OpenAI Codex lead on the new shape of product work | Andrew Ambrosino
Andrew Ambrosino, OpenAI's Codex app lead, discusses how AI is inverting product development, emphasizing "taste" and curation over implementation. He shares the vision for Codex as a universal home base for all knowledge work, integrating with other tools.
Deep Dive Analysis
16 Topic Outline
Introduction to Andrew Ambrosino and Codex App
AI's Impact on Product Development Process
The Role of Taste in Product Curation
Why AI Lags in Design Capabilities
Rethinking the "Dead" Design Process
Role Collapse and Team Structure at OpenAI
Planning Roadmaps in a Rapidly Changing AI Landscape
Building Features for Future Model Capabilities
The Ambition Problem and AGI-Pilled Products
Latest Frontiers: Loops and Autonomous Development
Andrew's Personal AI Workflows with Codex
The Vision for Codex as a Universal Home Base
Codex Interacting with Other Apps (Premiere Pro Example)
Lessons from Career Failures
Lightning Round
Post-Recording Insights on AI and Process
7 Key Concepts
Inversion of Product Process
The shift where implementation (building features) becomes cheap and abundant due to AI, making the expensive part the curation, steering, and selection of ideas.
Taste (Professional Skill)
Beyond aesthetics, taste in an AI-first workplace involves systems thinking, understanding wider context, knowing what to work on, how to present information, and selecting the right medium to achieve goals.
Primal Mark
The initial creation in a design or art process (e.g., the first mark on a painting) that establishes a starting point, to which all subsequent work responds and often anchors.
Role Collapse
The blurring of traditional job function boundaries (e.g., engineering, design, product management) where individuals increasingly perform tasks across disciplines, with their role defined by the average of their activities.
Zone Defense (Product Management)
A product management strategy where product people spread out to cover the entire company, identifying gaps and steering initiatives across a wide range of uncoordinated explorations, rather than focusing on specific, tightly defined areas.
AGI-Pilled
A state of being overly ambitious or optimistic about the immediate capabilities of AI models, leading to product designs that are too advanced for the current state of the technology or market readiness.
Dogfooding Loop
A continuous feedback mechanism where product developers actively use their own product for their daily work, identifying areas for improvement and driving the product's evolution based on personal experience.
7 Questions Answered
AI has inverted the product development process, making implementation cheap and abundant. The focus has shifted from de-risking expensive implementation upfront to curating and steering from a multitude of rapidly built prototypes.
AI struggles with design because grading design is harder due to the human aspect of taste, and historically, labs prioritized models that accelerate AI research (like code generation) over design. Design also requires novelty and an understanding of deeper abstraction layers in software.
While there's significant role overlap and blurring of boundaries, completely eliminating roles like product management is dangerous because it risks discarding established best practices and specialized knowledge within those disciplines.
Planning for the short term requires more detail, but long-term plans (e.g., 9 months out) must remain hazy to avoid false precision. The approach involves listing potential features, prototyping them, and waiting for model advancements to make them viable.
The ambition problem occurs when teams are "AGI-pilled," designing products that are too ambitious or advanced for the current state of AI models or market readiness, leading to early failures even if the core idea is sound.
AI tools like Codex can automate tasks such as aggregating daily briefs from multiple communication channels, answering questions, and updating status trackers, by setting up scheduled tasks and coaching the AI through iterative feedback.
The vision for Codex is to be a universal "home base" for all knowledge work, seamlessly interacting with other apps (like Excel or Premiere Pro) through connectors, computer use, or extensions, rather than trying to replace them entirely.
12 Actionable Insights
1. Prioritize Curation Over Building
In an AI-first world where implementation is cheap, focus shifts from building to curating the best ideas from many prototypes and aligning them with product goals and user needs.
2. Choose the Right Medium
Don’t default to prototypes just because AI makes them easy. Use documents for product clarity in vague areas and prototypes for stress-testing interaction patterns.
3. Decouple Medium from Process Stage
Be clear about where something is in the design process, regardless of how polished a prototype looks. Avoid over-anchoring on production-ready visuals if the underlying assumptions are still exploratory.
4. Embrace Role Overlap
Recognize that roles in AI-first teams have more overlap, defined by the average of what people work on rather than strict boundaries. This fosters agency and allows individuals to contribute across disciplines.
5. Avoid Eliminating Roles Entirely
While boundaries blur, don’t abandon the concept of roles, as it risks discarding accumulated best practices and specialized skills in disciplines like product management.
6. Adopt a “Zone Defense” for Product
Product teams should spread out to cover the entire company, identifying gaps and guiding initiatives from inception to product, rather than relying on top-down, long-term planning.
7. Plan with Hazy Long-Term Goals
For long-term planning (e.g., nine months out), keep details hazy, as precision is false and wastes time. Focus on what models might be capable of and iterate as model capabilities evolve.
8. Build Features for Future Models
Develop features that may not work well with current models, knowing that future model advancements could make them viable. Be prepared to re-release the same feature multiple times as intelligence improves.
9. Align Personal AI Use with Product Goals
Actively use your product for your own job, aligning your personal workflows with the problems the product aims to solve. This “dogfooding” helps identify issues and drive product evolution.
10. Coach AI Automations Iteratively
When setting up AI automations, expect to refine them. Start with a scheduled task, then coach the AI through steering and feedback to improve its performance and notification accuracy over time.
11. Be Ambitious and Experiment
Don’t be afraid to try ambitious things with AI, even if the product isn’t explicitly designed for it. Curiosity and intentional outcomes can reveal unexpected and useful applications.
12. Don’t Marry Your Process
Avoid getting too attached to your current processes. Instead, focus on the unique outcomes you can deliver and be willing to change your process to adapt to new tools and capabilities.
9 Key Quotes
The implementation is actually not the expensive part anymore. It's, dare I say, taste. But it's the curation process.
Andrew Ambrosino
If implementation is abundant, then it's really important to pick the right format for the point you're trying to make.
Andrew Ambrosino
Paul Graham clearly has great taste and wears cargo shorts, right? Like, you know, we gotta, we gotta like tease out what taste means a little bit.
Andrew Ambrosino
There's an amount of like novelty that is more important in design than it actually is in software engineering.
Andrew Ambrosino
The problem now is that you can pull all of the implementation into that. And there's a mismatch between, I think a lot of assumptions.
Andrew Ambrosino
I am very confident that the codex app that we released in February, if that had been ready in November, it would have absolutely failed in the market. And that the only difference was the models between November and February.
Andrew Ambrosino
If research is listening at any company, please make the models better at deleting code.
Andrew Ambrosino
I've tried to align my own usage of it with the problem that we're trying to solve.
Andrew Ambrosino
Do not get married to your exact process, get married to like, the outcomes that you were uniquely able to deliver and then do things like change your process to try things.
Andrew Ambrosino