This episode features Tara Seshan, product lead for Codex and ChatGPT Work at OpenAI, discussing how AI is transforming product development, the evolving role of product managers, OpenAI’s unique culture, and practical ways she uses AI tools in her daily work. Tara previously spent six years at Stripe as an early PM, led product at Watershed, and was a Thiel Fellow. The conversation covers OpenAI’s founder-like culture, the shift from theoretical to empirical product strategy, the “rowing vs. steering” metaphor for human-AI collaboration, why ambition matters more than ever, building for models 2-3 months out, ChatGPT’s Chat/Work modes, and specific tools like Sites and /visualize.
OpenAI’s culture is founder-led and radically open
Everyone at OpenAI operates like a founder in their area with minimal top-down direction, creating thin distance between teams and users.
Tara expected a “secret strategy room” but found OpenAI is genuinely open — internal thinking quickly becomes public product or messaging.
The cycle from idea to user-facing product is faster than anywhere else she’s worked.
Cultural memes include: “Is this maximally accelerated?”, “Are you mainlining it yet?” (using the product all day), and “feeling the AGI” — keeping the mission of beneficial AGI central.
AI product strategy requires fast experimentation over grand theory
In static markets (like payments), rigorous long-term strategy works because outcomes are predictable; in AI, the future is too emergent and fast-changing.
Being prolific and empirical matters more than being academic or theoretical — the key is forming a sharp hypothesis and testing it with users as fast as possible.
The “eigenquestion” (Shishir Mehrotra’s term): identify the single most important thing to test; everything else is noise.
You fail if you build for current models or for models a year out — the only viable horizon is 2-3 months, tied tightly to research roadmaps.
The PM role is shifting to hypothesis definition and rapid testing loops
The core PM job — defining the essential question, testing it, learning, and iterating — hasn’t changed but has become the only thing that matters; execution trappings have fallen away.
Everyone (PMs, engineers, designers, data scientists) now focuses on the problem-definition-and-testing loop.
Work increasingly looks like “steering” (setting direction, making opinionated calls) while agents handle “rowing” (tactical execution).
Steering operates at higher abstraction levels over time (line of code → feature → goal → vision), but human judgment on what to build remains essential.
If everyone has the same tools, the differentiator is human taste, ambition, and opinionation — software is more like filmmaking than real estate.
Agents are becoming persistent coworkers, enabling multiplayer collaboration
The next shift: agents as persistent teammates that work independently, sync at cadences, and incorporate feedback — like human coworkers.
Currently most agent work is solo; the frontier is multiplayer: “my agent collaborates with your agent” on shared tasks.
Internal OpenAI teams already share Codex threads; the goal is a natural multiplayer interface where agents work together under human steering.
Tactical enablers (not just intelligence) make this possible: local data access, cloud infrastructure, third-party system integrations — an agent without access to your tools is like a locked-in new hire.
Ambition must expand because execution constraints have collapsed
The most effective AI users don’t just automate rote tasks — they expand the set of things they can do (design, prototype, model, analyze).
Previously, “unicorn” people combined PM/engineering/design skills; now AI gives everyone that range, so ambition becomes the bottleneck.
The hardest part is remembering to try: “Could Codex do this?” must become a habit.
PMs should elevate team ambitions: “Isn’t the possibility ceiling higher? Can we try this faster? 10x bigger?” (Tyler Cowen’s framing).
Patrick Collison’s “fast projects” list (unreasonably ambitious things done quickly) should explode with AI — if they were possible before, they’re trivial now.
Building for models 2-3 months out is the only viable strategy
Kevin Weil’s line: “This is the worst the models will ever be” — hard to internalize but essential.
Product must get out of the model’s way; constructs should anticipate near-future capabilities.
Tight communication with research is critical: know which capabilities are being actively improved (e.g., coding, writing) and align product roadmap accordingly.
ChatGPT’s Chat and Work modes serve different starting points, but will merge
Three current entry points: Codex (dev-oriented UI), ChatGPT Chat (conversation/search), ChatGPT Work (Codex power with knowledge-work UI).
Work mode is Codex under the hood — same power, different UI (no work trees, less technical detail exposed).
North star: users shouldn’t choose — the system picks the right harness/model automatically.
Goal: bring agent power to ChatGPT’s billion users without requiring them to understand “harnesses” or technical concepts.
“Done is better than perfect” — shipping transformative capability early and iterating publicly beats pre-launch polish in this era.
Codex’s vibe shift came from sustained user obsession, not a single breakthrough
The team always operated with tight user focus, dogfooding/mainlining, and independent founder-like ownership.
External perception changed because the market caught up; internal operating mode didn’t.
Credit goes to individuals (Tibo, Andrew, desktop team) who notice gaps, build fixes, test internally, iterate, and ship.
Traditional roles are blurring as capabilities expand
Startup mentality: “everything and nothing is your responsibility” — accountability for outcomes matters more than role boundaries.
Someone must own “is this product used, wanted, high-quality?” (DRI), but execution can be picked up by anyone based on affinity/capability.
Tension: some craft skills are abstracted by models; craft shifts to higher-level application (e.g., engineers move from writing code to architectural steering).
No settled answer yet on balancing fluid collaboration with craft depth.
Humans remain uniquely valuable for accountability, expression, and care
Accountability: someone must own the outcome — quality, intent, regulatory/compliance needs.
Expression: software has artistry/opinionation (like film); what you choose to build and how it feels is human authorship.
Care: relating to teammates, collective enthusiasm, learning together, elevating each other’s ambitions — deeply human and more important than ever.
Tara’s personal AI workflow: Sites and /visualize
Sites: builds dynamic, shareable, auto-updating web apps via prompt (dashboards, games, trip planners, presentations) — realizes “malleable personal software” vision.
Available in ChatGPT Work, Codex, web, mobile; hosts publicly/team/private.
/visualize in Codex: pulls data and creates polished, interactive visualizations instantly — transforms “charts + narrative” workflow.
Both tools shift daily work from static artifacts (docs, sheets, slides) to living, interactive surfaces.
Writing to think vs. writing to report — only automate the latter
Writing as thinking: outlining, drafting, editing, iterating to clarify ideas — never automated; the act is the thinking.
Writing as reporting: status updates, launch plans, summaries — happily automated.
At Stripe, long briefs were the currency of alignment; at OpenAI, mocks/prototypes/results communicate better than docs.
Tara still writes hundreds of docs — but for herself, not as shared artifacts.
Best collaboration: write to 70%, share with rough edges, let others polish with you (perfected ideas repel input).
Avoiding AI brain rot: write the first and last draft yourself
Discipline: start the brief yourself, end it yourself; use AI only in the middle (research, data, pushback on ideas).
Personal rule: if asking others to read, invest at least the collective reading time in writing it first.
Don’t use AI for prose polish or first drafts — preserves thinking sharpness.
Sutter Hill taught her product marketing fit precedes product market fit
Sutter Hill (Snowflake incubator) has a repeatable playbook for B2B PMF, not luck/dark art.
Key insight: narrative/positioning/pitch should be tested before building — pitch 100 people, refine the story, then commit to product shape.
Mike Speiser excels at this; Tara previously underrated PMM as “glue,” now sees it as transformative.
Recruiting excellence (secret “Reticle” tool mapping talent networks) is another superpower.
Knowledge work differs fundamentally from coding — requires process visibility
Coding is output-verifiable (tests pass/fail); knowledge work requires trusting the process (reasoning, citations, inputs).
Product must adapt: show in-progress work, citations, chain of thought, so users can verify how the answer was reached.
Collaborator model: ChatGPT should be a thinking partner, not just an answer generator — surface the journey to the output.
Context access (email, docs, systems) is equally critical — an agent blind to your data is ineffective.
Lightning round highlights
Books: Barbarian Days (William Finnegan) — passion without mastery; Anna Karenina — re-reading reveals new layers at different life stages (growth metaphor for careers).
Movies: The Odyssey (Nolan) — secretly about AI and moral collapse; Rashomon (Kurosawa) — pioneering multi-perspective storytelling under constraints, reminder that tools don’t excuse low ambition.
Favorite external AI product: friends’ custom tools (cozy software movement) — e.g., Sebastian’s app that turns articles into private podcast feed, and “GATS” private social network for close friends.
Life motto: Toni Morrison’s four tenets — “Whatever the work is, do it well. Not for the boss but for yourself. You make the job. It doesn’t make you. Your real life is with your family. You are not the work you do, you are the person that you are.”
Thiel Fellowship: inflection point where someone elevated her ambition; cohort included Dylan Field (Figma); Ari Weinstein (now leads computer use at OpenAI) is a standout creative peer.
