20:45Tara Seshan： 用 AI 工具最有成效的人，不只是把重复性的活自动化，而是扩展了自己能做成的事情的范围。在 AI 之前，招聘里的「独角兽」是那种有产品感、会思考、同时恰好是工程师、也许还是设计师的人，因为他们压平了职能之间一层层的翻译，能自己很快做出或构想出一个东西，再和团队一起推进。现在我们都有了这种超能力。我可以随手出设计稿，搭一个初版原型，想出一套定价模型，把各种情景都算一遍。可能性的范围大大拓宽了，这意味着我可以更像一个作者，以更高的保真度实现我的构想。你的野心不再受限于你自己能执行什么、能讲清楚什么。最难的只是在一个短得不合理的时间里，扩展你对「什么可能」的想象。Patrick Collison 有一个页面，列了很多野心大得不合理、却在极短时间内做成的项目。现在让我觉得了不起的是，这些项目全都发生在这些工具出现之前。如果用过去的能力都做得到，那有了 AI 给我们的东西，这类项目难道不该出现指数级的增长吗？
59:43Lenny Rachitsky： 所以：想法、简报、计划，自己作为人来写。不从 AI 开始，甚至不用它来改进文字。
59:57Tara Seshan： 至少对我来说：开头我自己写，结尾我自己写。中间我可能用 AI 去研究某些具体部分、拉一些数据，或者反驳某些想法，但一篇东西的开头和结尾由我自己来，这样不会侵蚀我的思考。
1:00:22 · Sutter Hill，和「产品营销契合」
1:00:20Lenny Rachitsky： 你的职业经历里有一步很少见：在 Sutter Hill Ventures 做驻场创业者。这家机构自己孵化公司，Snowflake 就是一个例子。你学到了什么？
1:00:44Tara Seshan： Sutter Hill 是一家标志性的、刻意让外人看不懂的机构，它的网站上什么都没有。它尽可能安静、低调地运作，却是硅谷一些最标志性的成功背后的推手，靠的是 Mike Speiser 开创的一种不寻常的孵化模式。人们把找到产品市场契合、或者建一家几百亿美元的公司，当成一门靠运气和偶然的玄学，可 Mike 做成了好几次。显然有一套可以重复的打法：在下判断时一次次押对，以及那些每天复利、成就一家公司的事，比如怎么搭企业销售、怎么给产品定位、怎么组创始团队。他们的招聘无人能及；他们有个叫 Reticle 的工具，画出了他们打过交道的每个人，以及这些人合作过的最好的十个人。我的职业生涯就是尽可能多次地找到产品市场契合，而 Sutter Hill 已经摸清了怎么为 B2B 产品做到这一点，所以我想向他们学。
OpenAI's product lead lays out AI's three eras clearly, and explains from the product front line why knowledge work is hard to verify.
Part 1 of 10 · 1:17
Build it and test it; don't write a thesis
At OpenAI everyone acts like a founder and ideas quickly become public product. In a market this fast, grand strategy is useless; what matters is a sharp core hypothesis tested with users as fast as possible.
At OpenAI everyone acts like a founder and ideas quickly become public product. In a market this fast, grand strategy is useless; what matters is a sharp core hypothesis tested with users as fast as possible. Read this part →
Agents row and people steer, with the steering moving up layer by layer; but direction comes from opinionated judgement. Software is more like filmmaking than real estate. Read this part →
Treat agents as teammates: they work, people give input, they work again; after that, groups steering shared agents together. Data access, cloud infrastructure and reliability matter as much as intelligence. Read this part →
People who use AI best don't just skip chores; they widen what they can get done, and the hardest part is imagining what's possible. The new PM job is asking for the more ambitious version. Read this part →
Three mottos: is this maximally accelerated, are you mainlining it, feel the AGI. Build for today's model and fail; build for next year's and fail too. Stay close to the research roadmap. Read this part →
Work mode is Codex underneath with a different interface, so a billion users never have to pick a mode. The first era is chat, the second agents, the third the persistent coworker. Read this part →
Codex's reputation flipped because the team used it daily and iterated fast. Lines between engineers, designers and PMs are fading; whether the PM craft erodes, she admits she has no answer. Read this part →
Three places: accountability for outcomes, authorial expression, and caring for one another. She builds sites as everyday tools and uses /visualize to turn data into charts. Read this part →
Hand writing-to-report to the models, but write to think yourself; persuade with prototypes and test results, not long docs. From Sutter Hill: pitch 100 people and get the story right before fixing the product. Read this part →
Code can be checked with tests; knowledge work can't be trusted on a final “90% success”. The process, inputs and reasoning must be visible, so the product lays out work in progress and citations. A lightning round closes it. Read this part →
Indigo's conclusion
Two layers of first-hand value: from OpenAI's product front line, confirmation that taste and opinionated judgement are what humans keep; and, more important, she makes “verification can't be skipped” the central problem of product design.
How to read this An interview with OpenAI's product lead, and she is clearly speaking for her own products; to some degree the whole conversation promotes ChatGPT's work mode and Codex. But her observations about the craft are real insight: steer rather than row, write to think, ambition as the new scarcity, knowledge work that is hard to verify. Separate the product pitch from the practitioner's judgement, and discount the product vision and the timelines.
What to remember
Three eras: chat, agents, the persistent coworker. People move from rowing to steering, the steering keeps moving up, but direction and opinionated judgement stay human.
Knowledge work isn't coding: there are no tests to run and a final number can't simply be trusted; process, inputs and reasoning must be visible, so the product lays out work in progress and citations.
The “persistent coworker” is persistence in the product experience, not a model learning on the job, and it rests on data access and cloud infrastructure. The part to discount most.
Never hand writing-to-think to AI; hand off writing-to-report as far as possible. Persuade with prototypes, not long docs, and write the beginning and the end yourself.
Ambition is the new scarcity: easy things became extremely easy, so the gap is ambition; raising other people's ambition is the new core of the PM job.
What would change my mind
knowledge work gains a cheap mechanical check of its own, and verification can be skipped after all.
How to read this
An interview with OpenAI's product lead, and she is clearly speaking for her own products; to some degree the whole conversation promotes ChatGPT's work mode and Codex. But her observations about the craft are real insight: steer rather than row, write to think, ambition as the new scarcity, knowledge work that is hard to verify. Separate the product pitch from the practitioner's judgement, and discount the product vision and the timelines.
At OpenAI everyone acts like a founder and ideas quickly become public product. In a market this fast, grand strategy is useless; what matters is a sharp core hypothesis tested with users as fast as possible.
01:15 · What it's like inside a frontier lab
1:17Lenny Rachitsky: Today my guest is Tara Seshan. Tara leads product for both Codex and ChatGPT work at OpenAI, which I believe is the fastest-growing and arguably most important AI product for knowledge workers today. Before OpenAI she spent six years at Stripe, where she joined as one of the first five product managers, led product at Watershed, and was a founder and a Thiel Fellow. You've been at OpenAI about a year, which in AI time is a lifetime. What has most surprised you about working there, good and bad?
2:57Tara Seshan: A lot felt familiar, because I'd worked at other high-growth, high-talent, high-intensity places: amazing colleagues, very high urgency. What surprised me most is that every company I worked at before was founder-led, and OpenAI is founders-led: everyone inside the company, especially in their area, is in essence a founder. The level of top-down direction is extremely limited compared with places I've worked. I came from a founding journey, so I could keep feeling like the founder of my product area, and the distance between me and the market is very thin. You're doing whatever it takes to get product-market fit, like a founder would.
4:26Tara Seshan: The other side is that I came in expecting a treasure trove of secret OpenAI strategy, the way at past companies you find the payments bible that explains how everyone thinks. Actually, OpenAI is open. Every idea about how the world should look, how products should be built or how the model should behave very quickly becomes part of the public product or the public messaging. That changed my operating mode. What inspires me is that so much of what OpenAI does immediately becomes something users can touch in the product, and that cycle is faster than anywhere I've seen.
06:44 · From theoretical to empirical
6:42Lenny Rachitsky: You've been a PM and a PM leader at a lot of places. What do you lose in this new world?
6:49Tara Seshan: In a more static or slow-moving market you can do grand-strategy work, because it's more predictable. Payments is dynamic but established: if I take this bet, my competitor might take that one, and you reason rigorously from first principles through the next moves. That market mandates it; winners think more rigorously, and if you don't, it shows up as carelessness, because your decisions could have been predicted. In this market it's very hard to know what will emerge. It's emergent, fast-changing, and above all you have to stay tied to the research. So being prolific and empirical matters far more than being academic or theoretical. Instead of writing a long reasoning doc, almost a PhD thesis on the plan for the next stretch of time, it's: how do I get to something I can test with users as fast as possible? That switch was jarring. I wondered whether I was doing my due diligence. But you have to try things and learn as much as possible, which means the thinking you do must be as pointed as possible about your core hypothesis. Defining that hypothesis is the most important thing, what Shishir Mehrotra calls the eigenquestion; any other grand strategy you concoct is irrelevant.
9:07Lenny Rachitsky: So that's the part of the PM role that isn't changing, and it's becoming more important.
9:20Tara Seshan: There are many trappings around the PM role: running execution on time, writing docs and presentations. But the core has always been: what is the most essential question about your product, the thing that determines whether it works? How do you test it, read the results, and feed them back into a loop that refines the hypothesis? That involves understanding users, the market and the technology, combining them into the sharpest hypothesis you can, and making the test as fast and effective as possible. Not only has that not changed, it has become the most important thing at the company. EMs, engineers, data scientists and designers all think this way now, focused on problem definition and the testing loop. Many of the other trappings have fallen away.
02
Steering, not rowing
Agents row and people steer, with the steering moving up layer by layer; but direction comes from opinionated judgement. Software is more like filmmaking than real estate.
10:51 · Steering, not rowing
10:56Lenny Rachitsky: Loops were very hot on Twitter a few weeks ago: tell the AI what success looks like and let it go build until it gets there. Do you think loops will expand from software engineering to product management and all knowledge work?
11:24Tara Seshan: I think the future of work will increasingly look like steering rather than rowing. Agents will do a lot of the rowing, and your role becomes steering the ship and pointing it in the right direction. That steering may move higher and higher: it used to be writing a line of code and pressing tab, then directing something more comprehensive, maybe at the goal level, maybe higher still. It will keep moving up layers of abstraction. But ultimately it's still on a person to decide which direction to point it and, given feedback and data, where to take it next. Some of steering is what the data tells you, but a lot is making an opinionated call. We underrate intuition, even a kind of positive determinism about what we want the future to be: I want the product to look this way, not because the alternative isn't equally viable, but because this is the direction I'm pushing the world. That, at least for now, has to come from a person. Loops are great, and agents running in ever larger loops doing more of the rowing is great, but you still need to steer. Work will also look like steering together with other people over a group of agents you share.
13:25Lenny Rachitsky: If everybody has access to the same tools, what separates you is the human, basically. Otherwise we're all building the same thing, and the almost unfair advantage is the human brain.
13:42Tara Seshan: It reminds me of fashion. There are functional clothes everyone can wear, but so much of what you wear is a statement about your individuality, and it's compelling partly because it contrasts with other people's expression. A lot of the products we build feel similarly opinionated and artistic. Patrick Collison, or maybe John Collison, has a nice line: software is not like real estate, where you put money in and get value out. It's more like filmmaking. You can put a lot of money into a film and that doesn't guarantee it's good. There's an auteur statement, some opinionation and artistry, and that relies on you or your team having something interesting to say about your product.
14:59Lenny Rachitsky: Marty Cagan says your first idea for a product is rarely what it ends up being; there's a whole process to figure out what it should actually be, and you have to go through it as a human.
15:21Tara Seshan: For sure. Those loops are moving faster and faster, so your ability to form intuitions, get the information you need to form them, and then act on them with people and agents is the key.
03
Next: the persistent AI coworker
Treat agents as teammates: they work, people give input, they work again; after that, groups steering shared agents together. Data access, cloud infrastructure and reliability matter as much as intelligence.
15:46 · Next: the persistent AI coworker
15:48Lenny Rachitsky: You work around the most AI-forward people in the world. How are people working inside OpenAI that will become normal for all of us in the next three to six months?
16:08Tara Seshan: Two things. One is working with agents at higher and higher levels of abstraction: letting the agent do more independently, coming in to steer, and letting the agent keep cooking. People increasingly think of agents as persistent, as teammates and coworkers. You work with them the way I work with someone on my team: they do a lot of work, I give input, they work again, we sync at different cadences, look at each other's work in progress and give more feedback. That coworker model is where things are going, a much more natural interface for working with agents, and we already see a lot of it internally.
16:58Tara Seshan: The second is that my work with agents so far has been one-on-one: me and my agent, maybe spawning some sub-agents, but potentially divorced from what my colleagues do with theirs. There was a time when everyone internally was sending screenshots of their Codex threads to each other on Slack: here's how I got to this number. That's not the most natural way to collaborate. As more work gets done with agents, shouldn't we be able to work with our agents together? What's the most natural interface for that? That's what we're thinking about. Ideally work feels like a multiplayer game, all of us steering our agents together while they take care of more of the tactical rowing.
18:15Lenny Rachitsky: It's been a slow progression of trust: okay, go work longer, take on more. Everyone feared a fast takeoff, and it feels like we're in the slow-takeoff scenario, which is good. It doesn't feel slow, but we're not dealing with some 300-IQ AI.
18:48Tara Seshan: The models are incredibly smart, and what lets us hand agents higher-order work is certainly about intelligence, their ability to run long tasks and stay on task. But there are also very meat-and-potatoes things. Agents working locally are convenient because they can access all the data on your machine. Making an agent succeed in the cloud takes a ton of cloud infrastructure. And access to your systems: a colleague you hire and lock in a room with no access to Google Docs, Slack or the company database wouldn't be very useful, and neither is an isolated cloud agent. So a huge part of making agents useful is intelligence, but a lot is tactical: data access, cloud infrastructure, reliability. They feel much more prosaic than the big intelligence questions, but they matter just as much for how effective the agents end up being.
04
Ambition is the new scarcity
People who use AI best don't just skip chores; they widen what they can get done, and the hardest part is imagining what's possible. The new PM job is asking for the more ambitious version.
20:05 · Ambition is the new scarcity
20:07Lenny Rachitsky: Ambition keeps coming up on this podcast. Not only can we be more ambitious with these tools, we almost need to be, which isn't natural for many people. The easy stuff is super easy now, and what separates people and companies is how ambitious they can be.
20:45Tara Seshan: The people most effective with AI tools don't just automate rote tasks; they expand the set of things they're capable of. Before AI, the unicorn hire was a thoughtful product-sense person who also happened to be an engineer and maybe a designer, because they flattened the layers of translation between functions and could build or ideate something quickly themselves, then work with a team on it. Now we all have that superpower. I can spin up designs, build an initial prototype, figure out a pricing model and model all the scenarios. The set of possibilities has widened dramatically, which means I can be more of an auteur and realize my vision at higher fidelity. Your ambitions are no longer limited by what you can execute or communicate yourself. The hardest part is simply expanding your thinking about what's possible in an unreasonably short time. Patrick Collison has a page of unreasonably ambitious projects executed very fast, and what's remarkable now is that they all happened before these tools. If those were possible with the capabilities we used to have, shouldn't we see an exponential increase in such projects with what AI gives us?
23:57Lenny Rachitsky: The hardest part is remembering to even try, to say: let me see if Codex can do this. It's a new habit we have to build.
24:11Tara Seshan: Tyler Cowen says most people underrate the impact of going to someone and asking: what's the more ambitious version of what you're doing? Couldn't you try this faster, or at 10x the scale? When I think about what PMs can do that's incredibly effective now, elevating others' ambitions and reminding them what's possible is a huge part of the role. When someone says, here's how we'll get this done, here's the timeline, here's a first version, part of your job is to say: isn't the ceiling meaningfully higher? Shouldn't we be more ambitious? Couldn't we try this faster?
05
Build for the model two to three months out
Three mottos: is this maximally accelerated, are you mainlining it, feel the AGI. Build for today's model and fail; build for next year's and fail too. Stay close to the research roadmap.
25:07 · Three memes, and building for two to three months out
25:18Lenny Rachitsky: Nick Turley had an internal meme: is this maximally accelerated?
25:24Tara Seshan: "Is this maximally accelerated?" is totally an OpenAI meme. Another one Andrew Ambrosino and I love asking the team is "are you mainlining it yet?": are you using the product all day, every day to get your own work done? Combine that with asking whether we're being as ambitious as possible about scope and scale. Is this maximally accelerated, are we moving as fast as possible, and are you mainlining it, bringing all your taste to bear on whether it works and people really want it, tightening that feedback loop as much as possible. Those are the three memes of product development we need to spread. Another important one is feeling the AGI, being conscious that AGI is coming. What brings most people here is believing in the mission of AGI being beneficial and doing whatever it takes to make that happen.
27:17Tara Seshan: And in building products, a constant refrain in the back of my mind is: are we building for where the models will be in two to three months? You fail if you build for where the models are now. You fail if you build for where you think they'll be in a year. Both outcomes are equally wrong. If you're too early, you're wrong; if you build something overly focused on a past model's capabilities, you're entirely wrong. The only way to build is two to three months out, with the belief that models will get way better: model capability is the center of the product, and I need to get my product constructs out of the model's way.
28:05Lenny Rachitsky: How do you know what two or three months out looks like on an exponential? Gut feeling? Do the researchers give you a sense?
28:18Tara Seshan: Communicating tightly with research about where things are going is incredibly important. It isn't a black box: you know the teams are focused on making the model better at coding or writing in specific ways, so you tie product development as closely as possible to research's agenda and roadmap.
29:01Lenny Rachitsky: Kevin Weil said on this podcast that this is the worst the models will ever be. It's almost a cliché now, but it's true, and absurd.
Work mode is Codex underneath with a different interface, so a billion users never have to pick a mode. The first era is chat, the second agents, the third the persistent coworker.
29:28 · ChatGPT, Codex and work mode
29:36Lenny Rachitsky: I have ChatGPT open. There's a dropdown with ChatGPT and Codex, and a toggle between chat and work. What's going on, and where does this go?
29:50Tara Seshan: Our north star is that users shouldn't have to choose between all these options. You go to the box and type your task, say, build an app that helps my podcast guests research before episodes, and it picks the right harness and the right model to get it done. The choice shouldn't be on users, who would otherwise have to understand not only what they're trying to do but the limits and capabilities of our products. For now, choosing between ChatGPT and Codex is about whether you want a more development-oriented UI or the same power in ChatGPT. If you're a Codex user, keep using Codex; you're not missing anything. In ChatGPT, chat mode is the chat you know and love, for conversation and search, with better models each time. Under the covers, work mode is Codex, with some of the coding UI removed. You won't see a worktree pop up, but it has the same power to get things done, like generating a really complex financial model. Our corporate finance team uses work mode for things that used to be manual or required one person's deep expertise, and now the whole team can do them.
32:20Lenny Rachitsky: Is anything in work mode different from Codex, harness tweaks, or is it the same thing with a different UI?
32:24Tara Seshan: It's at the UI level. Ask Codex to build a financial model to price your product or predict revenue for the next six months, and it will do as good a job as work mode. The difference is what UI you want to see while it works, and how much technical detail you want exposed. Truly, our north star is to merge all of this so users don't have to make these decisions. The separation is about meeting people where they are, in the products they use and the concepts they know, so everyone can take advantage of working with agents, which has entirely transformed how every developer works. We should do the same for knowledge work.
34:03Lenny Rachitsky: One of the hardest parts of your job must be balancing ChatGPT, with its billion users, maybe the most successful consumer product in history, against Codex and other new things.
34:36Tara Seshan: One goal of launching work in ChatGPT on the web and in the desktop app was to bring more and more of the agents' power to those billion people. If the first era of AI products was chat, the second is clearly working with agents, so far mostly coding agents, and we want to bring that to more domains like knowledge work. The product challenge is to make it natural and easy to adopt, not a decision people must explicitly make, and to decomplexify it so a billion consumers don't need to think about things like harnesses. And the third era, which might come soon, is working with a persistent coworker who can get things done with you, maybe together with other people.
36:17Tara Seshan: A lesson that contrasts with my earlier product experience: at previous companies polish was king. Getting every interaction exactly right mattered more than shipping early, because time didn't change the outcome much, so if every corner wasn't polished you might as well not ship. In this era, getting the product into users' hands when you have conviction it's transformative is way better than perfect. We have a lot to do to make it more usable, especially for people using it for consumer tasks rather than productivity. But done is better than perfect.
07
A reputation flip, and roles dissolving
Codex's reputation flipped because the team used it daily and iterated fast. Lines between engineers, designers and PMs are fading; whether the PM craft erodes, she admits she has no answer.
39:13 · Why Codex's reputation flipped
39:17Lenny Rachitsky: On Twitter there's been a vibe shift from Claude Code to Codex in the past few months. What changed internally?
40:09Tara Seshan: There's a saying: before enlightenment, chop wood, carry water; after enlightenment, chop wood, carry water. The team that got the Codex app up and running was super user-focused, with a tight iteration loop, and mainlined the app as much as possible to get everything right. People externally started to notice, but the team was always focused on users and iteration; to some extent the market caught up. That process hasn't changed. Everyone building it is a developer using it for development, constantly fixing their own problems and listening to other people's problems in the company and from users. The operating mode is the same as before: are we elevating our ambition, are we maximally accelerating, are we mainlining it? Full credit to the team.
41:25Lenny Rachitsky: What's interesting is how human that answer is. It's you, Andrew, Tibo, the team obsessed with the customer and the product. AI wasn't the answer; the humans made the difference.
41:40Tara Seshan: The team deserves full credit. Almost everyone on the desktop team acts like a founder and cares about every detail. When they notice something should be better, they go build it independently, and if it doesn't test well internally, if people don't use it or find it useful, they iterate before shipping it externally.
42:20 · Role boundaries are dissolving
42:29Lenny Rachitsky: Roles are overlapping: engineers doing PM work, PMs shipping prototypes, maybe to production. I hear from many people, what is my job now as a designer, as a marketer? Are you dealing with that?
42:49Tara Seshan: What I've always liked about startups is that there are very few boundaries around your role: everything and nothing is your responsibility, and ultimately you're accountable for success. Stripe was very much like that; anyone could do anything. Now capability is finally catching up to that mentality. What I care about is that someone has core accountability for whether the product is used, wanted, high quality and effective. Whether that's an engineer, a designer or a PM, someone is the directly responsible individual, and whatever work is needed, people pick it up based on affinity and capability. The flip side is that I love the craft of being a PM, which people like Shreyas, Marty Cagan and Shishir have espoused so well. Do I lose out on polishing my craft with this fluid approach? I truly don't have an answer. We're all experiencing it together: some parts of our craft are being abstracted away by models that do them very well, maybe better than individuals, and your craft moves from that specific task to some other part of the product or discipline. I'm still thinking about how to balance my love of being on a team and using these tools with my love of the craft.
45:31Lenny Rachitsky: It's unbelievable how different the engineering role is now. You used to write code all day, and that's no longer the job, and it happened so quickly. People mourn the flow state of writing code by hand.
45:52Tara Seshan: It's a tough transition. Some people love it, some don't.
08
Where human brains still earn their keep
Three places: accountability for outcomes, authorial expression, and caring for one another. She builds sites as everyday tools and uses /visualize to turn data into charts.
46:02 · Where human brains stay valuable
46:03Lenny Rachitsky: Over the next couple of years, where do you think human brains will remain most valuable?
46:25Tara Seshan: First, as an entity of accountability. Who ultimately owns the outcome? You can think of the agent you work with as your report, but who owns the end product, whether it was high quality and did what you wanted, will remain a person, at least for now, especially in highly regulated industries or places that need a direct human interface. Second, expression. Software is not like real estate; it's more like film, where the greatest films aren't the ones with the biggest budgets. There's artistry, opinionation and expression in building software, a sense of authorship by a person or group, and what you choose to build and how it feels is such a human question. Third, how we care for each other and relate to one another. That part of my work has become more important than ever: talking to people on your team, figuring out together how to be enthusiastic about an area, how to learn and work together, how to elevate each other's ambitions. I can't predict what will happen with the models, but those feel incredibly human.
48:21 · How she uses AI herself
48:31Lenny Rachitsky: There's an overhang between what AI can do and what we actually do with it. How do you use AI in ways that might inspire people?
49:01Tara Seshan: I build sites all the time now. Sites is a really fun product: you can build a site in work as a presentational artifact, but I build them for literally anything. I built a game for the team, since sites have a database. I went backpacking recently and built a site of the route that tracked the elevation everywhere we were going, and everyone on the trip entered their food. Sites realize the dream of malleable personal software that Alan Kay described in the '60s: the true personal computer has personal software. People tried with tools like Notion and their blocks, but with a site it's literally a prompt: build me exactly the tool I need. They're shareable, they can auto-update, you can use internal data to build a dashboard, and rather than laboring over a slide deck, a site is a much more dynamic surface for presentation. You just say in Codex, create a site that's a mafia game for my team, and it does it and hosts it, public, shared with your team or private. It has changed my day-to-day, which used to mean creating lots of docs and sheets. You can do it in work, in Codex, on the web, on mobile, anywhere.
51:48Tara Seshan: The other thing I love is /visualize in Codex. Type "/visualize my ChatGPT usage" and it pulls in everything you've done and creates an amazing visualization. I've spent endless time thinking about how to pull in charts and data and present them in a way that's useful for the story I'm telling, and visualize makes that incredibly simple.
09
Writing to think is never automated
Hand writing-to-report to the models, but write to think yourself; persuade with prototypes and test results, not long docs. From Sutter Hill: pitch 100 people and get the story right before fixing the product.
52:54 · Writing as thinking, writing as reporting
52:44Lenny Rachitsky: Brie Wolfson, who knows you well, said a Tara brief is iconic. What makes your writing work, and any tips for people trying to get better at writing documents?
53:14Tara Seshan: I do two types of writing at work: writing as thinking and writing as reporting. Writing as thinking is a brief on why we should build a product or take a strategy, maybe a spicy take. Writing as reporting is summarizing what the team did this week or laying out the plan for a launch. Writing as reporting I happily automate; I use the models all the time to make it as simple as possible. Writing as thinking I will never automate. For me, outlining, turning it into prose, cutting, editing and iterating is one of the most important steps to get my ideas in line. Most people paint with a broad brush, either never using models for writing or always using them, and both are wrong. Use the models as much as possible for writing as reporting, and to the extent you think by writing, don't replace your thinking with them. My briefs came from going into a hole, writing a brief for a new idea, refining it for a long time, then shopping it around and having people attack it, poke holes and make it stronger, then taking it to the next person and doing the same. Stripe is a writing culture, one of the few places where a brief goes viral inside the company.
55:46Tara Seshan: At OpenAI I still write as thinking all the time, but the shareable artifact isn't a long doc as proof of work anymore, partly because a long doc no longer signals that you thought something through; you can easily produce a long doc that shows you haven't. The biggest, maybe most jarring change in my day-to-day: I used to think in a document, translate it into a presentational artifact, and that showed I'd thought through a problem and the team would move. Now it's mocks, not docs, or prototypes, not docs. Something people can try and interact with, or even better, results: we ran an A/B test, here's why we should go this way. That communicates far better than the doc. I still write hundreds of docs, but I write them for me.
57:12Lenny Rachitsky: I liked your tip of getting tons of feedback as you iterate.
57:33Tara Seshan: A former manager told me to write a doc to 70% and take it to the people whose buy-in you need to get it from 70% to 100%. I still do that all the time, because very few great people want to engage with a perfectly polished, finished idea. New ideas bounce off it. Something with rough edges they can polish with you brings them into the process.
58:08Lenny Rachitsky: How do you think about AI brain rot, over-relying on AI and losing our ability to read and write long documents?
58:29Tara Seshan: The writing-as-thinking discipline is one of the main ways I make sure I'm not atrophying my thinking. I outsource writing as reporting as much as possible, but writing as thinking I do myself. I also believe that if I'm going to make someone read my document, I should at least have read it that many times, and if I call a meeting, I should have prepared for the collective time people will spend in it. I don't rely on the model to polish my prose, which I don't think it does well, or especially to generate the first version. I do use it a lot for summarizing or translating content from one format to another.
59:43Lenny Rachitsky: So: write the idea, the brief, the plan yourself as a human. Don't start with AI, and don't even use it to improve the writing.
59:57Tara Seshan: At least for me: I start myself and I end myself. I might use AI in the middle to research specific elements, pull some data, or push back on some ideas, but I start and end a piece of writing myself, and that doesn't erode my thinking.
1:00:22 · Sutter Hill and product-marketing fit
1:00:20Lenny Rachitsky: You had an unusual career step: entrepreneur in residence at Sutter Hill Ventures, which incubates companies; Snowflake is an example. What did you learn?
1:00:44Tara Seshan: Sutter Hill is an iconic and intentionally illegible firm; its website shows nothing. It operates as quietly and modestly as possible, yet is responsible for some of Silicon Valley's most iconic successes through an unusual incubation model that Mike Speiser started. People treat finding product-market fit, or building a company worth tens of billions, as a dark art of luck and chance, yet Mike has done it multiple times. There's clearly a playbook for doing it repeatably, for being right a lot on calling shots and on the daily compounding things that create a successful company: how you set up enterprise sales, how you position the product, how you build the founding team. Their recruiting is unparalleled; they have a tool called Reticle that maps everyone they've interacted with and the ten best people those people have worked with. My career has been about finding product-market fit as many times as possible, and Sutter Hill has figured out how to do that for B2B products, so I wanted to learn from them.
1:03:06Tara Seshan: The surprising lesson: product-market fit matters, but I underrated product-marketing fit. How you talk about and market the product can come before building it. It should come from deeply understanding the technology and the enterprise sales process, and that narrative and positioning is the right thing to test even before the product experience. Pitch 100 people, refine the pitch, get the story of why this is transformative right, and only then commit to exactly what shape the product takes. I used to think of product marketing as glue between functions, and I learned how transformative it is when done excellently; it can be what makes a company succeed.
10
Knowledge work has no tests to run
Code can be checked with tests; knowledge work can't be trusted on a final “90% success”. The process, inputs and reasoning must be visible, so the product lays out work in progress and citations. A lightning round closes it.
1:05:06 · Knowledge work is fundamentally different from coding
1:05:09Tara Seshan: One thing we've been thinking about a lot, especially with ChatGPT work, is how knowledge work and coding are fundamentally different. A surprising lesson: coding is so output-oriented that when you ask for a coding task, you can verify whether it was done correctly with tests. You can try it and see if it works; there's a way to validate it from the output. Knowledge work is different. I can't simply look at the deck at the end, see "90% success", and believe it. I really need to think about the process, the inputs, the reasoning and how it got there. So in the product, a lot of our work is adapting to knowledge work: making ChatGPT more of a collaborator, letting you see all the work in progress, its citations and inputs, and taking you on the journey with the model so that in the end you know the output is right, good and useful. That shows up in the UX, but also in the reasoning and chain of thought. Should you see more citations along the way about how it reached that data? Is a thread, which suits coding so well, the right surface for knowledge work? There are so many big product questions. And as we bring human collaborators into your work, we also need to make the model more of a collaborator with you.
1:07:00Lenny Rachitsky: Picture an exec meeting where you pitch a plan: so much of it is showing the work you did to get there, step by step. So the AI needs to show you the same proof of work, whereas in engineering you don't need every architectural decision, just whether it passes the tests. And then there's context: can it see your email, your docs?
1:08:02 · Lightning round
1:07:45Lenny Rachitsky: What books do you recommend most?
1:08:08Tara Seshan: Barbarian Days by William Finnegan, about a New Yorker reporter who fell in love with surfing. What I took from it is that you can be deeply passionate and dedicated, and make something your life's purpose, without being good at it: striving for excellence while knowing you'll never reach it. And Anna Karenina. I've been rereading the classics. At 13 I understood the plot, at 17 the European history and class dynamics, at 30 that it's a story about a woman and about people. It reminds me that you can see the same thing through different lenses as you grow, which is the challenge ahead for all of us in our careers in this new era.
1:09:51Lenny Rachitsky: A favorite recent film or show?
1:11:34Tara Seshan: The Odyssey. My hot take is that it's about AI, Christopher Nolan's view of how AI transforms society. He has bridged artistry and commercial success like no other modern director. I also recently watched Rashomon, where Kurosawa pioneered telling a story through multiple perspectives so you never know what was true. It was made in the '50s, in black and white, with a man holding a camera, and it's so tasteful and innovative. I have a hundred times his tools in my iPhone, so what's my excuse for not elevating my ambitions and making better things?
1:13:38Lenny Rachitsky: A favorite AI product you've discovered recently, ideally not from OpenAI?
1:13:50Tara Seshan: Products my friends make for me, because now people can. I'm a huge fan of the cozy software movement, where you make tools for five of your friends and use them together. My friend Sebastian made an app that turns anything into a podcast and drops it into your podcast feed, and a private social network for our friends. People should make software that exactly meets their and their friends' needs.
1:15:14Lenny Rachitsky: A life motto you come back to?
1:15:23Tara Seshan: Toni Morrison's four points from her essay "The Work You Do, the Person You Are": 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; and you are not the work you do, you are the person you are.
Where Indigo landsFurther
Indigo's conclusion
Two layers of first-hand value: from OpenAI's product front line, confirmation that taste and opinionated judgement are what humans keep; and, more important, she makes “verification can't be skipped” the central problem of product design.
What to remember
Three eras: chat, agents, the persistent coworker. People move from rowing to steering, the steering keeps moving up, but direction and opinionated judgement stay human.
Knowledge work isn't coding: there are no tests to run and a final number can't simply be trusted; process, inputs and reasoning must be visible, so the product lays out work in progress and citations.
The “persistent coworker” is persistence in the product experience, not a model learning on the job, and it rests on data access and cloud infrastructure. The part to discount most.
Never hand writing-to-think to AI; hand off writing-to-report as far as possible. Persuade with prototypes, not long docs, and write the beginning and the end yourself.
Ambition is the new scarcity: easy things became extremely easy, so the gap is ambition; raising other people's ambition is the new core of the PM job.
Claims you can check later
Claim
Who
When we will know
How firm
The third era is “coming soon”: a persistent AI coworker that gets things done with users and their colleagues
Tara
Near future
First-hand product vision, partly speaking for her own products
Future work is steering rather than rowing, with the steering moving up layer by layer
Tara
Trend
First-hand judgement
ChatGPT's chat, Codex and work modes will merge, and users won't have to pick a model or a harness
Tara
Near future
First-hand product goal
Build products for the model two to three months out; this is the worst the models will ever be
Tara, Kevin Weil
Ongoing
First-hand method
We're in a slow-takeoff scenario, with trust built step by step, not a fast takeoff
Tara
Now
First-hand judgement
Back on the long-running theses
confirms
Verification can't be compressed: it becomes the bottleneck, the moat, the breaking point “Knowledge work isn't coding” confirms from the product side where this view holds: verification compresses where a mechanical check exists and not where it doesn't.
confirms
What you own is not the model: value moves up to what cannot be rented Taste, opinionated judgement and steering: testimony for this view from the product side inside a model company.
confirms
Software as media: the business outside the model's range “Software is like film, not real estate” nearly restates selling the taste and ways of working no one else has, and it comes from inside OpenAI.
adds to
AI job doomsday: from two-pole prophecy to task-level displacement Where human brains still earn their keep (accountability, expression, care) and dissolving role boundaries add inside testimony on how specific jobs change.
adds to
Ashwin Gopinath: the engine is commoditized, memory is the moat, the third wave is state Both describe a third wave on different axes: Ashwin's is state and memory in the technical foundation, Tara's is the product experience.
conflicts
The Dwarkesh RSI debate (Schulman, Millidge, O'Neill) The debate says from the engineering side that learning on the job breaks down in the details; Tara, from the product side, makes the persistent coworker sound already here.
What would change my mind
knowledge work gains a cheap mechanical check of its own, and verification can be skipped after all.
Finished. Indigo's take on this piece is in two places:
