Source: raw/The_end_of_git_OpenAI_CTO_on_the_future_of_coding..md

Creator: Nate B Jones, AI News & Strategy Daily, interviewing Vijay Raji, CTO of Applications at OpenAI, who joined when OpenAI acquired Statsig, the company he founded · URL: https://www.youtube.com/watch?v=S6SEgMooZpE · Duration: highlights only; the full ~45-minute interview is on Jones’s paid Substack and was not reviewed · Platform: YouTube (auto-captions, fetched 2026-11-09).

Raji describes what happened inside OpenAI after coding agents took off in February 2026. Engineers produced so many more pull requests that review, CI/CD and finally the code repository itself became bottlenecks, and OpenAI built a new repository system for its monorepo because git assumes developers work offline. He also covers why the next billion ChatGPT users will speak rather than type, how he now prototypes ideas himself before taking them to a team, how the ads team reached a billion-dollar run rate with about 50 people, and a test for founders: are you excited or worried when a new model ships?

Key Takeaways

  • February 2026 was the turn. Stronger models plus Codex running multiple agents in parallel with worktrees caused “a massive uptick” in PRs. As a rough proxy he says token usage grew 17x in a year at frontier labs against about 2x at an average software shop, and calls token count a one-dimensional measure.
  • Cheap code changes what gets thrown away. Less discussion happens before building; “more code is being thrown away now than before” because prototypes are so cheap.
  • Then the bottlenecks move downstream:
    1. Review. When AI-assisted engineers produce about 5x more code, review cannot keep up; OpenAI has an auto-review system teams can turn on.
    2. CI/CD. The infrastructure had to be rethought to keep pace.
    3. The repository. Agents now pull code, run tests and push back to main continuously; traffic through the monorepo grew more than an order of magnitude beyond what the system was designed for.
  • Why git had to go (for them). Git was designed for offline work: check out, branch locally, push when done. Agents are online all the time. A team rebuilt the repository on that assumption; Raji says it is “two orders of magnitude more capable” in throughput, and all OpenAI engineers now work on it.
  • Merge collisions on main are not the worst part. If you do not lock main, throughput handles them. The bigger problem is checkout volume: “you have to like solve sparse checkout problem first”, with caches so agents are not constantly hitting main.
  • How it was assigned. A team got the problem and a mission, but no instructions, because nobody knew the answer. Internal engineers are “our hardest customers”, and the team had to keep the old system running while moving everyone over.
  • The next billion users will speak. ChatGPT has 1.2 billion users on “a text box and a button”. The next billion will be less comfortable in English and with typing; on a trip to India he found people send WhatsApp voice notes instead of typing, and Indic scripts are hard to type. So voice is his growth team’s focus.
  • Leaders prototype first. Instead of bringing an idea to a team, he built a voice-first ChatGPT prototype with Codex: strip the app down to voice, open the microphone while it is still connecting so people can start talking, then replay what was recorded once connected. Only after it looked right did he take it to the team, while acknowledging production is “a completely different set of problems”. Jones’s version: ask whether you have to write something down, or can show it by building it. Raji: “code wins arguments”, and now not only for engineers.
  • The founder litmus test. If each new model release worries you because it may absorb what you built, “you’re not building the right product.” If it excites you because your product gets better, you are. He says the smallest model available now is ten times more powerful than the biggest one of a year ago.
  • The ads team as proof. About nine or ten months old, it reached a $1 billion ARR run rate within seven or eight months of starting from zero code, with “about 50 or so people” by his estimate. The pipeline (ads index, query, query rewriter, candidate retrieval, auction, ranker, serving) used LLM capabilities instead of a separately built ML model for each stage, and “the model keeps getting better, which means… the entire ads system keeps getting better.”

Try It

  1. Before scaling coding agents, profile the downstream steps: review time, CI queue time, and checkout and clone time per agent. Fix the one that saturates first.
  2. Turn on automated first-pass review for agent PRs and send only flagged or risky ones to humans.
  3. For agent fleets on a big repository, use sparse checkouts and a shared cache before worrying about merge collisions.
  4. Bring a working prototype, not a proposal, the next time you pitch an idea to a team.
  5. Apply the litmus test to your product: list which parts get better when a new model ships and which become redundant, and move effort toward the first list.
  6. Where a pipeline has many small ML stages, try one LLM call per stage first so model upgrades improve the whole pipeline.

Open Questions

  • OpenAI’s replacement repository is not named or described beyond its design assumption; no external release is mentioned.
  • The 17x versus 2x token-growth figures and the “two orders of magnitude” capacity claim are given without sources or methods.
  • The ads team size is Raji’s guess on air.