Episode Summary
Executive Summary: Andrew Ambersino argues that AI has inverted product development: implementation is now cheap, so the hardest work is choosing the right problem, medium, and product shape. He says Codex’s success comes from rapid dogfooding, strong taste, role fluidity, and timing with model capability. The episode explores how AI is collapsing traditional PM/design/engineering boundaries while preserving needed specialties.
Main Topics: AI has inverted the product process (Priority: 5/5): Andrew says teams no longer need to de-risk implementation upfront; they can prototype and build quickly, so product work shifts toward selecting what matters, what medium to use, and how to curate many parallel attempts. Taste becomes the core product skill (Priority: 5/5): With code generation and prototypes abundant, the scarce capability is taste: judging what to build, how to frame it, which artifacts to trust, and which ideas deserve production investment. Design is changing, not disappearing (Priority: 4/5): He argues the old design process is largely obsolete because implementation is no longer expensive, but design remains essential as a process overlay, especially for choosing the right stage, artifact, and interaction model. Role collapse and cross-functional work (Priority: 5/5): Codex’s team has blurred boundaries among design, engineering, and product. People are more hybrid, but he warns against eliminating specialties or assuming everyone can do everything equally well. Timing, model capability, and product-market fit (Priority: 4/5): Many product ideas only work when models are mature enough; the same shape can fail months earlier and succeed later. Codex’s release timing and model improvements were critical to adoption. Codex as a home base for work (Priority: 4/5): The long-term vision is a general work surface that can coordinate across apps, connectors, browsers, and desktop tools, serving as a hub for both specialized and general knowledge work. Dogfooding, automation, and AI-assisted management (Priority: 4/5): Andrew uses Codex for scheduling, briefings, research, and release coordination. The team’s product evolution is driven by internal use, with the app increasingly acting like an assistant for real operational work.
Key Arguments: Implementation is no longer the bottleneck; curation and judgment are. Rapid prototype generation creates many competing explorations, so teams need taste to select and combine the best parts. The medium matters: documents are still useful for clarity, prototypes for interaction testing, and polished artifacts can mislead if they imply a later-stage product than they are. Design is harder for AI to master than coding because feedback is subjective, culturally variable, and more difficult to grade. A product can fail at one model capability level and succeed later without changing shape; timing with model intelligence is crucial. Traditional roles are blurring, but specialties still matter; eliminating roles can erase hard-won best practices. Codex should be a home base that coordinates work across tools rather than a single-purpose app or a fully self-contained super app. Dogfooding and internal usage are central to shaping the product and identifying the next primitives to productize.
Data Points: OpenAI employee Codex usage: 90% of the entire company - Andrew noted internal adoption is company-wide, not just engineering. Codex weekly active users: over 5 million - The host cited current weekly active users for Codex. Usage growth since January: 6x - The host said Codex usage has grown 6x since January. Team size: double digits of engineers - Andrew described the Codex team as having double-digit engineers, about half as many designers, and a few product people. Prototype count: 90 different attempts/explorations - Andrew used this as an example of many uncoordinated teams working on the same feature area. Release timing example: November vs February - He said the Codex app released in February would likely have failed if launched in November due to model capability differences. Product release example: May-ish release - He referenced a release that added browser/computer-use/artifact creation capabilities and changed release management workflows. Company-wide role coverage: marketing, comms, finance, legal, and engineering users - He described internal adoption across many functions, not just technical teams.
Pivotal Quotes: "The implementation is actually not the expensive part anymore. It's, dare I say, taste." — Andrew Ambersino: On how AI has changed the product development bottleneck "If you are tied to the tools in the exact day-to-day specific of the process, then yeah, it's dead." — Andrew Ambersino: On whether the traditional design process still applies "It's the average of where they're working." — Andrew Ambersino: On how he thinks about a person’s role now across design, engineering, and product
Implications: AI teams will increasingly win through judgment, timing, and hybrid roles, not just execution speed. Product leaders should optimize for curation, dogfooding, and flexible tooling while preserving disciplined specialties and avoiding premature role elimination.
About Lenny's Podcast
Lenny Rachitsky interviews world-class product leaders and growth experts about building products and growing careers.