Key ideas
Scroll to read ↓Jason Liu explains how Pydantic models, OpenAI function calling, Instructor, and explicit validation turn language-model outputs into typed objects that existing software can inspect, execute, and maintain.
← All speakers
▶ Watch ↗
▶ Watch ↗
▶ Watch ↗
Related work and projects
- blog ↗github.com · Source ↗
- systematically-improving-rag ↗github.com · Source ↗
- Introductions - Jason Liu ↗
What's up? ¶ Start Here ¶ What This Site Is ¶ Investments ¶ Old Work ¶ Training ¶
jxnl.co
Bio, Work & Ideas
Jason Liu
Conference affiliation: Developer Experience, OpenAI · OpenAI · 2026
Jason Liu is a developer experience engineer on OpenAI’s Codex team and the creator of Instructor, the open-source library that helped establish structured outputs as a practical foundation for language-model applications. His work gives developers concrete ways to make AI systems predictable: typed schemas, validated responses, actionable evaluation, and agent workflows that preserve context without surrendering human oversight.
Liu studied computational mathematics and statistics at the University of Waterloo, worked as a data scientist at Meta, and became a staff machine-learning engineer at Stitch Fix. There, he built Flight, a framework for moving recommendation logic into observable, maintainable production systems. He subsequently founded 567 Studios and developed a consulting and education practice focused on retrieval, evaluation, and AI products; his teaching continues through Systematically Improving RAG.
Instructor distilled that experience into a lightweight interface built around Pydantic models, Python type hints, validation, and retries. Developers describe the object they need; the model returns structured data that conventional software can inspect and use. Liu released Instructor 1.0 in 2024, extending support across providers and programming languages. OpenAI later named Instructor among the projects that inspired its Structured Outputs API.
- Structure is an engineering contract. Liu treats schemas as executable specifications encompassing the prompt, response, and business rules. Validators can catch fabricated URLs, inconsistent receipt totals, and unsupported claims; informative errors enable targeted retries. Typed search requests can encode dates, sources, and dependencies, letting conventional software execute retrieval plans and parallel searches. His work on Pydantic and language models argues that developers should retain ownership of objects, functions, and control flow.
- Impact-weighted product evaluation. His open-source Kura project summarizes and clusters chatbot conversations to identify recurring requests, frustration, tool failures, and differences between user segments without exposing individual conversations. Combining usage with performance reveals whether a team should repair a heavily used weakness, promote an overlooked capability, or disregard an inconsequential edge case. Evaluation becomes a product-roadmap instrument instead of an isolated quality score.
- Persistent agent threads. At OpenAI, Liu applies the same discipline to sustained knowledge work. His personal monorepo template organizes project context, reusable instructions, and durable task records. He distinguishes fresh scheduled tasks from recurring messages to existing threads, enabling agents to retain earlier decisions, monitor unresolved work, coordinate delegated tasks, and respect stopping conditions.
- Computer use with permission boundaries. Liu differentiates native computer control, authenticated Chrome sessions, and the in-app browser according to the access each task requires. He emphasizes approval and organizational safeguards around consequential actions, recognizing that an agent capable of switching interfaces can also circumvent restrictions unless its permissions remain explicit.
Liu also writes about taste, attention, and judgment: as AI makes execution easier, recognizing quality and articulating what needs improvement become more valuable. His approach makes automation serve informed human decisions, not replace them.
Topics covering these talks
Talks by Jason Liu
5 talks▶ Watch ↗AI Engineer World's Fair 20261:15:02
Full Workshop: Setting Yourself Up for Success — Jason Liu, OpenAI Codex
Read the full talk →▶ Watch ↗AI Engineer World's Fair 202519:23
How to look at your data; what to look for, how to measure
Read the full talk →
▶ Watch ↗AI Engineer World's Fair 202435:21
What We Learned From A Year of Building With LLMs
Read the full talk →References
- github.com/jxnl/blog ↗
- github.com/jxnl ↗
- github.com/jxnl/systematically-improving-rag ↗
- ai.engineer/talks/full-workshop-setting-yourself-up-for-success-jason-liu-openai-codex ↗
- ai.engineer/talks/how-to-look-at-your-data-what-to-look-for-how-to-measure ↗
- jxnl.co ↗
- ai.engineer/talks/structured-llm-outputs-with-pydantic ↗
- ai.engineer/talks/pydantic-is-still-all-you-need ↗
- ai.engineer/talks/what-we-learned-from-a-year-of-building-with-llms ↗
- ai.engineer/worldsfair/2026 ↗