
The New Frontier: Supervised vs Unsupervised

Why OpenAI is merging Codex and ChatGPT and the future of knowledge work | Andrew Ambrosino
Topic 11 of 17
Published Jun 28, 2026 · Clip 39:21–42:09 (2:48)
Andrew jokes 'loops are so last week.' With last year's goalposts, 100% of the Codex product is AI-written code; the meaningful question is whether code is written supervised or unsupervised, and he welcomes the moving goalposts as product progress. Teams are exploring autonomously developed software and harness engineering — like an agent coming in overnight to garbage-collect the codebase — but a core problem is that all models increase complexity, so he begs researchers to make models better at deleting code. Full autopilot development isn't here yet: models must learn which feature requests to build, ignore, or reframe, and how to build the right abstractions; a self-improving loop listening to Twitter, Slack, and email isn't reality yet, though they're trying. Asked whether agents will one day just 'make me a billion dollars,' he says he's not in the business of saying never or always.