OpenAI Codex lead on the new shape of product work | Andrew Ambrosino

3:07 – 10:22

When building became cheap, product work flipped upside down

The striking image is not one engineer with an AI copilot. It is Andrew saying that inside OpenAI, nearly everyone is building. His blunt summary is that product work is now "backwards." A few years ago, teams treated implementation like the expensive part. You wrote docs, ran research, made prototypes, and tried to de-risk every decision before asking engineers to spend precious time building the thing. That whole logic depended on one old assumption: software was costly to make. Now, Andrew says, "anybody can build anything" from scratch with frontier models. That has changed the economics of product development more than most teams have emotionally caught up with. At OpenAI, if there is a feature people want, there might be "90 different uncoordinated teams" already trying versions of it. The hard part is no longer getting a prototype into existence. The hard part is deciding which of those 90 attempts actually deserves to survive. This is where he introduces the word that keeps coming back through the conversation: taste. He knows it sounds fuzzy, so he tries to pin it down. Taste is partly aesthetic, yes, but much more than that. It is systems thinking. It is knowing how something fits into the product, what theme it serves, whether the interaction matches the meaning, whether a feature belongs here or should be folded into something else. In a world of abundant implementation, taste becomes the filter that keeps teams from drowning in their own output. Lenny pushes on whether this means documents are dead. Andrew says no. In fact, he thinks people are overcorrecting. Non-engineers now rush into prototypes because they can. Engineers, meanwhile, still produce long docs that are often not worth reading. His real point is simpler and more useful: choose the right medium for the question. If you need product clarity in a fuzzy area, a doc may still be best. If you need to feel an interaction, use a prototype. The danger is confusing polished-looking output with validated thinking. A prototype can arrive so production-ready that everyone forgets it was only meant to explore an idea. That is the new failure mode.