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

Andrew Ambrosino describes a reversal in product development: models make it possible for many people to build many versions of an idea, so implementation is no longer the main cost. At OpenAI, he says, nearly 90% of employees use Codex, not only engineers. He wants the app to become a natural desktop destination, with a quality bar high enough that people open it without hesitation.

In the older process, teams researched, wrote documents, and prototyped to reduce the risk of expensive implementation. With AI, dozens of loosely coordinated explorations can appear before anyone has settled the product direction. Ambrosino says the harder work is now choosing what is good, combining the useful parts, framing the feature, and deciding how it fits with the rest of the product. He calls that work curation—and says it depends on taste.

Ambrosino pushes back on the claim that PRDs are dead and prototypes are in. Cheap implementation makes prototypes tempting, especially for people who previously could not code. But documents can still clarify an ambiguous product problem, while an interactive prototype is better for testing an interaction pattern. The important change is not to abandon one medium; it is to select the medium that matches the question.

A prototype can also create a misleading signal. In the past, a production-like artifact usually meant that research and design had progressed and key assumptions were checked, because reaching that fidelity was costly. Now an exploratory build can look finished before anyone has validated whether it reflects user needs or business goals. Teams therefore need to say plainly what stage a prototype represents instead of letting polish stand in for readiness.

Ambrosino defines taste as more than aesthetics. It includes judging whether an interaction fits its meaning, understanding how a feature belongs in a larger system, choosing the right presentation, and deciding what is worth building in the first place. His example of websites copying Linear illustrates the difference between reproducing a successful style and producing something novel. Design may need more novelty than software engineering, where reuse of established patterns is often desirable.