The end of git? OpenAI CTO on the future of coding.

Summary

The interview examines how AI coding agents are changing OpenAI’s engineering workflows and the infrastructure needed to support them.

AI coding agents and engineering bottlenecks

The guest describes a shift beginning in February 2026, when stronger models and parallel Codex agents led to a sharp increase in pull requests. Token usage rose 17-fold at frontier labs over the prior year, compared with about twofold at conventional software companies. He cautions that token use is only a proxy for AI adoption, not a direct measure of productivity.

Because code is cheaper to produce, teams can prototype more ideas—and discard more of them. However, code review, testing, audits, and CI/CD can become bottlenecks.

Repository infrastructure for agents

OpenAI’s monorepository traffic reportedly exceeded what its existing infrastructure was built to handle. The guest says Git’s offline-oriented assumptions are a poor fit for agents that continuously fetch, test, and submit code, so OpenAI created an internal repository system for its monorepo.

Other scaling challenges include agent traffic, sparse checkouts, and distributing caches so agents do not repeatedly burden the central server. Keeping the main branch stable is also difficult, though the guest considers throughput and repository access more pressing issues.

Voice as an interface for more users

ChatGPT is described as having 1.2 billion users, but the guest argues that a text box may not suit the next billion—especially people who are less comfortable typing or using English. He sees voice as an important interface, citing the use of voice messages and the difficulty of typing in some languages.

He prototyped a voice-first app experience with Codex, including activating the microphone immediately while the connection is established, then shared it with the product team for further development.

Prototyping ideas as a leader

Rather than only presenting concepts to teams, leaders can use coding tools to turn ideas into working prototypes, validate them, and communicate them through something people can try. The speakers describe this as making product development more accessible beyond engineers.

A test for products built on models

The guest suggests asking whether a product benefits when its underlying models improve. If each model release makes the product more valuable, that is a positive sign. If releases threaten to replace its core design, the product may be too dependent on capabilities that will soon be commoditized.

Advertising system and team

The guest says OpenAI’s advertising team reached a reported $1 billion in annual revenue within roughly seven or eight months. He attributes its speed to using LLMs across parts of the ad pipeline—such as query rewriting, candidate selection, and ranking—rather than building a separate machine-learning model for every stage.

He says the team is small, at about 50 people, and that improving models can make the overall system better.

Reviews, guides, or tutorials

No product review or step-by-step tutorial is presented. The video is an interview with analysis of engineering infrastructure, product design, and AI-enabled development. The host mentions that a longer interview and in-depth analysis are available to paid subscribers on Substack.

Main speakers and sources

• Nate B. Jones: Interviewer and narrator
• Vijay Raji: OpenAI’s CTO of Applications and former Statsig founder