
OpenAI's Sherwin Wu, who leads engineering for the developer platform (API), sits down with a16z's Martine Cassato to discuss how OpenAI operates at massive scale—now reaching roughly 800 million weekly ChatGPT users, or about 10% of the globe. Wu explains why OpenAI treats ChatGPT as a first-party app and the API as a horizontal platform simultaneously, framing the tension as manageable because Sam and Greg have been principled about it since day one, and because "growth solves a lot." A recurring theme is that models behave as anti-disintermediation technology: users know and care which model they're using (GPT-5, Composer, Cursor's tab model), making traditional software abstraction layers surprisingly brittle. The conversation traces how the industry's mental model has shifted—from expecting one model to rule them all toward a proliferation of specialized models like Codex, GPT-OSS, and vertical-tuned variants. Wu highlights reinforcement fine-tuning (RFT) as the recent unlock that lets customers actually leverage their proprietary data, citing examples like medical insurance coding and agentic planning. OpenAI is piloting data-sharing arrangements where customers who contribute data get discounted inference or free training. Pricing has stayed usage-based (a "one-way ratchet" per Ben Cott's framing) because it correlates heavily with outcomes at the test-time-compute level. Wu also defends OpenAI's open-weight strategy (GPT-OSS) as low cannibalization risk and ecosystem-positive, and dives into why the Agent Builder uses deterministic, node-based workflows: much real-world work is SOP-driven (customer support, sales, regulated industries), not the undirected, exploratory work software engineers do. He draws a parallel to NPCs in games, where the response catalog and game logic live in code while the model handles language understanding—exactly the constrained pattern OpenAI's agent tooling is designed for.
Original sources
Podcast The a16z Show
Episode:How OpenAI Builds for 800 Million Weekly Users: Model Specialization and Fine-Tuning ↗
YouTube The a16z Show
Video:How OpenAI Builds for 800 Million Weekly Users: Model Specialization and Fine-Tuning ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.