# ⚡️ The best engineers don't write the most code. They delete the most code. — Stay Sassy Latent Space · 2026-04-13 Stay Sassy PM and EM join Shawn Wang to discuss how AI coding tools are forcing managers to grapple with per-person token budgets that could reach $2.5M annually as consumption-based pricing replaces subsidies, making code review more critical, not less. They argue that managing these budgets is a new bottleneck, requiring companies to decide how much to spend on individual employees—a scale unprecedented outside department-level budgeting. On build vs buy, they caution that many products are more complex than they appear, and that the old frameworks of feature analysis, administration burden, and vendor lock-in still apply. They highlight code review fatigue from running multiple agents as a risk, citing Amazon's six-hour downtime from AI-generated code, and urge teams to maintain a culture where no single person can take down prod. Finally, they suggest automating executives' standard decisions first—since much leadership work is routine—rather than targeting only junior tasks, and predict that the hardest problems in 2026 will remain human. ## Questions this episode answers ### Why does Stay Sassy think managing AI token budgets per employee will become a major management challenge? The Sassy EM warns that as AI tools move from subsidized flat fees to consumption-based API pricing, annual per-employee spend could range from $500 to $50,000 or more, akin to laptops. Managers will need to make unprecedented per-person budgeting decisions, deciding who gets what and facing potential bottlenecks when teams can build faster than the company can convert output to revenue. [6:57](https://addtry.com/3dcd5cf8-2c81-4391-a141-e59860da0b73?t=417000) ### What is Stay Sassy's advice for deciding whether to build software internally or buy it, especially with AI? The Sassy EM advises listing all features of a product to realize its complexity; if it can be replaced by a spreadsheet, you can build it. The Sassy PM warns that homegrown software ties you to key builders, and maintaining uptime, scaling, and handling edge cases often outweighs the appeal of building. Old build-vs-buy principles still apply, but AI lowers barriers. [15:54](https://addtry.com/3dcd5cf8-2c81-4391-a141-e59860da0b73?t=954000) ### How did Amazon's outage highlight risks with AI-generated code, and what should engineering teams do to prevent such issues? Amazon's 6-hour outage attributed to 'vibe coding' shows the danger of skipping code review for AI-generated code. The Sassy EM stresses that teams must maintain a culture where no single person, new or AI-augmented, can break production; code review must remain rigorous. The Sassy PM adds that AI-induced fatigue from managing multiple coding agents risks exhausting reviewers, making them miss critical issues. [42:25](https://addtry.com/3dcd5cf8-2c81-4391-a141-e59860da0b73?t=2545000) ## Key moments - **[0:00] Intro** - [3:03] Stay Sassy saw their blog spread organically within top companies, with 20-30 employees subscribing from a single firm. - **[3:30] Growth** - [7:09] 2026 will see a shift from subsidized AI to consumption-based API pricing, creating a budget management cliff for companies. - [9:39] AI token budgeting will turn every employee's spend into a department-level financial decision, with direct ROI calculations. - **[10:00] AI Budgets** - [11:32] The move to API-based pricing has no precedent, leading to organizational blame and credit fights over token spend. - [13:34] An OpenAI researcher spends 1 billion tokens per day, costing approximately $2.5 million per year. - **[16:12] Build vs Buy** - [16:12] Shawn Wang considers building a custom tool for $50K in token spend instead of paying an external vendor $250K per year. - [18:50] "Everyone says everyone else's job is gonna get done by AI, but not mine." - [20:05] If you can replace a software product with a Google sheet, you can build it internally. - [22:58] The UI layer of internal tools should be end-user modifiable with AI, similar to Retool but personalized. - [25:44] Stay Sassy EM got frustrated navigating Google Cloud console to find an API key, wishing it were agent-compatible. - **[27:22] Automation** - [27:46] AI value falls into two categories: pure automation of tasks and decision-making for strategy. - **[31:20] Executives** - [31:20] Companies should consider automating executive decision-making before junior roles because execs are bottlenecks and much of their work is standard. - [33:42] In the next year or two, there will be fractional or even AI executives for standard business functions. - [35:15] 99% of managerial decisions are knowable; the hard part is having the willpower to execute, not the cognitive complexity. - [37:24] Stay Sassy was born after a meeting where they realized many management challenges are knowable but lacked written guidance. - [40:02] Within two years, the first company will try an AI manager, but it will be weird and likely fail. - **[42:12] Code Review** - [42:25] Amazon suffered a 6-hour outage attributed to vibe coding and their tool Hero, showing the risks of AI-generated code. - [44:18] "I could literally put a malicious actor in your team, and they couldn't take down prod because of required code review from knowledgeable colleagues." - [45:48] AI coding tools are leading to armies of one and code review fatigue, risking reliability as fewer eyes review mission-critical changes. - [49:02] Code review could become regulated like pilot duty hours, with mandatory rest periods to ensure software reliability. - **[50:45] Closing** - [51:37] Stay Sassy's core message: care deeply about your team's well-being; you are not alone in startup struggles. - [52:58] 2026 will be the most dynamic year in software, changing everything from engineering practices to business operations. ## Speakers - **Swyx** (host) ## Topics Coding Agents, Enterprise ## Mentioned Amazon (company), Anthropic (company), Netflix (company), OpenAI (company), Temporal (company), AWS (product), Chaos Monkey (product), Claude Code (product), Hacker News (product), LinkedIn (product), Retool (product), Stay Sassy (product), Substack (product), X (product) ## Transcript ### Intro **Swyx** [0:05] Okay. Hi, this is, uh, swyx in the remote studio with a very interesting podcast guest, uh, duo. We've done a anonymous podcast before, and this is our second one. This time, though, they are also are podcasters. This is the Stay Sassy crew. Say hi, guys. **Guest** [0:23] Hey, how's it going? **Guest 2** [0:24] Hey, what's up? **Swyx** [0:25] One of you is Stay Sassy PM, which, which, which one? **Guest** [0:28] Yeah, that's me. **Swyx** [0:29] This one, this voice. And then the other one is Stay Sassy EM. **Guest 2** [0:32] Yeah, and that's me. **Swyx** [0:33] How do you guys introduce Stay Sassy? What, what's the one-liner pitch? **Guest** [0:37] Say the one-liner pitch is that we are a blog, and now, now I guess a Twitter personality or X personality, a-as well as podcast personality, about building and scaling really technology businesses, especially from those early startup days all the way out to how far, however far you take it. **Swyx** [0:55] Yeah. I know you guys since my Temporal days, where, like, my former boss was so in love with you guys. He's like, "This is the best blog e-ever," and it's anonymous, which like, again, how do you grow an anonymous blog? Let, let's just start there, right? Like, because you don't actually want people to know. I mean, you're, you're somewhat doxxed to, like, people who do know you, but, like, how do you even start? **Guest 2** [1:18] I think it's a-- it, it was a really interesting journey. So I think at the beginning, the easiest way to start was just posting stuff and trying to share it in different locations, and so the, the simplest answer is just starting getting paint on the board and, and seeing if we actually get some writing that's actually useful and good. But then it was really about how do we distribute it, and how do we start to get eyeballs on it? And when you're anonymous, it's really, like, you can't go and, and tell your friends to post it on, you know, on Twitter or something like that. So I would say in the early days, Hacker News was actually, like, a really great growth engine for, uh, the blog. And we had, in its very early days, maybe, like, five to 10 posts a year that would get, like, front page, and that was a, a huge sort of, uh, boost to our, our kind of early growth of the blog. From there, we pivoted to Substack off of just having our own kind of website. Substack is a great community and, and that is, like, its own sort of place where we've started to get people engaging and kind of building, uh, and that was really great. And then we added Twitter on top of that, and that has really been, in the last year, like, for, I would say, like, the overall sort of engagement, a total kind of turbocharge on just, like, engaging the community and kind of being in this kind of, um, you know, whatever sort of, uh, ecosystem where now we have a blog and we have subscribers there, we have Substack, we have Twitter. So it was really just trying to find the first inkling that we could get anyone to listen to anything that we were doing, and it was thrilling when we had that. And then Hacker News was, like, a, a really great boost in the beginning. And then the only final thing is it's been really exciting when we see the subscribers come in to see the S- the Stay Sassy blog kinda run through companies. So you'll see, like, somebody come through, uh, at a company you know, and you'll see that subscribe, and you'll be like, "Oh, that's awesome. That's, like, really exciting- ... that this great company and this great person at that company is reading our blog." And then a day later, you'll see two more people in, from that company subscribe, and you're like, "Oh, some people are sharing this around." And there's some companies that, you know, we have 20, 30 people from, and you know, they're one of the best companies in the world, and you know, there's, there's handfuls of, uh, companies that we feel like we've, we've gotten in their sort of internal blogosphere. So that entire sort of, uh, maturation process has been very kind of probabilistic, but very gratifying. ### Growth **Guest** [3:32] Yeah. I think another aspect that has helped a lot is that we keep the same general subject matter, no matter what platform we go to, and we'll keep sort of the, the same soul, but we will definitely adjust the tone and talk about different sorts of things on different platforms. Like, our blog has always been much, much more serious. There will be something funny in, uh, a blog or two every now and then. My, my parents are sometimes like, "Oh, that was, that was kind of a funny one." But generally, it's we try to make it really right down the middle, really actionable stuff that you can use in your day-to-day. Our X i-is a lot wilder if, if people follow us. Like, there, there's a lot more stuff. **Swyx** [4:11] There's more shitposting. I, I'm, I'mma put it up on screen right now. **Guest** [4:15] Yeah. A, a lot more shitposting, a lot more, you know, talking about crazy stuff that happened to our, in our personal lives or, or the Sassy EM's, uh, personal life as a child. I think he had- ... a much wilder childhood than I did. And, and so we will sort of venture out there, 'cause that's sort of what the X ecosystem's about. Like, you know, it's a place- ... for fun as well, as opposed to Substack. You know, we're, we're kinda here to learn. We're, we're going to school when we open up Substack. **Swyx** [4:39] Yes, yes. Um, so I actually, I think I reached out after this specific one. My favorite insult is calling someone's job a cloud skill. Uh, we're gonna make a T-shirt, uh, out of this one. Is there a story behind, like, some of your, your, like, one-off tweets? **Guest 2** [4:54] Yeah. I would say, like, every tweet, you know, comes from different places. One thing that I, I sort of, uh, is, is always, I think, probably s- maybe surprising is, like, very little of this actually comes from my day-to-day work, and a lot of it comes from, like, talking to people about their jobs and friends about their jobs, and it's, like, very actually shocking how similar tech is to various, you know, mechanical engineering or finance or something like that. And so I think that one in particular, I was talking to a friend, and they're in finance, and they were just talking about this coworker who they were really slow in responding to things, and the only thing that they did was kinda respond with an email that, like, checked one thing and affirmed that you can do it. And so I was talking to this friend, and they were like, "I'm waiting. My whole job is waiting on this person to send an email that just says, like, one tiny thing, and there's, like, three sections that they could possibly sort of go with it." And, uh, I think that's where the inspiration where that came from, where I was like, "Oh boy, that's a-- I don't, I don't know that that's, like, a full white-collar job in the future." Um, and so yeah, it's-- And, and then from there, you know, I started to think about just, like- All the different things that kind of look like just replying to an email, the decision tree of cardinality four or something like that, you know? **Swyx** [6:07] My, uh, here, here's the shirt, uh, from our designer. Go away- **Guest** [6:13] I love it **Swyx** [6:13] ... uh, uh, sorry. Go away or I'll replace you with a very small skills.md. Yeah. That's the name. **Guest 2** [6:19] So you think- **Swyx** [6:21] Uh, yeah, I mean, you know, this is like, I guess my low-key way of announcing, announcing that we're working on a merch store. Um- **Guest** [6:27] I love it. Love it. **Swyx** [6:29] Okay. So, uh, ship posting aside, uh, you guys are actually serious people, uh, uh, working in, in a real company that I can vouch, right? You know, you don't- you are anonymous, but I can vouch that, um, I... You, you guys know what you're talking about. We prepped some topics, right? You're running teams, PM side, EM side. EM, like you, you maybe sent over a, a list of topics. Uh, we can go top-down if, if you want, or where would-- do you want to sort of give people the context to, to get into the discussion? **Guest 2** [6:57] Yeah. SassyB, I don't know if you have any preference there, but I think we ought to start at the top. I mean, I think the top of the list is, uh, you know, we're, we're talking about managing AI budgets and token budgets, and I, I think that's like very, very timely. I think when it comes to sort of the whole AI ecosystem, it feels like 2025 is where a lot of companies, not even just sort of like the super early adopters, but a lot of companies realize, "Hey, we can really very materially change our output using AI," and they're using these subsidized tools, and it was like great, and it was good times. It was like, "Hey, $100 per, per, uh, employee, and I can get this magic box that answers some number of my questions." Um, I think 2026 is gonna be insane. It's already a little insane on that front because we're seeing a lot of the big players move from kind of subsidy-based, request-based pricing to straight up API pricing, consumption-based pricing. And I haven't seen a ton of content online on this, but I think it is gonna be one of the biggest changes in how people work. And we're-- we talk a lot about, there's a lot of angst about sort of replacing, you know, people with AI and stuff like that. I'm actually a little less worried that I think, you know, Javin's Paradox, all sorts of things, like there's a lot of things that people can do. I think the task and the skill of managing of a possibly fluid AI budget per person in your company is gonna be an incredibly difficult thing for managers to do. Um, the kind of thing that I've, uh, likened it to recently is like imagine, you know, if your laptop, you know, for some people cost $500 and other people cost $50,000, you know, per year or more, how do you decide who gets what? And I think most companies have a very simplifying assumption of most of the things that most people use are fixed price or, or very little volatility. And I think there's a, a kind of management, uh, slash sort of business operation, um, cliff that's about to happen this year with how do you figure out the answers to all of that? How do you figure out how much budget do I give a junior engineer versus a senior engineer versus a non-engineer for an AI coding tool? If a great, you know, senior engineer says, "I need, you know, double the budget," how do you evaluate that? How do you even think through how you would make the decision on that? And that entire kind of budgeting per person, I don't think that there's a single other thing that I can think of in my entire history of management that looks like that, that has that shape of having to real time somewhat- ... fine-tune and evaluate how much money you're willing to spend on individuals. And, and I think that that is sort of like a, a very, very nascent, but will be very fast adopted sort of area that people are gonna have to think through. **Guest** [9:39] I'll actually say the best comparison that I could think of is how companies manage more executive budgets or departmental budgets. And I think that in a way, this transition is sort of emblematic of the fact that once you can add all the AI automation, but that automation is expensive, people, you know, in, in... Individual independent contributor on a team starts to sort of take on the, the space within the company of what would've been like a small team or department before. Because companies were already starting to make different judgments like, "Okay, our budget for engineering is this, our budget for offices is that, our budget for IT, our budget for sales." ### AI Budgets **Guest** [10:20] You know, and those budgets are going to shift up and down over time. But now we're going to start to see that happening at like the individual employee level. And I think similar to how you would do budgeting across, uh, departments, we're going to have all the same sorts of variables at play, like what is the literal job that has to be done here? How efficient is this group or is this individual that's aided by AI overall? How business critical is it? How much are we able to do a really, really direct calculation of the exact value that we're getting out of the cost that goes in? All of these things that were already happening for large groups of people, they're now going to be happening for every single individual across an organization and, and that's a very different way to start to view scaling. **Swyx** [11:05] Yeah. It's, uh, scaling a non-human dimension, I guess. Uh, because you can be- eventually run your own department, right? Uh, if you are good at it. **Guest** [11:14] Yeah. Uh, absolutely, and I think that that scaling up is something mainly that's happening really, really quickly, and the speed with which it's happening, of course, i- is going to be very destabilizing because in a situation like this, I think you can make first principles arguments to start to change your approach in very different ways. You can make the argument that all automation is good, and we're, we're just getting efficiency, and we're just going to pour everything into it. You can also make more of a risk management type of argument that you need to be very, very conservative and that everything needs to be financialized and underwritten to the letter in terms of how much token spend we're going to allow people to use, and everything in between. And there's really just absolutely no precedent for this. And once you go into a world of no precedent, that means that, you know, some people are gonna be right, some people are going to be wrong, and the way that that shakes out in organizations, even down to like, you know, who gets blamed, who gets credit, that's going to be really, really interesting to see. **Guest 2** [12:07] The other thing is Budgeting at every company that I've ever been at is like very time-intensive and very controversial and very messy. And so the idea of scaling that to your entire employee base and thinking on an individual level, maybe as recently as, as common as every month, you have to kind of think about that question. I don't know how people are gonna answer that question right now. I think it's also gonna create the sort of world where building is much, much cheaper. It's gonna create really weird bottlenecks in businesses. 'Cause I think there are already businesses, and we're seeing this in industry, where they can build faster than they can distribute, they can build faster than they can get customers. And so what do you do when that happens? What do you do when the AI gets so good that you have an engineering team and it's like, "Listen, we could spend more tokens, we could spend more money here," but it's not returning to the business in a way that is fast enough and valuable enough. And, and that's another thing that as we're moving to API-based pricing, there's gonna be some teams in industry that have to face that reality of, "Hey, we could spend $100,000 to go and build more product, but we literally cannot turn that into revenue fast enough." So what do you do with that? Uh, and how do you keep high performers on a team if you end up in that situation? So I think there's all of these sort of butterfly effects that are gonna come into the situation when you start to get a world where, you know, an engineer can reasonably lock up a $50,000, $100,000-a-year AI bill. Like it's, it's not an insane thing to think could happen these days. **Swyx** [13:34] PM, I'm actually gonna tell you that you're going too low with your numbers. Uh, so one observation that I'll say is like, you know, the closest thing you have is, uh, you know, department heads, where you have to budget for department heads, right? Uh, well, uh, and then the, the other thing, uh, it... you know, researchers also have to do this, right? 'Cause the researchers have GPU, uh, research budget, right? So it's, it's like the engine- the research budgeting process is now transferring over to the engineering b- budgeting process and product budgeting process. But then this guy, uh, which we just interviewed, we haven't released it yet, uh, from OpenAI, is proudly saying we-- you know, he's, he's personally spending one billion tokens per day. The rest of his three-person team is spending like the, the remaining 1.5 billion. One billion per tokens every day, if you take like some blend of like input and output, is going to be roughly, let's call it, uh, 2.5 million. **Guest** [14:27] That is a lot. **Swyx** [14:29] A year, right? So like actually 50K is nothing. That's what you're gonna factor in as the fully loaded cost of an agentic engineer in the very, very near future. 50K, fine, you know, office equipment, whatever, but like 2.5 million is, is, uh, starting to, to really, uh, be chunky. **Guest** [14:46] It, it's really something you need to, you need to really think about from first principles from the start, like before you even decide to do something. And I think that's a... it's a very, very different world than the one that we've been living in for, for decades with software, where there was always this idea of software is expensive, but it's bounded by the number of people who you deploy on it. And also you have to get the people first before you even know what, what exactly it is that you're going to necessarily build. Because in a lot of cases, the, the effort to go and, and build all these things is going to be so high, and that's going to be like the determining factor of the risk that you're taking on. A lot of that's getting flipped on its head, which is, which is neat. And what I think is also really neat about that for companies that are scaling up is that whenever the playing field gets more complicated, your ability to really distinguish yourself go- goes up because it's, it's sort of like if we've been playing soccer or we've been playing basketball for decades and decades and decades, and then they change a third of the rules on you, whoever is the cleverest, whoever is the most agile, probably not necessarily the person who's most experienced in the old game, those are the people who are going to have those really quick advantages. **Swyx** [15:54] Yeah. Okay. Uh, I'm gonna m- transition topics very slightly because, you know, a budget on employees is one thing, but abstractly as a company owner, I am dealing with this absolutely right freaking now on build versus buy, where I can spend on an external vendor. ### Build vs Buy **Swyx** [16:12] Uh, literally I have an external vendor I'm paying for, uh, my conference business, uh, about 200, 250K a year. And I'm go- I'm now going, "Well, okay, if I s- if I give my engineer, give myself 50K to go build the same thing, it's more personalized and I spend less." Right? So like I could justify the e- the, the token spend, uh, budgeting the other way by also making a build versus buy decision. I know this is a sensitive topic for you guys, so I, I don't know, but like, you know, you guys clearly have some view about it 'cause you, you, you, you put this, the build versus buy thing out there and like I'm going through this right now, so I figured I would just promote it. Uh, also I think like budgeting is, is, is holistic, right? Like it's not just about, um, "Oh my God, my, my engineers are out of control getting LLM psychosis." They're trying to do something productive and sometimes, you know, the other way of accelerating things is to buy things historically, and maybe you don't have to anymore, right? **Guest 2** [17:04] Yes, and as soon as you go-- I mean, this comes down to a lot of the angst on, um, you know, software in general and stuff like that. And, uh, you know, I think that build versus buy is a question that's been happening in industry for a very long time. And, and I think a lot of the framing is like it's a new thing with AI. And I would say it's more acute with AI. There's more things to think about with AI, but it's not a new question. It's existed since the beginning of software and when people went to the cloud, that had an entire bunch of people saying like, "I have all these cloud resources and I can do this thing and not paying for it," all that kind of stuff. So I think one answer is that a lot of the old frameworks of thinking through this exist, and those frameworks are how complicated is it? How hard would we be able to manage it? How fast could you build it? Now, I would say like a lot of the times when I talk with people about products that they think, "Hey, I could just build this internally," one of the first things I always say is like, "Just describe what the features are." Just like- Tell me what you think actually the features are of this product. And what you find is like that usually curbs some amount of, you know, sort of like over-enthusiasm on building things. 'Cause you can't, you can describe the feature set and it's like not even... I'll tell you this, it's not even describe it now in a meeting. It's take three days and describe what you think is a PRD for this thing, and even with that, you come back with like a lot of like, well, it's complicated, and there's actually a lot of things it doesn't account for. Um, so I think one piece of the puzzle is I think AI is absolutely accelerating a huge amount of things and it's gonna, it is the, it is the future, it is the now, it is everything. It is still the case though that many products are just very sophisticated and have, you know, a, a large amount of complexity that people don't realize, and that's part of why you see everyone saying everyone else's job is gonna get done by AI, but not mine. It's part of the like, I don't know exactly what you do, but I think it looks easy. I don't know exactly what this software do, or I know this software, but I haven't thought through all of the things. I think it looks easy. So I think one piece of it is, you know, a lot of mature software is more complex than it looks. Another piece is administering software takes a lot of time. Um, having software that scales and has good uptime takes a lot of time. Um, and one of the things that I've just always kind of thought about as a leader is a lot of the software that I buy, I wanna have nothing to do with thinking about what features I wanna add to it in the future. I don't wanna PM that software. I don't wanna host that software. I don't wanna have to think about the uptime of that software. And it's all kind of, um, even if you could build it, it's of-often all fun until the first time you're like, "I have a huge crunch. I have incredibly critical, uh, you know, business things that I need to get done right now today, but somebody ran a DB migration on my internally hosted version of some software that I only have 40% of the features on, and all the new buyers are complaining that it's not like the thing I've used at the other place and nobody can actually..." So I, I think it changes the calculus a lot, for what it's worth. But I think a lot of the old reasons why you would buy versus build still hold up. And so I think companies that are going through this, I think just need to think about it carefully. I think a very good, for what it's worth, heuristic for some software is if you can replace it with a Google sheet or a spreadsheet, you probably can build it internally. And then there's a lot of software that actually looks like a Google sheet or a spreadsheet with a bunch of stuff on top of it. So that's my kind of first thing of like, hey, if you could literally replace it with an MVP with a spreadsheet, maybe you could build it. If you can't, now you're in this next tier and there's kind of tiers that go up and up and up. But I will say that we have one piece of internal software that, that I use that I'm just like, "Oh, I wanna just rebuild this thing 'cause I like, there's parts of it that I don't like." **Guest** [20:38] Do it. What's holding you back? **Guest 2** [20:39] But I tried, I... Well, I tried. It's complicated. It's more complicated than you think. Like I, I, I spend a good number of tokens trying to do it, but you're like, yeah, there's, there's stuff here, there's stuff there. So, um, you know, I think as the models get better, there will still be opportunities, but I think it's just people gotta be judicious about it and think about it. And I think as you try some of it, people also learn really quickly, like what the kind of m-more mature sort of, um, pathways of thinking about that and some things you can write off immediately and some things you can't. **Guest** [21:07] I also think there's something interesting with homegrown software, which is that homegrown software inevitably ends up very, very tied to the people that are building it, and there's a whole category of key man risk that goes along with that. And there's another whole category too, which is that if you really think about a code base, it's sort of like just a big collaborative document that all sorts of people are potentially contributing to. And in something like that, especially for something that's really complex software where there's really important things that it has to do, having more cooks in the kitchen is not necessarily a good thing. And that is the kind of thing that I think organizations really need to reckon with. But it's not a problem u-unless you've managed large software teams over a long period of time and had to own certain processes or certain systems for a long period of time, it's not something that you necessarily think about upfront. And so I think a lot of the discourse on X gets very, very heavily indexed onto things like, wow, I just typed a bit of cloud code and it just one-shot this thing that looks very, very impressive and the UI always looks decently good on, on the one-shots. And it indexes a lot less on things like, all right, what happens if there's a major reorganization of how my business does something, or we change some other major vendor or some other major system? How is all of this going to change? Because for better or worse, these companies, all, all companies end up as fairly interconnected webs over time, and that adds complexity. But on the other hand, I will say the barriers to entry on this certainly are dropping, that something has changed. This is not just business as usual and, and that's that fundamental tension that's going on right now. **Guest 2** [22:46] I think really good takes from you guys. Um, I'm happy to mostly leave it there. I, I will maybe comment two things or maybe one thing and I show you one, one other thing. I think the UI layer should be end user modifiable, right? Like I'm so tired of like having to hunt around and, and like navigate settings and what have you and like, you know, that's not harming anyone else. That's not risking any database data loss or anything like that. Like I do think like you can sort of chop up a product into different layers and some layers, uh, should be restricted to admins and others. Vibe code, whatever you want, I don't care, right? Like just take it away, right? The other thing is I, I just wanted to show you like, uh, you know, you know, um, uh, PM the, the, uh, the exercise that you said, like list your features. I, I just did that yesterday because I'm like trying to rip out the SaaS that I have. Um, and I'm like, I was like, "Give me three things." She's like, "Blah, blah, blah." And I, and I, and I'm like, "Okay, first one's a good one. Second one, come on. This is like a, you know, this is like a spreadsheet. It's like a retool. It's like an air table, right? Third is this is another syncing thing." So, so I'm like, "Okay, like if I prove to you these three things, we do it, right?" Like so, you know, maybe one, one way is like, okay, you know, maybe there's like 100 features, but there's three that really, really matter and have the most tech risk. So let's, let's try and do a proof of concept there. **Guest** [24:03] For sure. And I think that also the mo- the model improvements and how fast the models are improving, that's of course the wild card. It, it's like we're trying to figure out where the horizon is while we are actively in the middle of an earthquake right now. And so kind of knowing what is going to happen is, is very, very hard and very challenging. I will say on the, the custom UI point though, i- is that something that you need, that you need AI for? I'm curious what, what you're thinking there. Is that something that was strictly not possible? **Swyx** [24:31] No, it's, obviously it's possible. It's just, it's, it's nice to have, and now people are sit- are attuned or woken up to the fact that they can just prompt things. **Guest** [24:40] Right. **Swyx** [24:40] And so if you keep the back end the same, keep the data the same, if the UI, if you don't like it, you mo- move the menu around, like rearrange things or, like, have a button that does two things instead of one thing, does that really matter, right? As long as the, the, literally the, a- all the endpoints stay exactly the same, your permissioning system stays exactly the same, then you're just... Like, your UI is just getting in the way of your user because some designer just somewhere decided they had to design for the median and not, not you. **Guest** [25:04] Right. **Swyx** [25:04] So why can't, uh, me, especially for internal tools, why can't I just customize my UI, right? Like, all UIs should be, like, entirely personal. **Guest** [25:11] Yeah. That's not too far away also from some, what some mature products out there do where, you know, you've got custom dashboard. But certainly, yeah, the, the way the state of the art could go. **Swyx** [25:21] I, I mean, this is Retool. **Guest** [25:22] Exactly. And so but there's, there's a lot more room to run, and I think the economics of it, the economic model of that, of course, is completely flipped on its head with AI, which is one of the really compelling dimensions. **Guest 2** [25:32] Yeah. And to your point of, like, just there, there is this whole new suite of things where we have AI, and it's, like, incredibly frustrating that you can't just do a thing that seems like it should be so easy with, with AI. I think I had a similar thing the other day where it was like I had to log in to, like, a, a Google API project and, like, pull a key out of it, and I had to figure out, like, where in the project the key was and all this stuff, and it was, it was another one of these things where I, I do think there'll be more UI flexibi- flexibility. I also think the ability to do, like, any simple find and grab information from a dashboard, like, I, I think that just has to be agent compatible, like, this year for companies because that is one of those things where if you spend 10 minutes searching for a key that you can't find because the project layers are weird, it's just this-- It makes me want to go build something that, uh, replaces it. **Guest** [26:19] Yeah. And for what it's worth, I, I have a slightly hotter take too that in some ways AI might be really, really good for DoorDash and Uber Eats and Instacart. **Swyx** [26:28] You named DoorDash. You, you know, you know, you know what you're referencing. **Guest** [26:32] Oh, no, I'm not, I'm not talking about that. I, I was saying I think that this could be really good for, for any sort of service that makes-- th-that allows people to be kind of, kind of lazy because I think that AI has a very strong ability to just sort of reduce people's willingness to do work. It's like you could offload so much to it- **Swyx** [26:50] Yeah **Guest** [26:50] ... that I think that it-- we are training people in a certain w- for a certain set of consumer behaviors the same way that social media has sort of trained brains to accept or to expect a, a very, very heavy stream of dopamine. I think that we're ex- we're training people to not want to exert large amounts of cognitive effort on tasks, and I think that some of the, the winners of that, other than I guess, you know, people who are trying to write essays in, in college or whatever, are any sorts of services that allow you to just reduce the total cognitive load of whatever task it was that you were trying to perform. ### Automation **Swyx** [27:22] I think this maybe maps closest to the topic about just, like, a- adding value and getting, getting agents to be full employees. Is there a step? Is there a hierarchy that w- that you've sort of mapped out in your head? **Guest 2** [27:33] Is this in terms of... So can you restate it? Just in terms of trying to get AI to be as-- acting as much of an employee as possible and kind of- **Swyx** [27:41] Yeah **Guest 2** [27:42] ... doing? So CPM, I'm curious your, your thoughts on that first. **Guest** [27:46] Well, I think that there's really, at a very large scale, there's two directions where AI is really helpful. One direction is just the pure automation, like just doing things, whether- And those things could be writing all of your unit tests, or they could be generating, generating some image or generating some video. There's, there's a whole range of different kinds of things that can, that can just get automated away, and that obviously has a lot of economic value and has a lot of capacity to replace, like, certain tasks that were being done. But then there's a whole different category around, like, making decisions, so things like what should the overall sales strategy or go-to-market strategy be for a company, or w- how should we change our supply chain to account for some sort of, like, unrest, uh, geopolitical unrest or something like that. And, and that's completely different, and I think that the, the models have given us a certain way that we could think about that between the reasoning models and the, the models that aren't doing reasoning. But just thematically, there's a really big difference in terms of the overall types of economic value that are being performed. And at least what I see right now, just out and about in my life from seeing what people are doing, I think that generally speaking, people are still quite focused around the automation and doing tasks side of things, which is in a way very, very similar to prior kinds of, um, technology changes in the past. Like, once you get personal computing and y-you can use spreadsheets and things like that. The question of sort of like judgment and decision-making, that's a really, really different one, and I think one of the ways that this can start to manifest, uh, that's kind of interesting is this question of, like, what's gonna be most efficient? Is it going to be to have an AI that tells humans what to do, and then the humans can kind of use- A smaller level of judgment or where the AI is the router and the orchestrator, or should this be flipped where the humans are the router and the orches- orchestrator, and the AI goes and does all of the, the manual tasks? Which is gonna come first, and does it all eventually become machines? But I think that that kind of continuum in parallel, like how good is the automation and how good is the, the decision-making and the, the strategy setting, those two factors are really gonna determine kind of how far for any one type of role or any one type of, you know, agent that can partner with people, how much trust it ends up getting, uh, from an organization. And I think that as a result, what, what happens is that organizations need to be really, really thoughtful about the level of quality of decision and the level of quality and perfection of automations that they need for every single task across the organization. Historically, I think a lot of companies have not needed to think about this. It's just like I hire the best person or the best person that my budget accommodates for every single task that I can find. But I mean, going back to the economic point, that question of what is good enough and what is perfect and where does that fit, that's really gonna determine how much and where the reins get handed off. **Swyx** [30:48] I've been plotting a little like, well, well I like to think visually and just plotting a little like, you know, when, when I asked for a hierarchy, I was like, it's kind of like, like this where, you know, there's like levels, levels and, and you sort of like step through it. I, I don't, I don't know if we have one, I'm just kind of plotting it out. You said AI to human and human to AI. I also think of like AI to human to AI, where like basically that's human in the loop, and you can use AI as a cheap drafter, uh, and the human judgments kind of just apply there, and then, and then you can sort of kick off e-everything else, right? So that's kind of how I would respond quickly to, uh, PM's point. ### Executives **Guest 2** [31:20] Yeah, I will say I, I think one of the interesting things, I don't know if the SaaS PM said this explicitly, but a lot of people think about the org chart that, you know, you have a tree and they think about I can start to automate leaves of the tree, you know, sort of like the most junior bottom rung of sort of your organization. And that's a very like classic kind of, kind of bullshit executive thing where they're like, "Oh, I'm gonna start to like, you know, improve my, you know, whatever." But I think I heard the SaaS PM say this, of like there's sort of like replacing automation that maybe you have most junior people and stuff like that on. And I think we've seen companies try to do that successfully. We've seen companies try to do that and fail. And you know, sometimes those employees are literally talking to your customers and they're the face of your, of your company. And people are, I think, in industry sometimes like misunderstanding the value that that actually has. But I think you have sort of like intern, full-time employee and kind of, yes, ending what the SaaS PM said, I think actually I would say like, what about executive? And I always say like from a leadership position, I don't think I do that much complicated stuff, and a lot of stuff that I do is like very, very, like I wanna be part of the pack. I don't want to do-- I don't want every decision to be weird and exotic. And so I actually think that if you're looking at, you know, where you can automate things in a co- in a company, a lot of people are looking at the bottom of the tree. I would start to think at the top of the tree, and a lot of leaders and a lot of executives are bottlenecks for their entire organization. A lot of what they're doing, 70% of what they're doing should be things that are pretty standard in industry. And a lot of what sort of they end up trying to think about is like, "How do I do the thing that everybody else does so we're all on the same page?" Why not automate that with AI? I've personally automated a good amount of decision-making with just decision trees that's not AI at all. It's just literally like I only do five things when I do this specific thing in recruiting. It's not that complicated. You know, one of the things that I always coach my managers on is like people say, "I can't delegate that. It's too s- it's too complex." And I work with them like, "Write down how many things you actually do here." 'Cause most of the time people think too complex is actually like a case statement with more than five things in it. They just like, it's you do maybe 21 things, and that's like complex, but it's not that complex. It's not undelegatable. And if you look and you back test against everything you've done in the past five years, I personally have many things where I've done one of 15 different options, but most people just can't get past, "Hey, I... It's not for, you know, so it's, it's too complex." So I think actually maybe one of the most interesting things is companies are looking at sort of trying to automate away sort of your more junior staff, but like look at your executives and maybe there's a world where in like, you know, the next year or two there's fractional, you know, executives, maybe there's AI executives, and you, you're the CEO and you say, "Hey, actually for this role, I want 80% of things to be standard and the 20% that I want to be exotic, like I can think about that for first principles, and it's like I can hone in on that much more easily than I can actually have quality control over 500 people in a location that's at the leaf of a tree of an organization." That's something that I think as we talk about that sort of evolution, I think the technology is already there, and I think leaders should be thinking about that, of how I make myself less in the path of, you know, decisions and bottlenecks and all that stuff, especially where it's kind of bog standard. So anyway, all that to say, I think the tech will get better and everybody's job will be able to be sort of more automated and it gets... Then they'll take on new tasks. But I think many companies are sort of not thinking enough about how do I get my executive team to be more augmented, supplemented, or in some cases even sort of replaced by AI. **Guest** [34:51] This also goes to a post that we had from many years ago. I think it was called You Know What to Do. And the inspiration for that post that I wrote back then was basically just saying that if you really look at most of the questions that are facing you as a manager, as a leader at a company, from what I've seen, 99% of the time you, you know exactly what to do. The number of truly, truly difficult questions that require you to know some sort of advanced theory or advanced philosophy or run some huge amount of calculations that are very non-straightforward to do, that is the extreme minority of different questions that actually face you as a leader. In most cases, you know what to do, and it's much more about having the willpower to go through with it And I think that a lot of the power of AI for that sort of decision-making characteristic of leadership is about the fact that the AI, well, at least right now, depends on how many Anthropic blog posts you read, but at least right now, the AI does not have any emotions, and it also really just tends to know the standard way of looking at a problem. And so in a lot of ways of what I-- the way I could see this shaking out is that the role of the human is to identify, okay, this time it really is different. This time is not just that amalgamation of reading the entire internet and, for better or worse, reading every single Reddit comment on the Internet and then using that to generate your version of intelligence. This is a unique problem for me, but in many other cases, you can get all of that time back, and you can get all of that cognitive energy back. And I think that's actually going to lead, in many cases, to better leadership, to better management decisions, because the standard questions will get answered in a more straightforward way and those really hard and unique problems that are facing different organizations, AI will have solved all of the easier peripheral questions around that, and people can really, really focus. I mean, if you think about your personal life, if you're thinking about questions like, "Who do I marry?" or, "Where do I live?" It's like the really deep personal questions. They could be exhausting to deal with, but they're very important. And when it's in your personal life, you can really invest all of your energy into that. I think that there's a huge amount of power in that, and if that gets brought sort of into the workplace and into management of teams, I think that that's very, very valuable and, and just a net positive overall. **Swyx** [37:10] I love the sort of point-blank and blunt way of doing this. I, I feel like y-you guys are almost like management coaches here, which, uh, uh, I mean, you don't do as a day job, but you could be. In some ways, you, you, you are doing through your Stay Sassy, uh, work. **Guest** [37:24] Yeah. Well, that was the genesis of the blog actually, was-- I actually remember where it was. We were having some meetings, some touch-base meeting with-- And afterwards, we were saying, "Wow," like "This is def-- Like, I'm glad we talked about this, and we need to go do something with this, but also, I really wish that there had been someone who could just tell us what to do in this situation because- ... we're sort of guessing, and I think just from debating, we've gotten to a good conclusion. But this was knowable beforehand by somebody with more experience." And then we were starting to look back at some of our past experiences and say, you know, "We should really just start to write more of this stuff down, if anything, just for, for our future selves." Because to that point, a lot of this stuff is knowable, or at least the principles are knowable, and then you have to take them and apply them to whatever situation you're in when it comes to management and comes to leadership. **Guest 2** [38:13] Yeah, and I think- **Swyx** [38:13] Yeah, yeah **Guest 2** [38:14] ... one of the most interesting things that I've seen in the modern era that has been surprising has been all of the things that we learn in series A, series B, you know, all that kind of stuff, like it's super applicable to modern AI companies. Like I, I kind of, as I started to s- talk to more leaders of mo- you know, sort of like the latest and greatest AI companies, I thought I was gonna hear crazy different things and just sort of like, "Hey, it's wildly different than anything you've ever seen, and you can't even comprehend it 'cause you're not at one of these state-of-the-art, absolute, founded eighteen months ago and is, you know, a trillion-dollar company." But I talk to the people, and they're like, "Hey, I'm at like a Series B hot AI startup," and they have the exact same problems we had, you know, when we were at a Series B, when I was at a Series B company, and they had the exact same problems as at Series C. And so it's like, uh, it is, um, it's incredibly interesting how many of these problems are durable through technology shifts. And I think when we think about like what AI can automate and what can't, one of the most common things that is extremely consistent amongst all of these companies is we're dealing with humans. It's people. And I think one of the places where good leaders are really good at is just thinking about people, thinking about how to manage individuals and humans. And, and I don't think, even though I think like, "Hey, how do we like do outbound recruiting?" Like, that's a very like automatable task with principles, but how do I manage a team of humans, and they might have off days, and how do I build their trust? I don't think that AI is particularly close to that. And I think that that is like probably one of the things that is most consistent amongst any company that we see, problems that are in that realm. And it's also something I think AI is probably the farthest from. And it's also one of these weird things where even if you started to have something like an AI as a manager and like your manager is an AI, people will start behaving differently. And so it's like, it's one of these Heisenbergs where like the minute you even try to do it, you sort of, you, you lose the race and stuff like that. So I think that will also be interesting. And, and I, I really think that in the next two years, we will see the first company that tries an AI manager. They won't try an AI executive 'cause I think that's like a whole different can of worms, but I think we'll see an AI manager, and I think it'll be weird. And I actually don't think it'll work particularly well at all. **Swyx** [40:18] My sort of side business is, uh, small AI, and I have bought the domain small.ceo for specifically this reason, because I don't want to, I don't want to have people tell me what to do. It's nice. **Guest** [40:28] What will also be really interesting there is going to be how some of the societal customs around this stuff come to be. Like, for example, I could see a world where people start to say, just as an example, something like, "AI therapists are better." Like I, you know, I would rather talk to an AI about my issues and have it, you know, be very, very patient and go as deep or as shallow on any sort of, uh, thing that's on my mind or thing that's bothering me that I want. Like, I could see that being something that's very, very popular. But as the sassy EM said, I could also see a world where being told what to do by an AI feels terrible. I mean, I, I don't, I don't want that. And even-- And I would probably even be willing to tolerate a certain amount of lower quality management coming of me as long as it was coming from a real human. Like I, I strictly prefer that the same way that I don't like, I, I don't like Impossible Burgers because it feels weird, and it feels like a bunch of chemicals just intellectually. **Swyx** [41:27] I love the Impossible Burger analogy. It's just like, you know, you have the Impossible Burger manager, you have the Impossible Burger en-en-engineer, and it's like not quite the right thing. But like, you know, someday. **Guest** [41:37] Yeah. But maybe it gets way better because also if you ask me in a pure sort of, you know, caring about the world and the environment sense, do you like the idea of Impossible Burgers? I say like, "Absolutely. I think we should totally... Like everything should be an Impossible Burger." And maybe that's the end state, but that's far away and, and there's a certain amount of work that the whole machine needs to go through or an adjustment the whole machine needs to go through if we wanna get there. And that's gonna be, I guess, the journey of, the journey of the next few years or next few decades or next few months, depending on who you ask. depending on who you follow on X **Swyx** [42:10] Or it, it could be, uh, you know, 2028. ### Code Review **Guest** [42:12] Yeah. **Swyx** [42:12] Okay, so one thing I wanted to touch on before, uh, we close. I, I don't think we're gonna cover everything, but I just, I needed to do this because, you know, you, you guys actually work at, like, a real company, and it's, like, very serious and, and all that. So, and I, I do think this week it's particularly of focus because Amazon went down for, like, what? Six hours because of vibe coding, right? And they were like, "Oh, this is..." They, they threw their own, um, software tool under the bus, uh, because it, it was actually apparently attr- attributed to Hero. And, uh, I'm like, wow, if, if that hits Amazon, like, we might be in trouble because Amazon's supposed to be the best at this. I don't know. Just, like, reliability and criticality, I just wanted to prompt you and set you off. **Guest 2** [42:54] Yeah, I think it's an extremely interesting space, and I think ultimately companies trust engineers. Like, there's no system that doesn't have some trust that engineers are the checks and balances for what goes out to prod, and that's just the reality. There can be AI, you know, progress. There can be AI CI and all this stuff, but any good system still has some level of judgment in it. And I think one of the things that has become very challenging is that as the level and the volume of code going out has exploded, it's coaxed some companies into thinking, like, "I don't need to look at this PR. You know, that's that person's thing. That's their kind of code going out the door." One of the challenges of that, and this is, again, this is not a new challenge in software, actually. One of the challenges as you scale a software business is you get to a place where you have an incident, and you go, "What happened?" You go, "Ah, it's the, it's the new guy, you know, who didn't know what he was doing." **Swyx** [43:48] Totally throwing under the bus. Like, it's... But no, actually it was you. **Guest 2** [43:52] Well, you know, it happened, but, you know, you then go like, "All right, well, we have some new guy budget for, for issues," but, and then you have five teams. Now you have people like, "Oh, it's the new guy on that team, the new guy on this team." So there's always this possibility in, like, a scaled software system where somebody who doesn't know what they're doing could break something, and as you scale a software company, you need to be clear with your teams. Like, there's no world in which somebody that doesn't know what they're doing can sneak through the cracks and break prod. You as a team are responsible for making sure that things stay safe, whether you have a low performer or a new person or whatever. And companies that have been successful and stable, they have that culture. They have that culture of I could literally put a malicious actor in your team, and they couldn't take down prod because they have to get a poor review from somebody that knows what they're talking about and want-- And, you know, you have to be in a place where, like, that is a check and balance in the system. You would need multiple people who don't know what they're doing to actually coordinate taking down prod in, in good software cultures. But what I think is happening is people going and saying, "We have to ship so fast. We have to ship so much AI that I don't even know what this PR really is, and e- even if I did, I can't do a risk analysis on it, and I don't know how to my good staff engineer to review that PR." So that whole invariant of the team is here to... And by the way, new people hate taking down prod. You shouldn't set them up to do that. You should have the team be a cocoon of safety of, like, "We got you. We will make sure that you get it out the door." And what we're seeing with AI is many companies and cultures are turning into armies of one. This idea of, like, I can do it all. I want to ship as much code as possible. And by the way, I, I don't know that I want to be here to be that person's, you know, code reader for them. And so what I think is happening is companies are trying to adapt in real time of how do I not unblock people that say they can and, and show that sometimes they can actually get it out the door, but then how do I make it in a world where somebody who can't is not just left to take down prod because nobody will take the time to actually review their code. I think it's a solvable problem, but I think it is testing the cultures of companies, uh, in real time. They're gonna have to adapt to it. I think AI can only go so far, and I think coming back to sort of, like, human and, and human judgment, uh, you know, keeping software sites up with many nines of, uh, availability requires judgment constantly, and that is something that I think the entire industry is gonna have to adapt to. So anyway, if you're out there listening to this, review your coworkers' PRs and, like, you know, don't let anybody able to take down prod. Like, if that is the, the culture on your team, you cannot do that. It's not sustainable. It's only gonna get worse with more code going out the door. Um, but, but I think that's what's happening is people are becoming fatigued and kind of breaking that invariant of, like, the team owns it and the team is stable. **Guest** [46:31] Yeah, I think for one thing very telling that this was Amazon, so a very, very large software b- company, one of the biggest software companies in the world that also does a lot of things that really are business critical and supports... Yeah, you know, of course, they have AWS, which is the business critical system that other business critical systems are set up on. So I think it's, it's very telling there because that's really showing the level of intensity that of pushing to use these new coding tools to ship as fast as you possibly can that's, that's going on out there because i- if you would expect anywhere to be more conservative and more risk-averse around this stuff, you would expect it to be a, a company a bit more like Amazon. I think that the other element of this that's going to be really interesting when it comes to code review, and particular code review of really nuanced functionality that's really critical or is very deep into the bones of whatever product it is, uh, that your company builds, is that this new sort of mode of working that you, of course, you read about on X all the time or increasingly on LinkedIn, but LinkedIn's what? Like, 60, 90 days behind X. Someone ran that, uh, study one time. **Swyx** [47:42] Oh, that's a real thing? **Guest** [47:43] Um- **Swyx** [47:43] Damn. **Guest** [47:43] I, I, I, I don't know. I'm making it up. **Swyx** [47:45] The joke is one week. **Guest** [47:46] One week? Oh, man, I would not give LinkedIn that much credit. It's at least four weeks. **Swyx** [47:51] So something goes viral on X, people go like, "Oh, LinkedIn's gonna be crazy when they find out about this next month." Like **Guest** [47:56] Yeah. So but the thing that people talk about is, all right, so I'm an engineer. I'm a cracked 10X engineer, and now I'm a, now I'm a 100X engineer with AI. I'm running 10 simultaneous agents all at the same time. One of the things that people quietly talk about is that that's That's very fatiguing, and I know some people who are operating that way. I've, I've met these people, and it, it is fatiguing. And what is fatigue really going to hit you on? Fatigue is really gonna hit you when now it's the end of the day, you have been y- tending to this, uh, herd of agents that you, that you've named and are, are chatting with constantly, and now someone from another team hits you with a pull request that's very, very complicated, that touches many different systems, where you need to basically page all of this context into your brain in order to review it. That starts to become very risky. Like, to somebody who is, you know, accountable, at least partially accountable to a system being sustained, that's very risky to know that somebody who's that exhausted is going to be reviewing really mission-critical code. That is something that I think we're going to see over time. I, I could even see crazy worlds where it starts to become like a pilot's license. Like, I can't have a beer for X number of hours before I fly a plane. You know, I need to have a, you know, Y amount of sleep, or I can only be in, in the cockpit for- **Guest 2** [49:19] Ah **Guest** [49:19] ... for so long before I sign out. **Guest 2** [49:21] Yeah. **Guest** [49:21] I wonder if you start to see something like that going on, if you're going to be the reviewer of, say, that 5,000-line, uh, merge request, pull request that's coming through of something really critical. **Guest 2** [49:31] I think it's either gonna go, like, PRs are totally automated review, and literally there's no human in the loop, and the, the tools are good enough, or the entire industry is gonna get to, we need to have pull review be, like, a serious thing that we're not taking for granted in the age of AI. But we'll see. **Swyx** [49:46] A couple things, uh, which we are now running out of time to talk about, but, uh, we did just publish a blog post on Latent Space about how to kill the code review, and you're s- you're saying do more reviews. I do think, like, there's different parts of the adoption curve. Extreme frontier edge right now is thinking about what people are calling the, uh, software factory or the dark factory, where you not only have no human-written code, you also have no human reviews. And it's, it's very alarming, but they, you know, there's people setting up systems to do this. I don't have the time to to, to dig into it, but just people should know. Your, your comment also reminded me of Netflix's Chaos Engineering and Chaos Monkey, which like, yeah, if you have a malicious actor, you randomly turn off services, is your organization resilient? And like, yeah, y- uh, people are accumulating a lot of key man risk when it comes to sort of the, the way that they approach AI engineering today. And, uh, I think, I think you guys are just, like, very sensible, which is applying EM and PM advice to building companies in the age of AI, and I think that's, like, really well, uh, w-welcome. Any last words before we, we, we, we end? Uh, we, we do have to go soon. I'm just kinda leaving it open in terms of topics. ### Closing **Guest** [50:52] I think one thing that I would just leave... Well, well I guess I have to say the, the Mr. Beast YouTuber line, right? Like, come and like and subscribe to our blog. **Swyx** [51:01] And I really wanna shout out your podcast as well. **Guest** [51:03] Oh, yeah. **Swyx** [51:04] Uh, I don't know if you guys are continuing, but episode one, "Winning from Third Place," I was like, "Holy shit, what a title. What a focus. You guys have taste." Like, yeah, go do a... check it out. But yeah, the- **Guest** [51:13] Yeah. Thank you **Swyx** [51:14] ... primarily the Twitter is, like, where, where you guys are blowing up, right? **Guest** [51:16] Yeah. I think that, like, the... Ultimately we are, I think at its core, a management blog or a, a management voice. And I think that the number one thing, maybe two things that I would say, just based on the, the reader feedback and different audience folks who've, who've reached us, to us in the past are, are two things. One is if you are a manager, it's just really, really important, I think, to really care, really be compassionate, and really care about the well-being of your team and the well-being of your, of your team in its entirety. And also, to anyone who's on a team, I would just say that what we consistently see from people who reach out to us directly is just how much many people care about their teammates. I think that that's been one of the coolest, really optimistic things that I have gathered from the time that we've been running Stay Sassy, is just the number of people who really, really care. That has just been really cool to see. The other thing that I would just say is that building startups, even with AI, even if you have Claude Code, building startups are really, really hard. And what we do see is the problems and the challenges that different startups run into, different founders, executives, you know, every engineers, anyone on a team that they run into, those problems really do rhyme. And so if you're feeling really stressed out by some business problem that's going on, just, just know that you're not alone. If there was a single kind of sentiment that I could get out there, it's really that. It's that, like, you know, don't worry too much. Someone else is dealing with this probably in exactly the same way in some other location right at this moment, and, and just don't be discouraged. **Swyx** [52:49] Very much, like, that's also my, my goal is, like, this is a shared experience. We're all going through it together, and like, uh, that's why I write, that's why you write. Uh, I think it's a, it's a common thing. **Guest 2** [52:58] Yeah. I would just say, uh, I think 2026 is gonna be the most dynamic year in software, like, in my entire career. I think the industry is, is, uh, really ready for a shakeup, and I think I, I heard a lot of people in 2025 say, "I feel like I'm late to this stuff." And I think my only thought is, like, I absolutely think nobody's late to anything. I think this year is gonna change every single thing that we thought about software engineering and running businesses and, and, uh, and I think the people that are ready for this year in particular are going to go into the outer years and just say, like, "This is where I made everything happen." So I think from a manager perspective, it's be ready, be ready to adapt, know yourself, know your limitations, ask for help, do the right thing, you know, all that kind of stuff. And from, you know, sort of a, a builder perspective, I think for a lot of people, the, the hardest problems are gonna be human. It's gonna be figuring out how do I work with other people? How do I make sure that, like, I'm not aiming at the wrong thing and then getting, you know, just sort of like building in a bubble and stuff like that. And so I think it's gonna be the craziest year for, for AI, and, you know, if that's, uh, if we're at the singularity, I guess that'll be true for every year from here on out. But, uh- ... but, um, I do think it's something where, like, it's gonna put a lot of pressure points and a lot of human different interactions, and I think everybody... There's a lot of, like the Sassy kind of said, there's a lot of prior art on that. There's a lot of things that you can look up about how to navigate that. But I think that is gonna be something where Everyone needs to take care of each other, think about how they can execute in a world where the AI is not taking over everything yet. There's a lot of people out there that you need to work with to get-- build great businesses, and I think people need to embrace AI fully if they haven't already, which I don't think anybody on this pod listening to this is gonna be, uh, guilty of that. But I think people also need to really not veer too far from dealing with human problems because those are, I think, where people are gonna also differentiate this year. **Swyx** [54:46] I'm both an optimist in like we will just invent new problems once the old problems are solved and also a pessimist in like, well, that's just because we are never satisfied. We keep moving goalposts, right? Like, and it's the same thing with, with what AGI means, what AI is. Like we're no longer impressed by GPT-4, which was so impressive when it just, just came out, right? Like, so like, yeah, you know, have faith. **Guest** [55:07] I have a slightly more optimistic take on that though, which is that I, I think that hedonic treadmill is very real. But the thing that is the saving grace on this, I think, is the fact that smartphone cameras, well, well really the age of the internet, but also especially smartphone cameras recently mean that history is now being so completely and exquisitely documented all the time that whenever there's a major lifestyle upgrade, I think people remember now in a way they didn't. I think that something-- a sentiment that I now see a lot more of versus when I was a kid, and maybe I just wasn't paying as much attention, is this idea of, you know, the example people sometimes give is if you can take a hot shower, that's something that if you were literally a king, if you were royalty 500 years ago, you would never have an 11-minute hot shower. And you know, I have an 11-minute hot shower every day, 12 minutes even. And so I think that there is a bit more reflection on the past just because since entering the information age, the past is so well documented. I can see the cars that they were driving in the '30s. That looks terrible. I don't wanna drive that car. I can go and hop in my Prius and I'm like, "Yeah, this is way better than that stuff. This is way faster. This is way more fun." Which I think is a good thing. It's, it's the same way that I think in some ways social media has made like the youth more compassionate just because you meet more people or you see more people across the world. So I'm very optimistic on that one actually. **Swyx** [56:27] Yeah. Yeah, totally. Totally. I had an example the other day. It looks like it was deleted, but I'm just gonna show it here because this is a, a prime example of how documenting things do-- does help you remember. I just think that most people have the attention span of a goldfish, so they don't remember anything. But- **Guest** [56:43] Yeah, that's, that's probably true. **Swyx** [56:44] But yeah, this is an example of like, well, it's easy actually to look up right now, 2016 versus 2026, and actually everything Apple costs lower. You know, you, you make fun of the Ma- the MacBook Neo, but actually it's a way better MacBook than you used to buy for $1,300 in 2016. So what are you complaining about? **Guest** [57:01] Yeah. **Swyx** [57:01] Anyway, it's a pleasure to talk to you guys. I'm a fan. It's always a pleasure. We should do this more often than once every four years. But like keep doing what you're doing and, uh, everyone go subscribe. **Guest** [57:13] Yeah. Thanks so much for having us on, swyx. **Swyx** [57:16] Yeah. Yeah. Thank you. --- This library is powered by PodHood (https://podhood.com), the podcast website platform.