AI era skills: Why cultivating agency matters more than job titles | Max Schoening (Notion)

[00:00] Before, it was very easy to always say, well, I will never be able to do this because insert skill issue. We're realizing that even if you have the skills at your fingertips, the thing that matters is agency. I don't think agency is very evenly distributed in the world. Do you have a piece of advice for someone that wants to develop this within themselves? I tell this to myself, if you drive Notion like it's stolen, one day you wake up and you realize the world is made up by people no smarter than you.

It just really awakens you to the idea that you can just change things. [00:30] a couple years ago. What's most changed? The first 10% of every project are now free. It takes almost no effort to now build the first version of a startup. Taste comes up a lot now. Taste actually means you're able to run a virtual machine in your head where given an idea, you can predict for a certain in-group whether they're going to like it or not. You just have to do reps. It's almost like training a model.

What do you think matters to building a successful product? All the great products have something tiny that is a superpower. One tiny core that is so exceptional [01:00] good. One of the biggest pitfalls is if you get into the loop of if I just add one more thing to the product it'll be finally great. That never works. Give this hot take on universal basic income. We already have universal basic income it's called knowledge work. [01:14] Today, my guest is Max Schoening. Max is a hard person to describe. He was a product manager at Google.

He ran the design team at Heroku. He was a design leader and an engineer at GitHub under Nat Friedman. He's also a two-time founder and is now head of product at Notion. He is one of the most successful AI forward product leaders out there. And as you'll soon see, one of the deepest thinkers on how AI impacts how we build and how we use software. Before we get into it, don't forget [01:44] com for an insane set of deals available exclusively to Lenny's newsletter subscribers. With that, I bring you Max Schoening.

[01:55] Max, welcome to the podcast. [01:57] - Thank you for having me. - I am so excited to have you here. I feel like there's this quote I think about when I think about you, [02:04] uh and you being on this podcast it comes from the bible [02:07] And just paraphrasing, the quote is, I was made for such a time as this. [02:12] I feel like there's all this talk about [02:14] Roles merging, designers becoming PMs, engineers, everyone's the same, the Venn diagram's collapsing. You've been that for a long time.

It's like hard to even describe what you are and what you've done. You've done all the things. [02:24] So I feel like you have such a unique insight into where things are heading. [02:28] I want to start with just kind of this broad question. [02:30] What have you seen about [02:32] where things are going for product teams, for product building, as AI becomes more powerful, as we integrate it more into our workflows. And I ask you this because [02:41] I've heard for so many people at Notion that you are the reason that [02:44] Designers are shipping code.

[02:46] PMs are shipping code. You're not just living in the future. You're like pushing the whole team and company to live in the future. [02:52] And so... [02:53] So coming back to the question, just like, what are you seeing about where things are going? What will change? Will people realize in the next few months, years that you're already seeing? Well, first of all, when you said a quote from the Bible, I was very curious where this was going to happen. It's the first time I've quoted the Bible on this podcast, I think.

[03:10] I wouldn't take credit for... [03:13] the designers at Notion and PMs at Notion now code. [03:17] I think that... [03:18] would have probably happened anyways. [03:20] But I can tell you the origin story of it, which is when I joined Notion, we were... [03:26] building a lot of chat interfaces and we were designing the chat interfaces in Figma. [03:31] and [03:32] There's this great talk by Brett Victor [03:35] stop drawing dead fish. [03:38] Thank you. [03:38] which essentially is [03:41] I mean, the static image of a chat is basically the dead fish here.

You have to feel the... [03:46] the AI [03:48] to some degree. [03:49] And so, uh, [03:51] two designers, myself, just [03:54] put together the worst possible playground you could think of, of a small code base that is very LLM friendly, used the tools that LLMs are very good at using. And then we moved all of our prototyping for the specifically the chat interfaces to that. [04:09] And just to understand this playground concept, essentially, this is an idea of people work within this separate kind of area with AI tools versus like their whole...

[04:18] Notion code base, making it really easy to get started and try stuff. [04:21] Yes, and that was the first version... [04:23] It sort of aligned with model capabilities at the time, [04:26] We... [04:27] don't always use maybe at Notion, sort of the main code base is not always the most agent, [04:33] friendly because iterations and a decade of patterns [04:38] And so we optimized for, okay, how can we make this the least scary and most one-shotable so that people would just have to overcome this sort of, oh, the fear of the terminal, but then it just becomes chatting.

Yeah. [04:49] And we recreated a bunch of the patterns and UIs that exist in that playground. [04:55] Now the good news is that's just to get people on the treadmill [04:58] because as model capabilities get better, [05:01] Now we have the same designers and PMs also just contributing to the production code base. [05:07] to a lesser degree, of course, but you can see where the trend is headed as model capabilities get better. [05:13] the amount of work that [05:15] you can do [05:17] is [05:18] They're obviously going to increase exponentially.

[05:21] This episode is brought to you by our season's presenting sponsor, WorkOS. What do OpenAI, Anthropic, Cursor, Vercel, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS. If you're building a product for the enterprise, you've felt the pain of integrating single sign-on, skim, RBAC, audit logs, and other features required by large companies. WorkOS turns those deal blockers [05:51] specifically for B2B SaaS. Literally every startup that I'm an investor in that starts to expand up market ends up working with WorkOS. And that's because they are the best.

Whether you are a seed stage startup trying to land your first enterprise customer or a unicorn expanding globally, WorkOS is the fastest path to becoming enterprise ready and unblocking growth. It's essentially Stripe for enterprise features. Visit com to get started or just hit up their [06:21] WorkOS allows you to build faster with delightful APIs, comprehensive docs, and a smooth developer experience. Go to com to make your app enterprise-ready today. [06:31] Maybe give us a sense of where things are today, like how much are designers shipping stuff, PMs, and then just...

[06:37] What do you see about where things might be heading, seeing all this actually happening at a company like Notion? I feel so uncomfortable predicting the future in terms of where things are heading because, well... [06:48] Predicting exponentials is really hard, but I'll take the stab at [06:52] It is... [06:53] Very, very useful for designers to move from... [06:58] manipulating Figma documents into code. [07:01] That has always been useful. I've always been camp designer straight code. In a previous life, I led design and product at GitHub. [07:09] And GitHub designers, before LLMs contributed to GitHub, I think, [07:15] and the top contributors to GitHub itself, like 10% were designers.

[07:19] Right now, processes are sort of breaking. [07:22] One is we have designers who now [07:26] Mostly. [07:27] code and prototype in code. And then they are asked by other teams in marketing and so on to reverse engineer that in Figma because they use that to create assets for videos and so on. And so obviously that is kind of silly, right? That seems like busy work. [07:42] On the pushing to production, I think it's a spectrum, obviously a small [07:46] changes [07:47] styling tweaks and so on. [07:49] It's a given that you can just do that now.

[07:51] I do have a general sort of [07:54] maybe issue with [07:55] Vibe coding in the sense of [07:59] I don't feel like the quality of software has increased all that much in the last 12 months. I think the [08:05] maybe the amount of software has, but it's very, very hard to find software that is [08:09] is reliable. [08:11] And so the way we see it is it's not so much about pushing to production and having designers deploy. It's about them thinking and designing in the medium that will actually end up being the real thing once engineering takes it over.

[08:25] There's always talk about designers should be shipping code, PM should be shipping code. [08:29] And then there's the flip side of because engineers can move so fast, there's so much more happening. Things are moving all the time. [08:34] designers and PMs are squeezed more and more because it's hard to stay on top of all these things that are constantly shipping and so maybe it doesn't actually make sense for [08:42] designers and PMs to be spending time coding and instead [08:45] Their time is better spent. [08:46] making sure things are moving in a direction that makes sense for the business.

It's cohesive. What's your thoughts on just that balance? [08:53] I actually don't care at all whether designers write... [08:57] code that lands in production. The reason I like [09:01] thinking in code [09:03] is... [09:05] because it forces you [09:06] to consider the medium. [09:08] If then all of that gets thrown out, [09:11] Great. So for example, I think the two extremes would be if a PM or a designer [09:16] knows how to tweak [09:18] with [09:19] Pick your favorite. [09:20] they're all the same codex, cloud code, whatever. If they know how to tweak small details of the UI, but they don't understand how an agent loop works,

[09:28] I would much rather take the designer RPM that deeply has an affinity for understanding how agent loops work and can design those, [09:35] than someone who can sort of write traditional software [09:39] and tweak styles. [09:40] And that's really hard because I think the only way that you can actually get to understanding agent loops [09:46] is if you build them in the material that they're made of, which is currently code, and increasingly so if you look at all the coding harnesses, [09:54] basically the operating systems of the 90s, right?

And so I think that's why I care that people code, not because of the utility of shipping to production, but because it forces you to really interrogate the material that you're designing with. [10:07] So it's more the prototyping use cases than we're just going to be shipping more features because we can. It tends to be that once you awaken someone to a new material, that at some point they also blur the lines and then write production code. But I think it's really important not to forget that the reason why is to become a master of the material, not...

[10:28] a [10:29] sort of [10:30] cog in the delivery mechanism for the idea. [10:33] That is really interesting. What do you find is... [10:36] key to people being successful in this new world? Like, you know, there's a lot of designers, a lot of PMs at Notion. What do you find is separating the ones that are [10:45] thriving and [10:46] will do well in this coming future versus ones that may fall behind. [10:50] I suspect that this is also something that has always been the case, and we would just categorize this as founder versus not.

Do you start a startup versus not? Which is agency. [11:01] I think before... [11:04] It was very easy to always say, well, I will never be able to do this because insert skill issue. [11:09] And I think [11:10] we're realizing that even if you have the skills at your fingertips because now [11:16] I don't know. [11:17] An AGI adjacent model helps you [11:19] The thing that matters is agency. [11:22] And, [11:23] I don't think agency is very evenly distributed in the world. [11:27] And I think people who have true agency [11:30] and they understand that the world around them [11:34] is malleable.

[11:35] Will do great. [11:36] and the folks who stick to what tell me really what does it mean to be a PM [11:41] What does it mean to be a designer? And like, what's my job as an engineer? I think that will be much harder. [11:46] and [11:47] Yeah, cultivate agency. I think that's the thing. [11:50] Is there an example of someone [11:52] Using agencies, a good example, I'd notion of someone just leaning into that and doing, and maybe shipping something, changing the way something. [11:59] was happening at Notion just to give us a like, oh, wow, I see what you're talking about.

[12:03] Notinos are, this was surprising to me, [12:06] especially on the design team, way above average agency compared to other places that I've worked at. And Notinos, by the way, are Notion employees. Yes, sorry. Once you're a boiled frog. [12:18] I would say one example would be someone like Brian Levin, who you should probably have on the podcast at some point. He was on our sister podcast, How I AI. We'll link to that episode. There we go. Yeah. You should cut this one short and have him on. Um... [12:31] I think the way I would describe it is and I tell this to myself as well, which is like, OK, do you drive?

[12:37] Notion like it's stolen, which is you know, we're not the founders where you know coming in and [12:43] after there was already insane product market fit, but you can still contribute to the company in a way that you feel agency and you're not sort of just like, it's what's your role. And so Brian obviously already blurs engineering and design, but he also [12:56] is probably our number one recruiter in terms of [12:59] "Hey, this is what the org needs. I'm going to go out and talk to people." [13:03] and find someone [13:05] And I think that is a thing that sort of just demonstrates [13:09] It's out of the day to day and it demonstrates, you know, I want to just affect change.

I don't care how it happens. Right. [13:14] Eric Liu is another one, the fact that he went from sort of writing a lot of strategy docs to, he asked me at some point, he's like, "Hey, look, at some point in the future, [13:23] If you started a startup, would you hire me? [13:25] And I said, well, not in the first 10. I don't need a product manager. [13:29] He's like, oh, okay, I'm going to work on the skills so that you would hire him in the first five. [13:34] And that led to first spending more time in Figma instead of, you know, writing long PRDs.

And now it's just why do I have to do the Figma thing? Can I just build the prototype and at least show you what I think and do the thinking in there? Right. And so. [13:46] Those are just sort of signs of high agency of I'm gonna change the role to to a [13:51] to how I think it should be. Something you mentioned earlier, which I love, this idea of just rethinking what is this role of engineer and what [13:58] What should it be if we didn't have this kind of

[14:01] meme already for it. [14:03] I wonder what we lose as these roles start to merge. [14:06] We used to have this clear engineer, product manager, designer, [14:10] And as people start to, you know, as you talk about malleable software, we'll come back to this. [14:14] but malleable rolls almost. [14:16] There's how many lose, like... [14:18] clear career paths and [14:20] design consistency, things like that. [14:22] I think if we're not careful, we will lose specialists. [14:27] And so the way I would describe this is I sometimes like to think about [14:30] software in terms of physical performance.

[14:32] metaphors, right? And [14:35] Physical metaphors somehow make it so much clearer what a prototype is versus what an engineered thing is. [14:40] and [14:41] If you and I were to build a hardware startup, well, we would make the first enclosures and prototypes with 3D printing and you would see all the layer lines. It would be very, very obvious to you that this is not a thing that you should just give to people to pay for. [14:55] And then there's a long, windy road all the way to the end, where at some point, if you're very lucky, you get to manufacture that product for...

[15:02] I don't know, 100 million people. [15:04] And so then the engineering is actually the how do I optimize the factory so that we have enough yield and so that we have enough precision? [15:12] and [15:14] That to me, I think, is very absent right now from most of the discourse in software, which is it's all about how many tokens can we spend and how many features can we ship? [15:23] I'm like, okay, but where's the engineering part? And the engineering part is the you make sure that this thing works for 100 million people for a billion people.

[15:31] And on the design side, I think there is the... [15:34] Yes, anyone can now very quickly take a design system off the shelf, [15:38] build a very usable user interface, get to the core of what's really important, [15:43] But where is the delight in craft? [15:45] And so I think we have to make sure that we in this sort of merging of roles don't lose the specialists on the on the edges. [15:51] And yeah, I would say that's something we could potentially be sad if we lost it. [15:57] I want to come back to this agency piece because I feel like people hear this word a lot on this podcast.

Yes, agency. [16:03] For someone that wants to [16:05] build this within themselves or even just understand, do I have agency? I don't know. I think I do. I imagine everyone listening is like, yes, I am a huge. I have huge agency. [16:13] I'm such an agent. I can do and I'll do what needs to be done. Do you have [16:18] A piece of advice. [16:19] for someone that wants to [16:21] develop this within themselves. Partially the reason why I'm in [16:25] software is [16:27] The thing that I care most about is the Steve Jobs quote, "One day you wake up and you realize the world is made up by people no smarter than you."

[16:35] There are basically people who realize this by themselves or they have an amazing experience [16:41] teacher early on in their life that encourages this. [16:44] And [16:44] The biggest through line I've found is making. [16:49] I think if you tinker and if you make things, [16:52] then you are now on this treadmill of just... [16:55] um, [16:57] Creating. [16:58] Then you're like, oh, it's actually not that hard to learn how to make that chair in my office or let me tweak it a little bit or [17:04] Maybe, I don't know, it's like a home-cooked meal is a form of tinkering, ironically, right?

Uh... [17:08] And I think the more you can do that in life, I think actually sort of making things is the innately human, like tool making, creating art and so on. So just do that versus I think when a lot of people hear agency, they think of themselves as they're in this big machine. [17:25] And they're like, oh, okay, I'm going to circumvent my terrible boss or manager or whatever so that I get X, Y, and Z. It's like, no, no, just... [17:32] Start by making things and usually when you get better at making things at some point people pay attention and [17:38] And it just really awakens you to the idea that [17:42] You can just change things.

[17:43] I love this. There's this meme on Twitter. You could just do things. I love this version of it. You could just change things. [17:50] Which is a good segue to something you've been a big... [17:54] I don't know, advocate of and proponent of this idea of malleable software, something you mentioned earlier. [18:00] It feels like something that [18:01] wasn't actually possible and now it's like, okay, I could see exactly what you're talking about now. Like you've been on this from before the AI revolution. Talk about just this idea, malleable software, why you think it's so important, what you think people need to be thinking about here.

Malleable software is... [18:18] The idea that software works closer to the interest of the people that use it than the interest of the corporation that makes it. Maybe that's how I'd frame it. [18:28] and in particular, [18:30] Like, I don't want to use software that is [18:32] specifically just designed by the ivory tower and Cupertino. And I say this as a huge Apple fanboy, [18:40] But, [18:41] Imagine you lived [18:43] in an environment where you do not get to rearrange your living room and the kitchen has to be exactly set up the way that someone else decided.

We would not take that. Right. But that is kind of the world that we have in software right now. [18:57] where we have this world of apps and apps are like this theory. Every layer is glued together of like the user interface, the data ownership and so on. And it's like this little square. [19:07] on your phone. [19:08] and [19:09] The moment you're like, okay, this is a really great app, but I just want to change a little bit. [19:14] That is usually not possible, right, the behavior. You have the flip side, which is you could run your own Linux distribution and go that way.

And I think then what happens is you realize, oh, okay, I like the malleability. [19:26] but I also have other things to do, and I don't always want to start from scratch and figure out why the trackpad doesn't work. And so to me, it just comes back down to, do you have ownership over your computing system? [19:38] life. [19:39] And I think increasingly we don't. [19:42] Now, [19:43] You brought this up presumably because I think you may have sort of not thought about malleable software too much before AI, but now you're like making your own tools, maybe for podcast recording, for prepping for shows or I don't know, whatever.

There's a myriad. [19:57] examples and people are awakening to this idea of like, oh, I can just make tools. [20:01] and [20:02] That is a form of malleable software. [20:05] but it has to be built on top of a... [20:08] platform or an operating system that encourages this. [20:12] Because otherwise we're just doing individual, like everybody has their own individual little tool. [20:18] And I don't know, I like working with people and I like communal tools and [20:23] I don't know, this is a thing that the folks at Ink and Switch are obviously sort of at the forefront.

I get to work with Jeffrey Litt every single day now. [20:29] That's... [20:30] spend a lot of time thinking about how would we make [20:33] software more malleable so that we feel more ownership over it. [20:38] without going back a long time and not having real-time collaboration and sort of the security aspects and so on. [20:45] I really love and I just want to make sure we highlight this idea you're sharing. [20:49] It's something that I learned also from Brian Chesky at Airbnb. [20:52] This idea that you can change things, that the things around you are just [20:57] made by other people that may not actually be smarter than you.

And it's just this really empowering process. [21:02] thing to always think about that. [21:04] Things can change. This isn't the way things have to be forever. [21:07] People, humans made this thing, like humans made this phone and you could, and there are better ways. [21:14] approaches that other humans that you can that you can come up with other people will come up with [21:18] What I think about as I think about this is there's a video that you pinned to your Twitter profile that we'll link to, which I think is Dita Roms.

Is that who the person is? Okay. [21:28] He's walking around, he's just criticizing all these designed chairs. [21:33] Talk about what? [21:35] that video is trying to why you pin that to your profile. [21:38] There's many reasons. One is, I think, maybe the only thing that I have in common with this [21:45] very accomplished person is that we're both German. And for sometimes I joke that I also aspire to disapprovingly just point at things with my walking stick. [21:54] and say, this isn't good enough, this isn't good enough. The reason why is because I think if you speak German, this is one of the funniest-- [22:01] clips that I've ever I just die laughing every single time.

I'm actually curious how you think about [22:07] how it ties to... [22:11] malleable software because the main reason why [22:15] I use that as a clip of reference is [22:18] I'm very much in the camp of design should be first useful and then beautiful. And I think a lot of the pieces there are predominantly things that you put in a museum [22:29] for display, and if you try to sit on them [22:31] You'd be like, what is this nonsense? What I felt there is just like you see, like you would think it was Frank Gehry and like all these famous designers pieces.

[22:39] put up in a museum. And I think to most people, it'd be like, wow, this is so incredible and beautiful. Like, you see somebody that has a status... [22:47] and a reputation, and you assume this is great. [22:49] And I love that he breaks that veil of like, no, this is so stupid. What is this? What is this bunch of cabinets tied together? It doesn't make any sense. Yeah, yeah. He said for that cabinet, I think he says something like it is neither... [23:02] orderly nor properly chaotic. [23:06] I understand the connection now.

[23:08] The timeless way of building and Stuart Brand's how buildings learn [23:14] I think idea [23:15] is also that it's very likely that the best [23:18] Homes for you are not actually built by an architect. They are the thing that over a [23:22] a long time adapt to exactly how you would love to lead your life. [23:27] and they learn over time versus immediately. And so then that is obviously a very costly version of malleability, right? If you have to rip out a wall or whatever. [23:38] But, um... [23:39] I think the [23:41] main thing [23:42] thing [23:43] that Dieter Rams points out there is it should be a thing that's useful and [23:48] a good way to figure out [23:50] how something is useful is if you can change it and tweak it [23:53] It all connects.

It all connects. I get it now. [24:12] Salesforce, you know, I mentioned some people are like, we don't need Notion, I'm going to build my own Notion. [24:16] You have a hot take there. Talk about what you think is going to happen. [24:19] If you just think about what SAS, the problem is the moment you have an acronym, [24:25] It means a lot of very specific things. And if you're going to say, "Hey, is this type of SaaS that we've built [24:31] in the 2010s. [24:33] just as relevant as it was in the 2010s, the answer would be, it would be silly to say, no, nothing's going to change, it'll be the same.

Because I think you can sort of say a lot of SaaS in the 2010s was a very, very fancy form. [24:48] around a spreadsheet or something more generic. [24:51] And the thing it did is it just guided people in the right direction to fill out that form. [24:56] as in it is less malleable than a spreadsheet and that sort of is the value. [25:00] The as a service part is, I think, the thing that actually matters. [25:03] which is I don't think most people actually want to maintain the full stack of software.

And so whenever I see someone, and I am someone here, say, oh, I just rebuilt this piece of software. I've tried rebuilding Notion in a weekend for myself just to push at the edges of frustrating things. [25:21] I don't think people want that. I think for the most part, [25:24] It's nice if you can just people don't want to go hunting either. They just want to go to Costco and have the steak in a styrofoam. [25:31] packaging and pretend that it wasn't hunting or, you know, an animal in the first place.

I think with software, it's like it's a it's like a Brett Taylor says this to you. [25:41] Software is like a garden, you need to tend to it. And the thing you pay for as a service is the maintenance. [25:47] and a bunch of specialists thinking really hard about a problem. [25:51] And so I don't think that's going away. What I would probably say is that tools will become more general. [25:56] I'm obviously biased. I work at Notion. I like Notion and [26:00] I consider Notion to be fairly malleable, not enough.

I think it should become more malleable. We internally joked, Joanna Stern, [26:09] A journalist recently tweeted something along the lines of, oh, thanks to Notion AI, I finally understand and use Notion. I don't know what that says about Notion. And to me, this is a great example of... [26:20] Notion wasn't SaaS in the traditional way. It's kind of hard to get started. [26:25] But because of AI, now people can sort of, they have a tutor essentially and can build more things. [26:32] And so I suspect that software will go more back into the 90s of general tools, word processor, spreadsheet, [26:39] FileMaker Pro, [26:40] that kind of thing.

[26:42] But, [26:42] those will still be as a service. And then you'll still have specialized tools around security and so on, of just people who go the extra mile to really solve a user problem. So I think to some degree, the SaaSpocalypse is greatly exaggerated. [26:55] At the same time, are things going to stay the same? [26:57] Of course not. Like, why would they? [27:00] I completely agree. I think people think about just the, okay, I'll create something that's pretty cool and close. And then they don't think about exactly as described.

Like I have to maintain this thing forever and I have to keep adding features, taking people's feedback. [27:13] One of the funniest things that I... [27:14] See you again and again. [27:16] I just had the head of product for Cloud Code on the podcast, Kat Wu, and she talks about how Slack is basically the OS for Anthropic. [27:23] Everything runs through Slack. And you think of all companies that would just like, we don't, we'll just build our own. What are we doing with Slack? Like, no, they're just... [27:29] They're using Slack like crazy.

And I think that's just one example of like nobody wants to rebuild a tool like Slack and WorkEthic is another example. [27:38] I don't know, I think it's maybe even more unique in the US, but one of the great things about the US is actually specialization. [27:45] It's that I get to spend dollars on something like Notion because it's not that expensive compared to me building it [27:52] And then why would I waste my time? I want to do other things with my life. Right. So I don't know. That's not going to go away.

Yeah, I agree. No, the guy people in the public there, their time is better spent on building AGI than trying to build better slack. [28:06] I also love the Slack example because there's this [28:11] graphic of [28:12] what it takes to deliver a notification, [28:15] in Slack, the sort of decision flowchart. And that is just something that you only get to when you have real users, real scale, [28:22] and [28:23] decades of just, yep, we understand the customer. [28:26] I want to come back to how... [28:28] product building is changing and how it's different.

I know you've done a lot of different jobs, but like your job, I don't know, a couple of years ago, what's most... [28:37] changed. [28:38] Like what part is most not something you don't do anymore? [28:42] where you do a lot more of now with AI technology. [28:44] emerging as a big part of your [28:46] process. I think the first 10% of every project are now free. [28:50] That's how I would describe it. [28:52] So there is no point for most things to, for example, write a, I don't know, the thing has changed.

I've never really been great at this, but like there's no point in writing a PRD if you can just do the janky version and sort of, you know, [29:07] do the, here's the demo of what I think we should build. So the first 10%, that's so interesting, just the [29:12] That's such an interesting way to frame it. The idea there is just like the thinking through of it. You can go a lot further really quickly. Yes, and if you look at a lot of the... [29:22] it takes almost no effort to now build

[29:25] sort of the first version of a startup, right? Or like the first [29:29] version 0.8. And then I think the last, or maybe even if you're generous and say the first 90% are now down, the last 10% are still actually 90%. That's always the hardest. So I think [29:42] It's cheaper to just explore a lot of paths. [29:46] you can now afford to say, I'm going to send off 10 agents to explore 10 different things and then see if I was right. [29:53] We used to say this at GitHub.

[29:55] in our product reviews a lot, which is demos, not memos. [29:59] And then we would say, give me something to react to. [30:03] Which is [30:03] Okay. [30:04] If you're gonna write a PRD, just write the changelog or the blog post that a user would read. Now it's much easier to give people something to react to. [30:13] As in, yeah, here's the version of the product. And it's like, OK, what if we did it this other way? Oh, yeah, here's that version. [30:18] And so I think that is just amazing.

[30:20] it sort of builds an iteration into the product much earlier, right? Like waterfall is sort of [30:27] Why bother? [30:28] What do you think is the next kind of leap or shift in how we build? What are you seeing as like, okay, this is now the new thing that's emerging that is going to change how we operate? [30:38] I'm very conflicted on this because on one hand I do want to, like I believe the never bet against plain text [30:44] It's a famous... [30:46] forum post at some point [30:48] Plain text, markdown, like it's just such a durable thing.

Code is such a durable thing. [30:52] I [30:53] think that [30:54] Expressing your thoughts in code is probably a really good thing. We can talk about why. But at the same time, I'm like, are we really going to just be chatting back and forth? [31:05] And so what is the future of Figma, for example, is like a really interesting example to me because on one hand, I do see like sort of a drop in usage of Figma in some designers at Notion. And then others are like, nope, these AI tools are wonderful.

It's very hard for me to predict of like, is direct manipulation going away because the agent is doing the direct manipulation? Yeah. [31:28] Um, [31:29] The other thing that I'm curious about is [31:32] There is this... [31:34] Automation versus augmentation [31:36] fork [31:37] If I look at the really, really fast models like Spark, and I forget what the anthropic variant is, [31:43] where... Haiku? [31:45] No, sorry. It's it's you still get a smart one. It's opus, but like opus fast or something. Yeah, like you very quickly run up a bill of like three thousand dollars a day.

[31:55] But... [31:58] the speed of inference really changes things. [32:01] If the inference is slow, then you're queuing up a bunch of jobs and then you're walking around the building thinking about other things and then come back and review. Versus if it's nearly instant, are you still going to do this? Is this sort of multitasking, the frenetic kind of thing that we currently have going on actually the thing that is sort of, you know, gives us flow state? Well, no. But if the inference becomes instant, [32:23] do we [32:24] get back to direct manipulation, right?

Do you instantly sort of like mold the clay that is the code? Right? I don't know. I think it depends on model capabilities. [32:33] which is do people [32:35] Is there a saturation on intelligence or not? [32:39] The analogy I like to give is a retina display. [32:43] which is... [32:44] After I can't see the pixels, I can't see the pixels. I don't need you to make them smaller. [32:49] Is it not the same for a lot of cognitive tasks, which is at some sort of level of intelligence? I don't need more.

And I instead I want a different modality and faster. So I don't know those things I'm excited about. Interesting. So the last point you're making is it's [33:03] Like smarter models will not significantly impact how teams operate because they've gotten so good and it's other... [33:09] blockers now like like ux essentially [33:12] Yeah, I think in general, I'm actually very curious. The labs sort of operate [33:18] I feel like they operate under the assumption that people will always want the smartest model. Like you want the frontier model. [33:25] And I think for certain domains that is probably true.

I think if we're going to do cancer research and so on and if we're going to spend millions of dollars on something, that's likely true. [33:34] But that's not how we run companies either, right? Like we don't have a PhD for everything. [33:39] And so I think for a lot of knowledge work tasks, probably sometimes we'll get to good enough. [33:45] And once you get to good enough then you can optimize other things like they run locally. They're cheaper. They're faster and [33:51] and [33:52] I don't know why the absolute intelligence thing doesn't interest me very much.

I think society is largely not capped by intelligence. [34:00] I think Tyler Cowen says something similar. I don't want to put words in his mouth. Um, [34:05] And so I'm much more interested in the [34:08] exoskeleton... [34:09] versus the I have a god in a box in some data center somewhere and we're all sort of [34:15] you know. [34:16] twiddling our thumbs. I have a bunch of questions along these lines, so interesting. [34:20] You talked about how this 1PM is the highest token spender. This is across all of Notion. [34:25] I would assume this may not include our automatic security vulnerability scanning and bug triaging.

[34:33] Human kicks off jobs. Yeah, what's your policy on token spent? Is it spend as much you want? Here's a limit everyone Do you keep track? Given that I don't know what the policy is. I think it is unlimited. I mean You can imagine at some point there would be but right now I think it's just the wrong thing to optimize for it's like when something new comes along and [34:53] It's worth... [34:55] letting people explore. [34:57] I. [34:58] do suspect in [35:00] six to 12 months from now, a lot of companies are going to actually start asking questions around RRI.

[35:06] And I think that will be an uncomfortable conversation for a lot of folks. [35:10] In terms of span, what are the numbers for, say, Eric or Rodley? [35:15] I am the wrong person to ask. [35:19] I would assume they pale in comparison to the folks at OpenAI and Anthropic, just by the nature of the work they do and so on. [35:26] but [35:27] It is definitely... [35:28] for an individual in the, you know, [35:31] I don't even want to put numbers in, but like thousands for sure, but like maybe tens of thousands.

I don't know. It depends. Yeah. I think just the fact that you're as had a product or not on top of that means that it's just let's not worry about this. Let's just see what we can do. And then we'll, you know, in six months, as you said, we'll figure out if this is ROI positive. Yes. I have the luxury to right now not care. Yeah. And I'm sure, you know, someone's looking at it. It's not going to be out of control. [36:01] It's on salary. [36:03] That's something people talk about now more and more.

Just like, should that be higher than your salary? Should it be lower? How does that all connect? Yeah, I think there's a real danger in... [36:10] sort of making the token spend the metric to like, [36:16] boast about [36:17] which is the same as when people boast about how many lines of code they've written in a day. And I'm like, I... [36:24] Why do you have so many lines of code? You have, I don't know, [36:30] The largest software projects in the world have... [36:34] it. [36:35] not that many millions of lines of code.

Why are we bragging about that? [36:40] I don't actually care about how many tokens someone spends. [36:44] Yeah. [36:45] It's not a metric that's useful. [36:46] Yeah, such a good point. I know Meta got some flack for this recently where they're trying to create a leaderboard of who's doing the most. [36:52] To be fair, I do understand why companies do that, which is I am surprised by how much work it takes to get people to... [37:01] to [37:03] identify the outer loop of their work. [37:05] and enlist an agent and build sort of the, I don't know, the term right now is like factory, right?

Like the software factory for the work that they do. [37:13] It is... [37:14] surprising to me how much [37:17] prodding you need to do [37:19] to get people out of their, the way they're used to working. And so if you're dealing with [37:24] tens of thousands of people at the scale of Meta [37:27] I have some sympathy for, okay, a good way to do this is just start a leaderboard and encourage people to do it. [37:34] they will find good things and useful things to do with that as they learn.

[37:40] It's such a good point. Like you have to over... [37:42] over index to change people's [37:44] default easy behavior. I'm just going to do things. I'm just going to write these PRDs the way I've always done that. I'm going to run the meetings the same way I've done it. [37:50] I think that makes a lot of sense. What have you seen actually work with the notion to get people to [37:55] significantly changed the way they work. [37:57] Depends on the role so roles that are [38:01] perhaps further away from engineering, [38:04] Actually, you don't have to convince them all that.

[38:06] much because they're like, "Whoa, I have superpowers now." [38:09] look at this amazing thing I've just built because the capability gap of what they were able to do before versus after is so huge that, [38:16] It's intoxicating. [38:18] And then you have to actually almost do the opposite, which is like, yes, but do you understand why we can't merge this PR? [38:23] I think on the engineering side, [38:26] something that Simon [38:29] last week. [38:30] talks about a lot is sort of any manual intervention [38:33] ENCODE. [38:35] is kind of bad, you probably did something wrong in the verifiability loop [38:41] and in sort of the software factory.

[38:43] This excludes, obviously, reviewing code, right? Like, I am still very much in camp [38:48] you should probably review more code than put more effort into reviewing code than you do. [38:53] but at least on the writing side. [38:56] Every time there is an intervention, a human intervention, it should feel a little bit like a bug. I think that's a good... [39:01] Litmus test for [39:03] how, I don't know, age-infilled you are. [39:06] I want to come back to the tools that you use. You mentioned Figma is kind of trending down within the design org, which is really interesting.

[39:12] Is there anything that's trending up, anything else that's trending down in terms of tools in the tool stack of your team? So I'm actually not positive that Figma is trending down. I think it's more that there is a... [39:24] There's two camps. [39:26] I could totally believe the Jevons paradox, which is Figma is actually going up. [39:32] And then of course, vibe coding is going up like I don't want to create. In general, I really, really dislike the the rivalry discourse that exists in [39:40] in Silicon Valley, which is for [39:42] Anthropic to win, OpenAI needs to lose and vice versa and like [39:46] that kind of thing.

So I, I, [39:48] I don't want to... [39:49] perpetuate that with sort of the Figma versus coding. [39:53] I think the terminal is actually surprising, which is [39:57] It's initially kind of scary for people and you could do so much, but now PMs are slowly [40:02] Once they're in cloud code or codex, everything is fine. And I generally encourage them to not use the GUIs. I encourage them to use the [40:11] the Tui's because I just know that over time they're going to be curious and pull at other threads [40:16] And one day they wake up and they're like, oh, I understand more of the substrate of how computers work.

[40:21] That is so interesting. So the designers are using the terminal? Yes. And then, I don't know, Conductor is another one. They're basically just mostly using developer tools. It's not that different from... [40:31] what developers use. I am so excited to tell you about this season's supporting sponsor Vanta. Vanta helps over 15,000 companies like Cursor, Ramp, Duolingo, Snowflake, and Atlassian earn and prove trust with their customers. Teams are building and shipping products faster than ever thanks to AI. But as a result, the amount of risk being introduced into your product and your business is higher than it's ever been.

Every security leader that I talk to is feeling the [41:01] increasing weight of protecting their organization, their business, and not to mention their customer data. Because things are moving so fast, they are constantly reacting, having to guess at priorities, and having to make do with outdated solutions. [41:15] vanta automates compliance and risk management with over 35 security and privacy frameworks including sock 2 iso 27001 and hipaa this helps companies get compliant fast and stay compliant more than ever before trust has the power to make or break your business learn more at

com slash lenny and as a listener of this podcast you get one thousand dollars off vanta [41:40] That's com slash Lenny. [41:43] - AI has completely transformed the work of a software engineer. Like two years ago versus today is completely different. [41:50] like almost all your code is now AI. And we've been talking about like, when will 50% of engineers in the world be writing 100% AI code? [41:59] It's probably like a year, which is... [42:01] Insane how much that job has changed. [42:03] Which role do you think AI transforms [42:07] Next.

[42:08] Is it marketing? Is it growth? Is it sales? Is it design? Do you have a sense of where things are starting to really change other than engineering? [42:15] Okay, this is maybe a hot take, and I actually don't have enough. [42:19] It's very likely that the labs are like, ha-ha, look at this guy. [42:22] Um, [42:23] My take, it's very clear, at least empirically, that models are getting better at coding [42:29] at some exponential rate, right? [42:32] And I don't think that's changing. [42:34] Now, [42:35] I'm not that impressed with the progress in any other domain.

[42:39] It tends to be... I don't think they've gotten significantly better at writing. I still very much hate reading... [42:46] sort of AI slot writing. [42:49] But the thing is, software, Andreessen, right, like software is eating the world. Well, if the cost of software and creating software and encoding business practices in code, and like I just literally mean the old like software 1.0 kind of code. [43:04] then [43:05] If that cost is very much going to zero, we will just have a lot more of it. [43:09] And so I think then in that case, [43:11] It's more that software engineering will go into all the other domains.

[43:16] Not necessarily that there is sort of some sort of [43:19] Yeah. Like, [43:20] I don't know, our [43:22] folks in HR are automating a lot of things because now they don't have to bug an engineering team to write that code. [43:29] And so I think that's how it's going. And like if you look at when [43:33] The model companies say, "Oh, we've made great progress in this other non-coding domain." [43:37] I was like, [43:38] "You just applied coding principles." [43:41] to this domain, which is wonderful. [43:44] But that's what it's getting better at, right?

And so I think... [43:46] I just think software is eating the world is going to accelerate. [43:50] That is a really interesting take. So it's basically software, just the acceleration of software eating the world. [43:55] versus it's like it's going to now do a different kind of job. [43:59] This makes me think about the header product for Codex said the same thing that every agent [44:04] There's all these different kinds of agents, and his take is... [44:06] Every agent that will win is going to be a coding agent. [44:08] that builds the thing it needs versus like it's come, it has like certain number of capabilities.

OpenClaus is such a good example. It's just like, I will build a skill for myself and now I know how to do this thing. [44:18] Yes, all agents are also like if you look at all the harnesses, whether it's the open source ones or the ones from the model companies, [44:24] ours as well, they all resemble a coding agent now. [44:27] I'm going to come back to the ROI piece. I think this is really interesting. As you said, there's just like, okay, we're going to spend, spend, spend, just see what happens, learn, accelerate.

[44:36] lean into all this stuff [44:38] You're saying that in maybe six months, something like that, you think a lot of companies are going to start really looking at the cost here. [44:44] What do you... [44:45] And he said you don't like to predict things, but what he predicts is going to start happening. I probably spend too much time than I should because I have literally zero impact on any of this, sort of how it plays out. [44:55] But [44:55] You can imagine a world where the labs, the delta between the labs and open weight models and so on, [45:03] widens [45:05] That is a world that I very much don't like because I hate centralization of power.

[45:11] But in that world... [45:13] I think the labs just kind of get to decide [45:17] what the world looks like. I think if that [45:20] gap doesn't widen, then you will just see a diffusion and people will get very comfortable running their own models, RL-ing their own models, right? Like you see this with Cursor, you see this with Intercom, [45:32] Notion is dabbling in it as well. I use dabbling right now, but obviously at some point we might become more serious about it. [45:39] And then you have like, it may not be the frontier, but for a lot of tasks it'll be good enough.

[45:45] And so I think in that case, that is just an ROI calculation, right? That is the [45:49] "Is it cheaper for me to send this task to a smaller model that is cheaper to run, where I remove the lab sort of profit margin kind of thing?" I think that may happen. [45:59] But it only happens if there isn't a fast sort of like, you know, [46:03] Oh yeah, the gap is now so big. The other one that's interesting is right now I think we're actually in one of the luckiest [46:10] possible timelines, which is we have at least in the US, [46:14] three competent labs that are all sort of duking it out, and like, who knows, maybe meta now, so four, [46:20] Maybe we can make it six at some point.

I think I would love a world where we have a dozen frontier models in the US versus having to always rely on... [46:29] on [46:31] uh... [46:32] other places in the world to do this. But that's sort of pretty good. If that chews... [46:37] Stopped I would be somewhat worried. Um [46:41] And then it's hard to predict what would happen. [46:45] But if that doesn't, then I think it's going to look similar to the Cloud Wars, which is [46:50] at some point layers commoditize [46:52] businesses are not going to want to lock in into one single provider [46:56] I don't know, in my past life I worked at Heroku and like Kubernetes was much more successful than Heroku, even though I think from a user experience perspective, it was.

[47:05] much worse, but the delta was, Heroku was saying, hey, we're going to [47:09] replace your ops team [47:12] and Kubernetes was we're going to make your ops team superheroes. And also we're not going to lock you into a cloud you can choose. And obviously that's what businesses want, right? Like businesses want [47:21] choice and so I don't know it's really hard to predict because [47:25] It depends. So it's so asymmetric in terms of model progress. [47:29] When you say that, [47:31] Products that win often are the ones that make you feel like superheroes.

I always think about Cathy Sierra. Do you remember that at all? That's the thing it rings up [47:38] Oh, okay. It's like from the olden days at this point, shows how old I am. She was just something that really stuck with me, and I think it's informed a lot of how people think about product, at least in the past, is just... [47:49] her whole pitch was [47:50] Instead of talking about your product and how amazing it is, it's about we will make you... [47:55] a superhero. Like it's like Mario getting the little flower and having superpowers now versus look at her incredible product.

I think it's actually a thing that the coding companies had to learn when they tried to move to like why do code review tools, automatic code review tools not work that well. [48:11] I think there's actually a subtle thing, which is you push your code publicly, or publicly, within your organization or your team, and then a thing roasts your code and tells you how terrible of a developer you are. Versus if you think about what Cloud Code and Codex does, [48:27] is, [48:28] You're coding. [48:30] And then you publish the work of you plus Claude, and you get bragging rights of how good of a developer you are.

And so I think the superhero stuff is definitely true. [48:39] Speaking of superhero, I wasn't planning to talk about this, but... [48:42] I've been hearing a lot about how much people... [48:45] love your agent, the Notion AI agent that you all released. Just like it's just coming up a lot of just like, wow, this is actually really useful. [48:53] with like a lot of different people. It'd be interesting to hear what you think made it so successful. I know it was like a long [48:59] time before you guys launched it, [49:01] Just like [49:02] What do you think is helping it be this useful and successful as a product out in the world?

I would like it to be... [49:09] Even better. [49:10] So I'm like my... [49:12] first critic, I guess, I've spent most of my day thinking about where it falls short, not how great it is. But I agree with you that [49:19] I'm actually surprised at how [49:22] This sounds so weird. I'm surprised how good it is. [49:25] if that makes sense. Notion has always been fairly at the... [49:30] forefront of AI. Like I think the first Notion Assistant was actually launched before ChatGPT. [49:35] And so it's not that like I think both Ivan and Simon [49:39] had the intuition of, hey, this is going to change a lot of things.

And so that's a huge [49:43] sort of reason why but there every company wants to become ai native now whatever that means i it's kind of like cloud native i'm like if you have to say it then are you really do you have a chance [49:54] But I'm surprised how fast that happened for Notion. [49:57] And I'll take... [49:58] Almost no credit in this. [49:59] I think what's good about it is agents need context to operate in. [50:05] Agents don't really like walls of like, oh, I have to go through this narrow orifice to talk to this other data repository.

And...