AI 的第三个时代：持久化 AI 同事崛起 | OpenAI 产品负责人 Tara Seshan

Brief Description

Tara Seshan, product lead at OpenAI for Codex and ChatGPT Work, joins the podcast to discuss the transition from simple chat interfaces to persistent AI coworkers. She shares insights into OpenAI’s decentralized and founder-led operating culture, why product managers must shift from theoretical planning to empirical experimentation, and how teams must drastically elevate their ambition. Seshan also explains the product architecture behind ChatGPT's work and coding modes, contrasts the feedback loops in software development with knowledge work, details her philosophy on writing as thinking versus reporting, and shares key career lessons from Stripe and Sutter Hill Ventures.

Table of Contents

• Introduction and Culture at OpenAI
• The Empirical Product Manager
• Steering vs. Rowing with AI Agents
• Elevating Ambition and Internal Operating Memes
• Demystifying ChatGPT, Codex, and Work Mode
• The Evolving Role of PMs and Human Craft
• Tactical Workflows: Sites and Visualizations
• Writing as Thinking vs. Writing as Reporting
• Product Marketing Lessons from Sutter Hill Ventures
• Adapting AI Interfaces for Knowledge Work
• Lightning Round and Final Advice

Introduction and Culture at OpenAI

Host: Welcome to the podcast. Today, my guest is Tara Seshan. Tara leads product for both Codex and ChatGPT Work at OpenAI, which is arguably the fastest-growing and most important AI product for knowledge workers today. Prior to OpenAI, Tara spent six years at Stripe as one of the first five product managers, led product at Watershed, and was a Thiel Fellow.

Tara, you have been at OpenAI for about a year now. In AI time, that feels like a lifetime. When you joined, you probably had a sense of what working at a frontier lab would be like. What has most surprised you about the actual day-to-day culture at OpenAI?

Tara Seshan: Many aspects felt familiar because I had worked at high-growth, high-intensity, hyperscaling companies before. The talent caliber and the urgency felt very recognizable. The most surprising part is that every company I previously worked for was founder-led, whereas OpenAI is essentially "founders-led." Everyone inside the company operates as a founder within their specific product area.

The level of top-down direction is extremely limited compared to anywhere else I have worked. On one hand, that was delightful because the distance between you and the market is razor-thin. You are not insulated from user feedback; you do whatever it takes to find product-market fit.

On the other hand, I came in expecting a secret vault of grand strategy documents, similar to how an established company might have an authoritative operational manual. In reality, OpenAI is genuinely open. The thoughts and ideas about where models and products should go very quickly make their way directly into the public product. There is no secret backroom with an AGI running a master plan; what gets built is immediately put in front of users to touch and test.

The Empirical Product Manager

Host: Given your extensive background as a product management leader, what shifts or changes when operating in this new fast-moving environment?

Tara Seshan: In static or slower-moving markets, you can execute grand, predictable strategies. In payments, for example, you can reason rigorously from first principles about how competitors will respond. Rigorous theoretical thinking is required, and failing to do so comes across as careless because the variables are well understood.

In this market, everything is emergent and dynamic. Staying closely tied to research is essential. Because of that, being prolific and empirical is far more valuable than being academic or theoretical.

It was a major adjustment to stop writing long, thesis-style reasoning documents and instead ask, "How quickly can I build something to test with users?" At first, it felt like I was not doing enough due diligence. But the real rigor now lies in formulating a sharp, pointed hypothesis—the core question that determines whether a product succeeds—and testing it as fast as possible.

Host: It feels like that core testing loop is the enduring essence of product management, even as everything else shifts.

Tara Seshan: Exactly. All the peripheral mechanics of the role—tracking schedules, producing status presentations, and writing long decks—fall away.

The core job has always been identifying the essential question, understanding user and technology constraints, forming a crisp hypothesis, and analyzing the results in a rapid feedback loop. Today, engineering managers, designers, and data scientists all share that focus, which allows PMs to double down on what matters most.

Steering vs. Rowing with AI Agents

Host: There has been significant discussion around autonomous loops where an agent is given an objective and iterates until it succeeds. How do you see that loop expanding across general knowledge work?

Tara Seshan: The future of work will look more like steering than rowing. Agents will handle the rowing, while human professionals focus on steering the ship in the right direction.

Over time, that steering moves up layers of abstraction. It used to be writing a line of code and pressing tab. Now, it is directing tasks at the goal level, and eventually, it will be at even higher strategic levels.

Data guides some of that steering, but human intuition and opinionated decision-making remain vital. Building software is not like real estate, where pouring in capital guarantees a predictable return; it is closer to filmmaking. An artistic point of view and a distinct voice determine whether a product resonates. Furthermore, work will become multiplayer, where teams steer collaborative swarms of agents together.

Elevating Ambition and Internal Operating Memes

Host: Ambition seems to be a major differentiator now. Because foundational tasks are easier to execute, the bar for what people and companies can attempt has risen significantly. How do you view this shift?

Tara Seshan: The most effective users of AI do not merely automate routine tasks; they dramatically expand their range of capabilities. Previously, a unicorn hire was someone with great product sense who could also design and write code, eliminating cross-functional translation layers.

Now, anyone can spin up initial designs, build a prototype, model financial scenarios, and execute complex workflows independently. Because execution friction has dropped so substantially, people must actively expand their mental model of what can be accomplished in an unreasonably short timeframe.

A primary responsibility of a PM today is elevating the ambition of those around them. When someone proposes a standard timeline or scope, the PM should ask what the 10x version looks like and whether it can be attempted immediately.

Host: Are there specific internal mantras or memes at OpenAI that capture this mindset?

Tara Seshan: There are three central memes that guide product development:

First, "Is this maximally accelerated?" which focuses on execution speed. Second, "Are we being ambitious enough?" which challenges scope and scale. Third, "Are you mainlining it yet?" which replaces traditional dogfooding by asking whether the team relies on the product all day, every day.

Another essential rule is building for where the models will be in two to three months. If you build for where the models are today, you fail. If you build for where you think they will be in a year, you also fail. Tight alignment with research roadmaps allows us to design product surfaces that get out of the way as underlying capabilities advance.

Demystifying ChatGPT, Codex, and Work Mode

Host: Looking at the ChatGPT ecosystem today, users see different options like standard chat, Codex, and the Work toggle. How should users understand these distinctions?

Tara Seshan: Our ultimate North Star is that users should never have to navigate menus or understand harnesses. You should simply specify what you want to achieve, and the system will automatically route to the right model and runtime environment.

Currently, the interface meets users where they are:

Chat mode is optimized for conversations, brainstorming, and web search. Codex mode provides an explicit development environment with technical visibility into the code tree and execution chain.

Work mode operates on the exact same engine under the hood as Codex, but strips away developer-heavy UI. It allows knowledge workers to generate complex artifacts, such as intricate corporate finance models, without needing to interact with a code terminal.

Host: How do you balance managing a consumer platform with hundreds of millions of users alongside rapid, experimental agent development?

Tara Seshan: The objective of introducing Work to ChatGPT is bringing the power of autonomous agents to over a billion users who primarily know chat interfaces.

In this era, shipping a transformative workflow quickly beats holding back for cosmetic polish. Getting capabilities into users' hands reveals how they actually interact with agents and enables rapid post-launch iteration.

The Evolving Role of PMs and Human Craft

Host: As functional roles blend together and AI writes more code, many people wonder what their daily identity looks like. How do you think about role boundaries and professional craft?

Tara Seshan: Having spent much of my career in early-stage startups and Stripe, I prefer environments where functional boundaries are fluid. What matters is having a clear directly responsible individual (DRI) accountable for product quality and user adoption.

It is natural to feel a sense of loss when the manual mechanics of a craft—like an engineer writing every line of code by hand—get automated away. Craft is shifting upward toward higher-level system design, framing, and problem selection.

Humans will continue to provide ultimate accountability, subjective artistic taste, and interpersonal care. Rallying a team, creating alignment, and elevating collective ambition remain deeply human endeavors.

Tactical Workflows: Sites and Visualizations

Host: What are some practical, innovative ways you integrate AI into your daily product workflow?

Tara Seshan: One of the most powerful workflows is building ephemeral web applications, or "Sites," directly inside Codex and Work mode.

Alan Kay originally envisioned malleable personal software that users could construct on the fly. Today, with a single prompt, I can spin up a fully functioning interactive dashboard with its own database, build a live itinerary tracker for a backpacking trip, or create a custom multiplayer game for my team. Instead of sharing static slide decks or spreadsheets, sharing dynamic, hosted sites has become my primary presentation medium.

Another useful feature is using the /visualize command within Codex. By asking the model to visualize specific datasets or usage histories, it automatically generates clean, narrative-driven charts that make complex data immediately legible.

Writing as Thinking vs. Writing as Reporting

Host: You are known for producing exceptionally sharp product briefs. How do you approach writing in an age where models can generate text effortlessly?

Tara Seshan: I draw a strict distinction between two types of writing: writing as reporting and writing as thinking.

Writing as reporting includes status updates, launch plans, and weekly summaries. I happily automate those using AI.

Writing as thinking is the process of working through product strategy, evaluating trade-offs, and defining core hypotheses. I never outsource that to an AI. Outlining, drafting prose, and editing are how I refine ideas.

At Stripe, long documents served as durable artifacts of thinking. Today, long text is no longer definitive proof of deep work because anyone can generate a verbose document. Mocks, prototypes, and empirical experiment results are now far more persuasive. When I do write briefs, I aim for roughly 70% completion and then share the draft with colleagues so we can stress-test and polish the ideas together.

Product Marketing Lessons from Sutter Hill Ventures

Host: You also spent time as an Entrepreneur in Residence at Sutter Hill Ventures. What key insights did you take away from their incubation playbook?

Tara Seshan: Sutter Hill showed that building multi-billion-dollar enterprise companies is not pure luck; there is a repeatable methodology.

My biggest takeaway was the critical importance of product-marketing fit, which I had previously undervalued. The narrative, market positioning, and customer pitch should often precede building the software.

Testing your positioning against one hundred enterprise buyers allows you to refine why the solution is transformative. Once the narrative and buyer pain points are crystal clear, you commit to building the exact product shape that fulfills that promise.

Adapting AI Interfaces for Knowledge Work

Host: What is a fundamental challenge in expanding agent capabilities from coding into broader knowledge work?

Tara Seshan: Coding is highly output-verifiable. You can run automated tests, compile the program, and execute the code to verify that it functions properly.

Knowledge work cannot be validated solely by looking at the final artifact. An executive cannot just look at the final slide of a strategic plan and assume it is correct without understanding the underlying reasoning, inputs, and assumptions.

Because of this, AI interfaces for knowledge workers must focus heavily on collaboration, transparent reasoning steps, and clear source citations. Users need to follow along with the journey of how a conclusion was reached to build genuine trust in the output.

Lightning Round and Final Advice

Host: Let's jump into the lightning round. What are two books you frequently recommend?

Tara Seshan: Barbarian Days by William Finnegan, which explores dedication and passion for a craft regardless of whether you achieve conventional mastery, and Leo Tolstoy's Anna Karenina, which reveals new layers of human nature and growth every time you revisit it.

Host: What is a recent movie or cultural work that inspired you?

Tara Seshan: Christopher Nolan's The Odyssey, which offers a compelling lens on technological transformation, and Akira Kurosawa's Rashomon, which demonstrates how radical creativity and timeless storytelling can flourish under tight production constraints.

Host: What is your guiding life motto?

Tara Seshan: Toni Morrison's four rules on work:

1. Whatever the work is, do it well, not for the boss, but for yourself.
2. You make the job; it doesn't make you.
3. Your real life is with your family.
4. You are not the work you do; you are the person that you are.

Host: Any final advice or calls to action for listeners?

Tara Seshan: Download the desktop app, try Work mode on the web and mobile, and experiment with building custom sites and visualizations. Kicking off an asynchronous cloud task on your phone and returning to find completed work is a great way to experience the transition to persistent AI coworkers.