The Golden Age of AI Engineering — Alexander Embiricos & Romain Huet & Peter Steinberger, OpenAI

OpenAI's Dev Day 2024 demo ran on an o1 preview model that could not run or check its own code, so Romain Huet had to cross his fingers live on stage. A year later, the same kind of demo ran a full camera and lighting rig, because the model could now test its own work. Alexander Embiricos and Huet use that jump to show how fast Codex is moving: new model releases went from every 15 months to about every 6 weeks.

They walk through why Codex is built on the same responses API, open source harness, and AGENTS md file format that ships to every developer, plus new numbers on cost and speed: frontier level intelligence at $1 per million input tokens and $6 per million output tokens, and a model generating 750 tokens a second, fast enough to produce a real pull request in about 10 seconds.

Speaker info:
- https://twitter.com/embirico
- https://www.linkedin.com/in/embirico

- https://twitter.com/romainhuet
- https://www.linkedin.com/in/romainhuet

- https://twitter.com/steipete
- https://www.linkedin.com/in/steipete

Timestamps:
0:00 - Introduction
0:13 - The World's Fair analogy
1:10 - The role of AI engineers in the future of work
2:14 - Accelerating model development cycles
3:09 - Evolution of build-and-test model loops
4:03 - Scaling engineering capabilities through agents
5:30 - Defining the desired AI engineering product experience
7:59 - The design philosophy of the Codex app
9:44 - The open-source stack and building with API primitives
11:43 - Expanding the ecosystem with Apps Server and plugins
14:20 - Optimizing for "Value Maxing": Cost and Intelligence
15:48 - Achieving high-speed inference for real-time workflows
16:51 - Future outlook: Removing the local/cloud distinction
18:16 - Special guest introduction: Peter Steinberger
18:56 - Shifting from manual orchestration to managing agents
20:02 - Three key changes for scalable agent loops
21:19 - Redefining the bottleneck as human attention
22:08 - Workflow example: Automating open-source issue resolution