Lenny's PodcastOpenAI Codex lead on the new shape of product work | Andrew Ambrosino
At a glance
WHAT IT’S REALLY ABOUT
AI makes building cheap; taste, curation, and agency become scarce
- AI has flipped product work so implementation is no longer the bottleneck, and teams now struggle more with curation, coherence, and deciding what to build than with writing code.
- Documents and prototypes both remain valuable, but the key is choosing the right medium for the intent and avoiding over-anchoring on polished prototypes that imply false maturity.
- “Taste” is framed as multi-dimensional judgment—context, systems thinking, novelty, and framing—not just aesthetics, and it becomes a primary differentiator in an era of abundant AI-generated output.
- Frontier models lag in design because design is harder to grade and requires human taste feedback, plus deeper abstraction-level understanding (semantic component systems) beyond surface visuals.
- Codex is evolving from a developer tool into a “home base” desktop app that can orchestrate work across many tools via connectors, in-app browser, computer use, and extensions, while product planning adapts to rapid model capability shifts.
IDEAS WORTH REMEMBERING
5 ideasTreat implementation as abundant; prioritize selection and coherence.
Ambrosino argues many teams can now stand up working versions quickly, so the scarce skill is choosing which attempts matter, merging the best ideas, and keeping the product coherent across explorations.
Don’t declare “PRDs are dead”; pick the medium that matches the question.
Use docs for clarity and alignment in vague spaces, and prototypes when you need hands-on stress tests; the failure mode is using the wrong artifact and letting it anchor decisions prematurely.
Polish no longer indicates readiness—teams must re-signal “stage of process.”
Because AI can generate production-like prototypes early, the old heuristic (“looks real” means “late-stage”) breaks; teams must explicitly label exploration vs. decision vs. ship-ready work.
Define “taste” as contextual judgment, not visual preference.
Taste includes systems thinking, novelty, framing, and understanding where a feature fits strategically—plus the smaller but real layer of interaction/animation semantics and UX details.
AI design lags because feedback loops are harder than code correctness loops.
Code can be graded with compilation/tests, while design depends on human cultural preferences and novelty; models also struggle with deeper semantic abstractions that make design scalable (e.g., rebranding via shared tokens/components).
WORDS WORTH SAVING
5 quotes90% of people at OpenAI use Codex. Not 90% of engineers, that was 90% of the entire company.
— Andrew Ambrosino
It's, it's backwards, right? The implementation is actually not the expensive part anymore.
— Andrew Ambrosino
It's, dare I say, taste.
— Andrew Ambrosino
I think design's a little bit harder to grade... because the human aspect of taste is, is like part of the feedback mechanism you need.
— Andrew Ambrosino
If you're married to the exact process you have right now, like that, I don't know what advice to ever give, but if there's one piece it's like do not get married to your exact process. Get married to like the outcomes that you are uniquely able to deliver.
— Andrew Ambrosino
High quality AI-generated summary created from speaker-labeled transcript.
