He is answering a question about research, not product, and he compresses about four steps into one sentence. Unpacked:

1. The bottleneck was researcher attention. OpenAI's research effort at that point was pointed at reasoning models trained with RL — his “clean RL.” Making a model good at the messy, interruptible business of coding with a person was a different target, and nobody senior was pushing on it.

2. One researcher was. Michelle Pokrass — the transcript spells it Pokress — who works on post-training. She wanted to push interactive coding capability.

3. But the Codex team had nothing to offer her. This is the bit that reads as a non-sequitur. At that moment their only product was Codex Cloud, which is asynchronous: you hand off a task and come back. There was no interactive surface, so there was no usage, no evals, no feedback loop to hand a researcher. Hence “arm's length, just vibes, kind of hanging out a little bit” — they could talk to her, but they could not collaborate, because collaboration between a product team and a post-training researcher means giving them something to measure against. So her work was mostly driven by other OpenAI customers asking for better interactive coding, not by Codex.

4. Then the CLI gave them the surface, and they inherited her momentum. Once they shipped it, the improved interactive coding capability already in flight had somewhere to land. “Leverage that momentum” means they didn't have to win the argument for that research — it was already happening for someone else's reasons.

5. And that inverted the relationship. Codex became first a significant and then the biggest consumer of those models. Being the biggest internal customer is what buys you researcher attention — which is the same flywheel he described one paragraph earlier at turn 232, seen from the research side instead of the morale side.

The short version: they didn't get the model by asking for it. They got it by building the one surface that made the model's improvements visible, and the improvements were already coming.

Transcript

Keyboard