Key Moments
OpenAI Codex lead on the new shape of product work | Andrew Ambrosino
Lenny's PodcastWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
OpenAI's Codex app has grown 6x since January and is used by nearly 100% of employees, blurring traditional product roles.
Key Insights
Codex has seen usage grow 6x since January and boasts over 5 million weekly active users, with nearly 100% of OpenAI employees using it weekly.
The traditional product development process, where implementation was expensive and ideation was cheap, has been inverted; now, implementation is cheap, and 'taste' (curation and judgment) is the expensive, valuable part.
AI models are currently not as adept at design as they are at coding because grading design quality requires a human aspect of taste, making the feedback loop more complex than code compilation.
The concept of 'role collapse' is accelerating, where traditional boundaries between product, design, and engineering are blurring, leading to roles being defined by the average of tasks performed rather than a fixed title.
The future of product work may involve individuals who are 'agentic' and possess 'taste,' capable of taking ideas from inception to completion and steering AI-driven development, rather than relying on rigid role definitions.
Building with AI requires embracing ambition and iterative development, with the understanding that a product's success can be heavily dependent on the underlying model's capabilities, necessitating a willingness to release and refine multiple times.
The inversion of product development: From expensive implementation to valuable taste
The traditional product development process, characterized by extensive research, ideation, and extensive documentation before costly implementation, has been fundamentally inverted by AI. Andrew Ambrosino highlights that with AI, the implementation phase has become significantly cheaper, allowing for rapid prototyping and iteration. This shift means that the truly valuable and expensive part of product work is no longer building the thing, but rather 'taste'—the ability to curate, judge, and select the best ideas from numerous AI-generated prototypes. At OpenAI, this is evident as employees, not just engineers, are empowered to build and explore ideas, leading to potentially dozens of uncoordinated explorations for a single feature. The challenge then becomes rigorous curation, identifying what's good, and synthesizing these ideas into a cohesive product. This inversion necessitates a new skillset focused on discernment and strategic decision-making, moving away from the sole emphasis on execution.
Codex: The app that's changing how work gets done
The OpenAI Codex app exemplifies this new paradigm. Launched in November, its usage has surged, growing 6x since January to over 5 million weekly active users. Remarkably, nearly 100% of OpenAI employees, spanning all disciplines including marketing, finance, and legal, use it weekly, underscoring its broad utility beyond engineering. Initially conceived as a developer tool, its adoption by non-technical roles has highlighted its potential as a general knowledge work tool. This widespread internal adoption, despite the app having a UI hostile to non-coders, demonstrated strong product-market fit. The success of Codex is attributed to its ability to empower individuals with 'agency' and 'taste,' allowing them to rapidly prototype and iterate, leveraging AI to bring ideas to life with unprecedented speed and flexibility.
AI's current limitations in design and the 'primal mark' principle
While AI excels at code generation, its application in design is lagging. Ambrosino suggests this is partly because design quality is harder to grade objectively than code. The human aspect of taste is a critical feedback mechanism that current AI models struggle to replicate, making it more challenging to train them on what constitutes 'good' design. Furthermore, historically, AI research has prioritized areas that accelerate research itself, like writing correct code, rather than design. The concept of the 'primal mark' from art is also relevant; the initial output of an AI can unduly influence subsequent development, potentially limiting exploration of broader or different ideas. While AI will likely improve in design capabilities, current limitations mean human judgment remains vital for assessing aesthetic and functional quality.
The evolving nature of roles: Towards 'builder' mentalities and role collapse
The rapid pace of AI-driven development is leading to a significant blurring of traditional job roles. Ambrosino describes 'role collapse,' where the lines between product management, design, and engineering are becoming less distinct. At OpenAI, many designers and product managers can write code, and engineers possess strong product intuition. This shift is leading to roles being defined by the average of tasks an individual performs, rather than a static title. While some fear this could devalue specialized disciplines, Ambrosino emphasizes that it signifies a move towards a 'builder' mentality where individuals are valued for their agency and ability to deliver outcomes. The 'member of technical staff' designation, historically used in research labs, is becoming more common, reflecting this fluidity. However, he cautions against completely eliminating roles, as disciplines still have knowable best practices that can be lost.
The 'primal mark' of prototypes and the enduring value of documents
The idea that 'PRDs are dead, prototypes are in' is a popular sentiment, but Ambrosino offers a nuanced perspective. He acknowledges the temptation for non-engineers to jump straight to prototypes when implementation is cheap. However, he stresses that prototypes should not be the sole 'primal mark.' If a prototype is the first and only artifact, the team might anchor too heavily on its initial form, preventing exploration of different or better concepts. Documents still hold value for clarifying strategy and exploring problem spaces, especially when the core need is conceptual clarity. The key is choosing the right medium for the message: documents for broad ideation and clarity, prototypes for stress-testing interaction patterns. Mistaking a prototype for a final product can lead to shipping features that look polished but aren't strategically sound or meeting user needs, highlighting the continued importance of discernment.
Ambition, iteration, and the challenge of AI roadmap planning
Building with AI requires a greater degree of ambition and a tolerance for iterative development. Products might need to be released multiple times in different forms as AI models evolve. Ambrosino uses the example of the Codex app's evolution, where its success was contingent on model advancements between November and February releases. This means roadmapping has become more fluid; shorter-term plans require more detail, while long-term plans must remain intentionally hazy to avoid false precision. The success of a feature might depend less on its inherent design and more on whether the underlying AI model is sufficiently advanced. This necessitates a strategy of building and testing ambitious ideas, understanding that some may not work yet but can be revisited as AI capabilities improve. The challenge is balancing the need for exploratory, ambitious projects with maintaining the quality and reliability of existing products.
The vision for Codex: A generalist home base for work
The long-term vision for Codex is to become a 'home base' for all knowledge work, moving beyond its origins as a developer tool. While it can interact with other applications like Excel or Premiere Pro through connectors, computer vision, or extensions, it aims to be a central hub where users start and end their workdays. This 'super app' concept, though not explicitly named, suggests a unified experience that leverages specialized tools rather than replacing them entirely. For instance, Codex might not build a better video editor, but it can interact with Premiere Pro to perform editing tasks. This approach acknowledges the importance of specialized tools while empowering AI to orchestrate and automate workflows across various applications, creating a more seamless and efficient work environment. The goal is to integrate AI into daily tasks, allowing users to automate work and leverage the most suitable tools for any given job.
Embracing failure as a stepping stone to success
Ambrosino shares that his career has been marked by significant periods of failure, particularly as a startup founder and in highly regulated industries. These experiences, though challenging, were crucial learning opportunities. He notes that success, like with the Codex app, often arises from a confluence of skill sets, passion, and market timing. The internal feedback culture at OpenAI, where rigorous critique of product ideas is common, plays a vital role in refining offerings and preventing stagnation. This iterative process, involving cycles of "this sucks" feedback, eventually leads to better external products. The enduring lesson from these experiences is the importance of perseverance, continuous learning, and not becoming overly attached to a specific process, but rather to the desired outcomes.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Books
●People Referenced
Common Questions
AI is inverting the traditional product development process. Implementation, once the expensive part, is now cheap, shifting the focus to 'taste' – curation, alignment, and deciding what to build among many AI-generated possibilities.
Topics
Mentioned in this video
An AI-powered application developed by OpenAI, initially focused on coding but expanding to general knowledge work, aiming to be the best desktop app.
A company whose website design is mentioned as a benchmark and whose product is praised by Andrew Ambrosino as a favorite software product.
Mentioned alongside Figma as tools that allow for fast-forwarding insights by pulling interactive prototypes earlier into the design process.
Mentioned as an example of a product with a 'baby version' that approximates production app interactions, making it quicker to iterate on.
Mentioned as a comparison point to Codex, particularly regarding its desktop application and the potential for integration.
Mentioned as a complex interface requiring setup for tasks like API triggers, which Codex can now handle through computer vision.
Mentioned as a tool users might set up for personal systems or a 'mind palace,' suggesting that Codex could offer a more streamlined 'memory feature'.
Mentioned as a tool users might set up for personal systems, suggesting Codex could offer a more integrated 'memory feature'. Also mentioned as a SAS app that can be run within Codex.
A tool whose latency is contrasted with the speed of Codex's in-app browser when controlling websites.
Mentioned in the context of keyboard shortcut mapping challenges for Codex's browser functionality.
The video editing software used by Brent, an OpenAI videographer, who utilized Codex to build an extension for it and edit videos.
The AI research and deployment company where Andrew Ambrosino works. It is highlighted for its internal adoption of Codex and its culture of open feedback.
A sponsor of the podcast, providing enterprise features like single sign-on, SCIM, and RBAC through APIs for B2B SaaS products.
Mentioned as a tool that helped bring interactive prototypes earlier into the design process, simulating production.
Mentioned as a company that may have originated the 'Member of Technical Staff' title, which OpenAI also uses.
A sponsor providing banking services for entrepreneurs, highlighted for its product-centric approach and new 'Command' feature for financial operations.
Mentioned in the context of his taste and simple attire (cargo shorts) to illustrate that taste is not solely about aesthetics, referencing a tweet from the head of product at Linear.
A podcast guest whose prediction was mentioned: that users will start using Codex to run their SAS apps inside it.
More from Lenny's Podcast
View all 38 summaries
99 minBuilding the most AI-pilled engineering team in the world | Fiona Fung (Anthropic)
100 minThe hidden pattern behind successful products | Mark Pincus (FarmVille, Words with Friends, & more)
96 minFather of the iPod and iPhone on building taste, judgment, and creativity in the AI era
80 minA rational conversation on where AI is actually going | Benedict Evans
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free