AI’s third era: the rise of persistent AI coworkers | Tara Seshan (OpenAI’s product lead)
In a Nutshell
Tara Seshan describes AI's third era: persistent AI coworkers that act as teammates, handling work at higher abstraction levels while humans provide steering and direction. The key to building in this fast-moving market is staying 2-3 months ahead of model capabilities—neither building for current limits nor a year out—and operating through rapid empirical testing rather than theoretical planning. The competitive advantage shifts to human ambition, opinionation, and expression, as the gap between what AI enables and what people actually attempt is the biggest constraint on progress.
These notes were generated by AI and may contain inaccuracies.
The first era of AI products was chat. The second era involves products working with agents. The third era, which may come soon, is working with a persistent co-worker who is able to get things done with you.
There is an overhang between what AI is capable of and what we are actually doing with it. It is hard to understand what will emerge in the future. You fail if you build for where the models are now. You fail if you build for where you think the models will be in a year. Both outcomes are equally wrong. The only way to build is 2 to 3 months ahead.
Being prolific and empirical is way more important than being academic or theoretical. Rather than writing out some long reasoning doc, the approach is to get to something that can be tried out and tested with users as fast as possible. It feels like not only are we able to be more ambitious, we almost need to be more ambitious, which is not natural for a lot of people.
Elevating others' ambitions or reminding them of what is possible is a huge part of the product management role.
Tara Seshan has been at OpenAI for just about a year. In most places this would be a very short amount of time, but in AI time it is like a lifetime.
Many things about working at OpenAI felt familiar because she had worked at other high growth, high talent, high intensity, hyperscaling mode places before. The part that actually felt most surprising is that OpenAI is not founder-led in the traditional sense. Everyone inside the company, especially in their area, is in essence kind of a founder to some extent.
The level of top-down direction at OpenAI is extremely limited relative to places she worked prior. The distance between her and the market is very thin. She is doing everything it takes to get product market fit for her product akin to how a founder might.
She came into the company expecting that there was a treasure trove of OpenAI secret strategy. Instead, OpenAI is open. Every sort of thought that exists in terms of how the world should look, how product should be built, or how the model should operate very quickly becomes a part of the public product or a part of the public messaging.
So much of what OpenAI does immediately becomes something that users can touch and feel in the product, and that cycle is faster than anywhere else she has seen.
When a market is more static or slow-moving, there is the chance to do grand strategy work because it is more predictable. In payments, if you take this bet, your competitor might take this other bet. You can reason from first principles very rigorously through what all the next actions might be. The nature of that market mandates that you do that. Winners will think more rigorously than everybody else.
In the AI market, it is so hard to understand what is going to emerge in the future. It is very emergent, very fast changing, really dynamic. It is very important to stay tied to the research. Being prolific and empirical is way more important than being academic or theoretical.
Rather than writing out some long reasoning doc, almost like a PhD thesis of what the plan should be, the approach is to get to something that can be tried out and test with users as fast as possible. The switch from theoretical to empirical felt very jarring at first.
The thinking needed is being as pointed as possible about what the core hypothesis is. The hypothesis definition is the most important thing. Using the phrase from Shashir Rahul, the iigen question is the specific most important thing to test. Any other grand strategy concocted is not relevant.
The core of the PM role has always been about what is the most essential question you need to ask about your product. What is the thing that will determine whether your product works or does not work? How do you test that? How do you look at the results? How do you feed that back into a loop of refining your hypothesis and running it again?
This involves trying to understand users, trying to understand the market, trying to understand the actual technology being built, pulling those three things together to make the most sharp hypothesis possible, then making the test as fast and effective as possible. This has not only not changed, but has become the most important thing at the company to be able to do.
Engineers, data scientists, and designers are all thinking this way. Everyone has moved to focus their efforts on problem definition and the testing loop.
The future of work will look more like steering than rowing. There will be agents that will do a lot of the rowing, and the role increasingly becomes steering the ship in the right direction and pointing it in the right direction.
Steering might grow to higher and higher levels. The steering used to be at the level of writing a line of code. Now it is directing something more comprehensive, maybe at the goal level, maybe even higher. The steering will continue to go up layers of abstraction.
Some of steering is about what the data tells you, but a lot of it is about making an opinionated call. The power of intuition or positive determinism of what the future should be is underrated. Picturing the product looking a certain way not because the converse is not viable, but because of wanting the world to look like the direction being pushed.
If everybody has access to the same tools, the thing that will separate people is the human element. The person basically. The human brain is the almost unfair advantage.
This reminds of fashion. There are functional clothes that everybody can wear, but so much about what you wear is about what statement you want to make about your individuality or how you want to reflect to the rest of the world. What makes that compelling is how it contrasts with other people's expression.
Patrick Hollen or John Collin has a statement about software: software is not like real estate where you put money in and get value out. It is a little bit more like film making where you can put a lot of money into a film but that does not guarantee that the film is successful or good. There is some statement or opinionation and artistry that goes along with it.
Marty Kagan is big on the idea that when you have an idea for a product or feature, rarely is that idea the thing that ends up being. There is a whole process to go through to figure out what it actually should be. You need to go through that process as a human to understand what it really is and what people actually want. It is never going to be like, okay, got it, go build this thing, I got it from the beginning.
Those loops are moving faster and faster. The ability to form intuitions, get the information needed to form those intuitions, and then use that with people and agents to put that into action is the key.
There are two aspects to the next shift. One is continuing to work with agents at higher and higher levels of abstraction. Letting the agent do more independently, coming in providing steering, then letting the agent continue to cook at higher orders of abstraction feels like the way people are increasingly thinking about agents.
Persistent agents feel like teammates and co-workers where you can work with them the way you might work with someone on your team. They do a whole bunch of work, you provide input, then they do work again. You sync up at different cadences, look at each other's in-progress work, and provide more feedback.
The second aspect is that a lot of work with agents has been one-on-one. That is potentially divorced from what colleagues are doing with their agents. The question is, as more work gets done with agents, shouldn't we be able to get work done with our agents together? What is the most natural interface to make that happen?
It feels like we are on the slow takeoff scenario where AI is slowly iterating. The models are incredibly smart, but a lot of the things that have enabled working with agents together or having agents take care of higher order abstraction things are about the intelligence and their ability to perform long-running tasks.
There are also very meat and potatoes tactical things that make this possible. Agents working locally are really convenient because they have access to all the data on your machine. To make an agent successful in the cloud, there is a ton of cloud infrastructure that has to be built.
Access to systems is critical. How can agents talk to all these third-party systems that have all of your data? Just like a colleague who you hire, if you lock them into a room and never give them access to Google Docs, Slack, or the company database, they would not be that useful. Similarly, a cloud agent that is similarly isolated will not be that effective.
A huge part of making agents useful is on the intelligence side, but a lot of it is also tactical like data access, cloud infrastructure, and reliability pieces that matter just as much for end effectiveness.
It feels like not only are we able to be more ambitious because of these AI tools, we almost need to be more ambitious, which is not natural for a lot of people. Everybody can now do all these easy things really easily. The easy stuff is super easy. The hard stuff is easy. The thing that separates people and companies now is just how ambitious they can be.
The people who are most effective at using AI tools do not simply use it to automate rote tasks but use it to expand the set of things that they are capable of doing. Back in the day, the unicorn person was someone who was a really thoughtful product sense person who also happened to be an engineer who may also have been a designer. That person was able to flatten the layers of translation needed between all these functions.
The most compelling thing is expanding the set of things that are within their range of possibilities so they can start realizing more of what is in their head into reality the way that someone who was previously a jack of all trades was able to do. We all have that superpower now.
You can spin up a set of designs on something, go build an initial prototype, figure out the right pricing model, model out all the scenarios. The set of possibilities have widened dramatically. This means having the ability to be more of an author as you try to get something done and realize your vision to higher fidelity.
The hardest part about doing this is simply expanding your thinking. The capabilities have expanded so dramatically. It is really expanding your thinking of what is possible in an unreasonably short time frame. Patrick Hollen has on his website patrickzen.com/fast all of these projects that were unreasonably ambitious that were executed in a really short time period.
AI has dramatically expanded what's possible for individuals. People can now request summaries of complicated texts or books instantly, execute tasks that were previously impossible, and create CAD models of ideas that were beyond their reach. This represents a fundamental shift where capabilities that were truly beyond individual reach are now accessible.
The hardest part is remembering to even try asking AI for help. Tyler Cowan's observation on his site is that most people underrate the impact of going to someone else and asking for the more ambitious version of what they're doing, or whether they could try it faster or at 10x bigger scale.
A huge part of the product management role now involves elevating others' ambitions and reminding them of what's possible. When teams propose timelines or approaches, PMs should ask whether the possibility ceiling is meaningfully higher and whether they should be more ambitious about what's being attempted.
"Is this maximally accelerated?" is a core OpenAI meme, represented with an emoji in Slack. The second meme that Andrew Embraino and Tara love asking the team is "Are you mainlining it yet?" - meaning are you using this product all day every day to get your work done.
These three memes combine to form the core of product development: being as ambitious as possible about scope and scale, moving as fast as possible, and using the product extensively while bringing all your taste to bear on whether it works and whether people want it.
"Feeling the AGI" is another important cultural meme - being conscious of AGI coming. A huge part of what attracts people to the company is believing in the mission of AGI being beneficial and trying to do whatever it takes to make that possible, both the realization of AGI and ensuring it's beneficial for humanity.
A constant refrain is building for where models will be in two to three months, not where they are now or where they'll be in a year. Building for current capabilities means failing because the models will quickly outpace the product. Building for a year out means failing because you're too early. The only way to build is two to three months ahead, with model capability as the center of the product, getting out of the way of the model in terms of product constructs.
Communicating tightly with research on where models are heading is incredibly important. The team has focused efforts on making models better at specific capabilities like coding or writing in particular ways, and product development needs to be tied as closely as possible to research's agenda and roadmap.
Kevin Wheel's quote that "this is the worst the models will ever be" captures the reality that models are improving so rapidly that current limitations are temporary.
The ChatGPT app currently has three modes: ChatGPT, Codex, with toggles between chat and work modes. The north star is eliminating the need for users to make decisions between these different options. Users should simply type their task and the system will pick the right harness and model automatically.
Picking between ChatGPT and Codex is a choice between staying in a development-oriented UI or having the same power in ChatGPT mode. Codex users should continue using Codex as they're not missing out on anything. ChatGPT users wanting new agent capabilities should probably be in ChatGPT mode.
In work mode, Codex operates under the covers with some coding UI elements removed. Work mode enables generating complex financial models and other sophisticated tasks. The corporate finance team uses work mode to execute tasks that were previously manual or required deep expertise from one person, now becoming executable by the whole team.
Work mode and Codex mode have the same underlying power. The difference is purely at the UI level - what kind of chain of thought visibility and technical detail gets exposed. Codex users aren't missing out by not switching modes.
The goal is to merge all these interfaces so users don't have to make decisions. The current separation meets people where they are in terms of the products they use and their familiarity with concepts, enabling everyone to take advantage of working with agents that have transformed how developers work.
Balancing ChatGPT's 100 billion MAU with Codex and other new products presents unique challenges. One goal of launching work in ChatGPT web and desktop app was to bring agent power to the billion people already using ChatGPT.
The first era of AI products was chat. The second era is working with agents, primarily coding agents so far. The goal is to bring this to more domains like knowledge work. The third era coming soon involves working with persistent co-workers who can get things done collaboratively with other people.
In the pre-AI era, polish was king and getting every UI interaction completely right was more important than shipping early. Now, getting products in users' hands when there's conviction they're transformative is way better than perfect. Urgency and introduction matter more than perfection.
The team iterates quickly and responds to feedback. Getting products out, figuring out what's not working, how people are using them, and iterating quickly is the current model. There are things that can continue to iterate and get feedback prior to launching, but iterating as quickly as possible and listening to the right signals is key.
The phrase "before enlightenment carry wood, chop water; post enlightenment carry water, chop wood" applies to the Codex team. The team was always super user-focused with tight iteration loops, dogfooding and mainlining the app. The change was merely the market catching up - users on Twitter and externally started noticing what was happening.
The mode of operating hasn't changed internally. Everyone still constantly uses the app, and the three core questions remain: Are we elevating our mission sufficiently? Are we maximally accelerating progress? Are we mainlining it as much as possible?
At startups, there are very few boundaries around roles - everything and nothing is your responsibility, with ultimate accountability for success. This mentality has always been appealing, and now capability is catching up to that approach. What matters is having core accountability for whether the product is being used, whether people want it, whether it's high quality and effective.
Some questions arise about whether taking a more fluid approach to teamwork means losing the ability to polish craft. This remains an open question. Some pieces of craft are getting abstracted by models that can do tasks more effectively than individuals. The craft moves from executing specific tasks to applying expertise to other parts of the product or discipline.
Humans will continue to be most valuable as entities of accountability - ultimately owning the outcome and whether the end product was high quality and achieved what was intended. This makes particular sense in highly regulated industries or places requiring direct human interface.
The human brain is also really valuable for expression. Software is not like real estate - it's more like a film where spending money doesn't guarantee a great outcome. There's artistry, opinionation, and expression in building software where there's authorship by a person.
Even as AI capabilities expand, choosing what to build and how it feels remains fundamentally human. The human brain continues to be valuable in how we care for each other and relate to one another. The part of work that involves talking to other people on your team, collectively figuring out how to be enthusiastic about an area, how you learn and work together, and how you elevate each other's ambitions feels like a human thing to do. Those pieces feel incredibly human regardless of what happens with the models.
There's an overhang between what AI is capable of and what we're actually doing with it. One of the biggest gaps is people asking what they should do with AI. There are two buckets to consider: how the PM job has changed most thanks to AI, and super interesting creative uses of AI that people might not have thought about.
One of the most exciting ways to use AI at work is building sites all the time. Sites is a product that lets you build a site as a presentational artifact, but also for literally anything. Examples include building a site for the team as a game using the database functionality, and building a site for a backpacking trip that tracked elevation and had everyone input their food. Sites realized the dream of malleable personal software that Alan Kay flagged in the 1960s about the true personal computer having personal software. With a site, you literally prompt it to build the exact tool needed and it does it. Sites are shareable, can auto-update, can use internal data to build dashboards with metrics, and serve as a more dynamic surface for presentation than slide decks.
To use sites, you tell Codex to create a site for something like a mafia game for your team, and it will just do it. It knows what sites are, hosts them for you, and you can choose whether they're public, with your team, or private to you. The easy reach of building a site has changed what day-to-day work looks like, replacing creating lots of artifacts like docs and sheets with making sites all the time. You can do this through Work, through Codex, through the web, or on mobile.
Another useful feature is slash visualize in Codex. You can ask it to visualize chat usage until now, and it will pull in all the things you've done and create an amazing visualization. This addresses the infinite number of times people think about how to not only pull in charts and data but present them in a way that's understandable and useful for the story they're trying to tell.
There are two types of writing at work: writing as thinking and writing as reporting. Writing as thinking involves writing a brief about why to build a certain product, why to take a certain strategy, or a spicy take. Writing as reporting includes summarizing team status, plans for launches, or announcements. Writing as reporting can be happily automated using models, but writing as thinking should never be automated. The act of outlining something, turning it into prose, cutting and editing it, and continuing to iterate is one of the most important steps for getting ideas in line.
Briefs are created by going into a hole to write about a new idea or product, spending time refining it, shopping it around with people to have them attack the ideas and poke holes, then taking it to the next person and doing the same. At Stripe, this was done many times to kick off new product areas, suggest big changes in direction, or analyze problems and suggest paths forward. Stripe is incredibly oriented as a writing culture, and a brief can go viral inside the company. Writing as thinking is really prized there.
At OpenAI, writing as thinking still happens, but the shareable artifact is not really a long doc or proof of work. Times have changed, and a long doc is no longer a signal that you thought through something because you can easily produce a long doc that indicates you haven't. The biggest change experienced personally is moving from thinking in a document and then translating that into a presentational artifact to being way more on mocks not docs, or prototypes not docs. If there's something people can try and interact with, or even better, results from an A/B test, that's a way better communication tool than the doc itself. Hundreds of docs are still written, but now they're done for personal use rather than for other people.
A manager advised writing a doc to 70% completion and then taking it to people needed for buy-in to get it from 70% to 100%. Very few great people want to interact with a perfectly polished finished idea because their new ideas just bounce off of it, versus something with more crags and rough edges that they can polish together. Bringing people into the process where a doc is an underlying artifact is one of the best ways to collaborate.
The writing as thinking discipline is one of the main pieces employed to avoid overly atrophying thinking abilities. All writing as reporting should be outsourced to the model as much as possible, but writing as thinking must be done personally. There's a personal belief that if someone is going to read a document, it must be read first that number of times. This also applies to meetings: if calling a meeting with a set of people, the collective amount of time people will spend in that meeting must be prepped before the meeting. When keeping thinking sharp, writing for the document is done personally first, investing the collective amount of time expected for people to read it.
Start and end personally with a piece of writing. AI might be used in the middle to research specific elements, drop in some data, go pull some data, or help push back on some ideas, but starting and ending personally with writing doesn't deteriorate thinking.
Sutter Hill is an iconic firm that is intentionally a very illegible firm. The website shows nothing because it's a firm that doesn't operate loudly and tries to operate as under the radar and modestly as possible, yet is responsible for some of the most iconic successes Silicon Valley has seen. They have an unusual incubation model started by Mike Spiser that has rolled out success after success. The thing most iconic about Sutter Hill is that people look at finding product market fit or building tens of billion dollar companies as a dark art involving luck and chance, yet Mike Spiser has done it multiple times, showing there's clearly a way to do it and a roadmap for making that possible. There's a playbook inside the firm, and they've figured out how to be right a lot in calling shots, making bets, and the daily compounding things that create a successful company.
Recruiting at Sutter Hill is unparalleled. They have a secret tool called Reticle where they have a map of everyone they've interacted with and the 10 best people that those people have interacted with, which helps them be very effective. Sutter Hill was joined because the career has been about finding product market fit as many times as possible, whether as a founder, starting new products at Stripe, or joining a startup like Watershed. Sutter Hill is a place where they've figured out how to find product market fit on B2B products.
One surprising thing learned at Sutter Hill is that product market fit is important but product marketing fit was really underrated. The way you talk about the product and market it can precede even building the product. It should probably come from bringing together understanding the technology deeply and understanding the enterprise sales process, and that product marketing fit narrative and positioning is actually even before you build a product experience the right thing to test. The approach is to pitch 100 people, figure out how to refine that pitch as much as possible, get the marketing narrative of why this thing is transformative right, and then only commit to the product shape. Mike Spiser is unbeatable at this art. Previously, PMM work was underrated as just the glue between functions, but it was realized how transformative that work done excellently is to a company's outcome and can be the element that makes a company successful.
Knowledge work and coding are fundamentally different. Coding is so output-oriented that when asked to do a coding task, you can verify whether it did the task correctly or well via tests and trying it out. There's a way to validate based on the output. Knowledge work is different because you can't simply look at the deck in the end and see the numbers and believe that. The process, inputs, reasoning, and how it went along the way must be considered. A lot of work has been done and must continue to be done to adapt the product to knowledge work, which means way more focus on making ChatGPT a collaborator, allowing visibility of all the in-progress work, seeing citations and inputs, and helping go on the journey with the model to get to that end output. This shows up in the UX of the product and should show up in things like reasoning and chain of thought, such as seeing more citations along the way for how it got to the end state.
The surface of a thread which is so suited to coding may not be the right place to see all of that for knowledge work. There are big important product questions. As human collaborators are brought into work, how to make the model more of a collaborator as things get done together must also be considered. Context is important: does it have the context it needs to do the thing wanted, does it know, can it see your email, can it see all your Confluence docs.
Barbarian Days by William Finnegan is recommended because it's about a New Yorker reporter who fell in love with surfing as his passion. The takeaway is that one can be deeply passionate and dedicated and have something be your life purpose without being good at it. It's about the art of falling in love with surfing and striving for excellence and perfection whilst knowing you'll never reach it.
Anna Karenina is loved because it's a book of layers. A huge part of what will have to be done in this new era is transform ourselves or take ourselves on a journey to do different things than what we were used to. When read at 13, basically the plot was understood. When read at 17, the European history dynamics and class warfare were understood. When read at 30, it was seen as a story about a woman and humans. It reminds that growth is possible to look at the same thing through multiple different lenses as you continue to grow.
Simon Hazel's Substack does a slow read of important books chapter by chapter, which is the only way to read something like The Power Broker or War and Peace. There's a 99 Press invisible book club breakdown of The Power Broker with 13 episodes, an hour or two each, going through a couple chapters at a time with special guests like Pete Buttigieg and AOC and folks that lived in that area.
Tara Seshan describes Christopher Nolan's The Odyssey as an incredible film about AI and Nolan's view on how AI transforms society. She highly recommends watching it, noting Nolan has bridged artistry and commercial success in a way no other modern director has achieved.
Seshan also recently watched Akira Kurosawa's Rashomon, the film that pioneered the technique of telling a story through multiple people's perspectives where the truth remains unknown. She reflects that this was made in the 1950s in black and white with a guy holding the camera, yet remains perfect and tastefully innovative. She notes having a hundred times the power and tools on her iPhone that Kurosawa had, and asks what her excuse is for not elevating her ambitions and making better stuff.
Seshan's favorite AI product outside OpenAI is products her friends make for her. She is a huge fan of the cozy software movement where people make software tools for five friends to use together. She mentions a friend named Sebastian who made an AI app that turns anything into a podcast and puts it in a little podcast app, and also made a private social network for their friends called GATS.
GATS is a private Twitter-like platform for a small group of friends. Seshan learns the most interesting things on that product. She describes it as exactly what the future should be - people making software that exactly meets their and their friends' needs.
The podcast app Sebastian made drops podcasts into your Apple podcast feed with new episodes of things you want to read. It works by adding actual podcasts to your existing podcast feed rather than creating a separate app.
Seshan cites Tony Morrison's four takes on work 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.
- You are not the work you do, you are the person that you are.
She has this pinned to her Twitter profile so she sees it every time she opens the app.
The Teal Fellowship was an inflection point in Seshan's life. She wouldn't be where she is without it. It was a moment where someone elevated her ambitions and said she could do something different than the path she was on. She is eternally grateful for that intervention.
One Teal Fellow Seshan works with is Ari Weinstein, who founded a company called Sky that was acquired by OpenAI. Prior to that, he founded a company acquired by Apple. Ari is one of the most creative thinkers she's seen and is the expert on cool things you can do on a Mac. Ari leads computer use stuff at OpenAI and his creativity, joy in what he does, and love of craft inspires her.
The Teal Fellowship gives people $100,000 to not go to college and instead build something they want. About 20 people participate each year as it's the 20 under 20 program. Seshan was part of the batch that had their entire fellowship documented as a CNBC documentary, including her pitch on stage presenting her idea, which is available on YouTube.
Dylan Field is one of the most amazing and successful people from Seshan's fellowship batch. She describes him as an incredible talent and very kind person.
Seshan recommends using the ChatGPT desktop app, trying it on web, and toggling to the work feature. She suggests asking it to build a site about yourself or create a visualization block of your ChatGPT usage as ways to experience the power intimately.
The most transformative experience is using ChatGPT on mobile. Users can go on a long subway or MUN ride with no service, and when they emerge, the work is completed. This cloud-based work feature runs in the cloud rather than locally, keeping work going until done, then allowing users to chat with results. This is described as massively powerful despite seeming simple.
Keep Lenny's Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
