The Future of Software And Data Engineering Teams (AI + Junior Problem)

And we are live. Hey everyone. Welcome to the Seattle Data Guy channel with me today. We have Dorian aka code trap for people who know him online. How are you doing today, Dorian? Doing amazing, Ben. Yeah. It's Friday. It's the end of the well, I think you're on West Coast so it's you got a little more but getting close to the end of the day. Yeah. And no no meetings tomorrow morning. I get to like chill out and have some coffee. — you're still going to be working man. Knowing you now that I've You never stop working. You know this being here too. You never stop working. Yeah, yeah. I'll get messages — they call that? 996? I don't know. You do like 997. Uh you're like I'll get messages from you on Sunday sometimes and I'm just like I mean one day off. Uh this 996 thing sounds like a vacation at this stage so — yeah. Um so I mean just I think maybe some people know maybe some people don't know. What is new with you? What are you currently what are you trying to solve uh on the in the tech world? I mean the same things you are. — Yes, yes. So like I think we're all kind of going around this uh merry-go-round of um agent harnesses and how do we leverage AI coding and all of this sort of stuff so um you know I think that's right now we're all kind of fighting the same battles and also like try like my the biggest thing I'm also interested in is like what is the future of the junior software engineer and um this like labor displacement kind of theory going on and like all of that stuff. Heavily focused on that as well. Yeah, well I think Meta just did like 8,000 uh laid off like 8,000 people and Microsoft's trying to push for 7% or at least offering 7% or something. Um what do you call it? Severance. So yeah, there's a lot of shifting going on. Um yeah, I mean you know there's definitely a tension, you know I I will say I'm like I'm busier probably than I usually am so I don't know uh it's one of the things maybe I'm just fortunate to be busy. Um but yeah like with this tension for engineers kind of getting replaced by AI like where do you think it's going? Where do what where do you even think like AI actually does replace engineers today? Um yeah, I mean I don't think it replaces engineers at all. I think that maybe in the short term people are experimenting with this idea. Like we saw um what's that company? Block laid off like 4,000 people. I think about 35% of that were software engineers. They said that they were using it as a forcing function because they were seeing like internal AI usage decline among the engineering teams. Yeah. And so they're like oh, let's get rid of some of the engineers and see if they use AI more. Yeah. But I wish they would have stopped and asked like well why are the teams not using AI as much and maybe it's because it's not capable of doing some of the things they want it to do, you know? Or maybe it like to your point like um there's a certain thing it's good for but like the amount of calories you burn to like use AI to do something might not be useful in a lot of day-to-day coding exercises, you know? Yeah. Uh so like my prediction is essentially where I see it going is it's more of an enabler for certain types of workflows. It can help in bug triage for example enormously. It's very good at apparently finding zero-day threats, you know? It's like the way I look at it is like it's extremely good at looking at a lot of information and digesting that in the context window, you know? Yeah. Um however, it's not so good in out of the box and guided workflow experiences. It's not so good out of the box when I need to make very focused changes. It's also very good at creating lots of bloat in the code base. Um so like the role of the engineer needing to reduce software entropy and manage complexity growth has never been more important in a lot of ways. Yeah. Yeah, so like I don't see it replacing engineers. I see it being another tool in the tool chest. We're all sort of figuring out the best way to apply it and like what does it need to sit on top of and that sort of thing. Um so yeah, I don't see the role going away. In the interim though there's going to be a lot of this experimentation like what Block is doing to see like what works and what doesn't work. Yeah. Um and so you may see like oh these people eliminated some roles. These people did not. And these IBM is leaning into this so like um I don't know. It in the short term it may look like the roles are going away but I think actually in reality if you look at the capabilities of the technology, there's just going to be a lot more software in the world and there's going to be a probably a lot more software engineers in the world. Yeah. Yeah, I do I did like when you brought up the fact that like it helps just get through information faster and like that's definitely something I find myself using when I'm using like tools that maybe are either unfamiliar or just especially any tool that's like click ops like and you there's like a gazillion things you can do. Like sometimes you're like I know how to do this in code but how am I going to do this in your tool that's you know one of a thousand and how do I find this path? Especially some of these tools.

I'm not going to name any names. Some of these tools like they you got to like well and if you go into the VS code version of it then this thing works but if this other version it doesn't work and you're like I or actually you know I will Snowflake I think had this problem for a while with uh Snowsight and doing like if I wanted to share um like just a sheet that I had. I think I had to go into like one of the specific UIs but it didn't wasn't in the other UI. I'm like I the share button's right there. Like why does it work for one and not for the other? Totally. I the agent I want is preference management agents. You know cuz like VS code, right? How do I change this thing over here so it does this thing instead of this thing, right? And it's like — Yeah. well go into the settings and then find this thing and then go over here and then filter by this and then click on this other thing. It's like can I just say what I want my VS code experience to be like and it can go configure the thing for me, you know? Like Uh that problem exists I think across every piece of software we use is like the preference management problem. Yeah. Uh and then the click ops problem is like yeah, I think any platform that heavily relies on that UI layer to drive the user experience or like the outcome you want to see. That's one that's been particularly troublesome because it's like to your point that the way the product team sort of like where do I put this new button becomes an issue. So it's like logically it's just whatever. How do I find it? AI doesn't need a UI either, right? So like it's interesting about AI is like that for a little while it seemed like they were trying to go down this like uh what was that thing that they were doing a little while ago where they would like build automation to drive really shitty software. It's like UI path that did a lot of this stuff. — Oh, yeah, yeah, yeah. What is that? Robotics process automation or RPA? — RPA. — Yeah, yeah. For a little while it looked like they were trying to make AI into the next gen RPA, uh you know? Oh look, it can use your computer. Like no. First of all, I don't want to use my computer. have access to that. The other is that's just super slow. Yeah. Right? So like getting AI to drive UI is the slowest way to do things. If you reduce it to tool calls it can perform and orchestrate the API back end, it can do things way faster. Right? So imagine like in the future you don't need to remember that button is or like remember the preference setting. You're just like typing or you're just saying like hey, can you do this for me? Send this email to Bob or can you please you know run my job or tell me what the status of the pipeline is or whatever and the agent could just figure it out, orchestrate the tools, come back with the answer and at a certain point UI collapses. Like all UI just kind of collapses into chat or something like that. — Yeah. And that's what you know Salesforce last week they launched uh Headless. — Headless 360. — headless. Yeah, something like that. Uh — Yeah. Yeah. Yeah, I think that makes like it's like again you kind of give these tools you don't have to see it. It makes me think of like when I would if you're like printing out um steps as you're coding like especially if you're doing like lots of data processing. If you print out it takes like 2 minutes. If you just let it run it's like done in 5 seconds. It's like oh, I don't have to print it out anymore. I'm just going to process it and then it's going to be done. Um so yeah, like getting rid of any form of UI I can see that happening. Uh for like mediocre teams how do you think or you know just assume you have like the average team which you know most of us. Uh what actually improves and what actually doesn't like when you give them AI? Um I would say like AI is an amplifier, right? So like if you set AI on top of dysfunction, it will amplify dysfunction. And you set it on top of something that's highly functional, it tends to amplify that. So I don't think there's a universal answer here. I think it's more relevant it's more like how good is your team already and — Yeah. I think also what helps is like patterns and architecture repeatability. Like a lot of good high-performing teams bake that into the cake because they're interested about like deploying frequently and deploying stable change sets and like high-performing team kind of characteristic. So like if they're doing their job of like good architecture, good design, uh ensuring standards and practices are followed, they have good amounts of test coverage or I don't want to say test coverage but they have the right test that kind of can predict future behavior, um AI can totally amplify that success. But if like the team is all operating different ways, they've all implemented different patterns, they've all they don't communicate very well, maybe they don't have good visibility when things break or they don't know when things break. Like well, what is AI going to do? It's going to light that on fire, right? Because you're going to deploy more, you're going to introduce more change, you're going to have higher context boundaries for both humans and agents to get around. So like it's mostly an amplifier at this stage. Uh so what is it going to amplify is a better question I think at the end of the day. I I like that view because I've I think I want to say maybe it was you that brought up this, right? Like it's so freq or it's so common that you often have like three, four, five, six different kind of coding styles at a company. Um and then if you just stick your AI on top of it and then it kind of almost like starts producing even more of its own versions from there cuz it's like oh here let me

take a little bit from style A, style B, let's make a new version. Kind of have little essentially a child of of a different or of these two different kind of code bases. So Yeah, I think I like the idea of like if you can set it up correctly in the first place and just keep it going and then making sure it doesn't go too far out of there too frequently. There's some benefits. — that's a problem for humans, right? Like a lot of people don't understand what that problem is you just explained because it's like If you have a bunch of people like let's just say you have multiple teams. I've got three teams, right? And maybe there's three people on each team and team A programs one way, team B programs another and it's not just the style. It's like they have different dependencies. They have different deployment processes. They're using different infrastructure. I can't move those people around effectively. You know, like team A is team A and if team A walks out the door like wow, I'm kind of screwed or if anyone from team A leaves like how do I replace them? So like this problem of being interchangeable has like real effects for the business as a whole and like it's by through standardization you people can actually work better together. There's more code sharing. There's the benefit of them being interchangeable work across teams. If product A that team A maintains goes out of business, we can move to product team B, right? But like AI again is an amplifier of like all these problems. Yeah. Yep. Yeah, you'll just you'll keep promoting bad habits. And like yeah. Those keep happening. I think one thing I've seen a lot recently is a lot of people discuss kind of the concept of like spectrum development. You know, one post I think I said I saw or said something like you know, I take like a day or two now to develop specs before I give it to the AI to build it. Do you have any thoughts there in terms of like as people are trying to take this long time just to plan out kind of what they're building? Well, I do. I mean it makes sense in certain cases. I think it's like a trade-off though like Um You write a lot I mean you will spend a lot of time writing specs. So I think it's like yes, spec driven development works but like how do I spend less time writing the specs? And so automating that process becomes really critical. Yeah. You've seen like how we do prompt expansion, right? Like where you input the initial prompt but then you're walking the user through like a checklist of questions or whatever that's like Okay, I see you want to do this thing but here are all the things you're missing in order for me to be able to do this thing intelligently and then the final spec we're having the AI write that based on like training data and things we've taught it. So I think it's like yes, spec driven development works but spec driven development takes a lot of human time if you're relying on the humans to write all the specs and then again, there's this problem of like well, various people will write the spec in various ways. How do we make sure that when we write a spec for a given part of our stack, it's consistent in the things that it covers cuz otherwise again, you wind up back in this amplification problem if team member A writes their specs one way and team member B writes their specs another way, the outputs could be vastly different. It could one could be better than the other. So what we've invested in a lot is like how do we automate that spec creation process? How do we learn from it? How do we track the artifacts produced those artifacts evolve over time? That's telling us like where we're good specs, where we're bad specs, where we're like the processes that generated that spec, where do they need to be refined? But I don't want my team members sitting there all day writing specs. That's also kind of not a good use of their time. Yeah. Yeah, and for anyone who's wondering we're kind of we're both working together at this point at Coded Strap to kind of build some tooling around that. So when Dorian's say is saying it like we both know it. That's what we're both kind of working on. And one of the things that I think the focus is is around like developing good specs and helping for example junior engineers have a good spec, maybe think about things they didn't think about so they can go back to you know, the business and be like hey, actually I'm getting questions from the AI that I'm like I don't know how to answer this. Maybe I should have thought about this before I go and build it. Um But on that note of like kind of the junior side, like you know, I think a lot of people kind of have this idea of like there being a junior problem. Could you like summarize it in a sentence? It's the ROI problem. So like the thing with junior engineers that businesses have always had a problem with is the when do they become ROI positive? So it's that ramp from when you hire them to when they become something that is generating, you know, return on the initial investment and part of the tax you pay is like how much time do they take from the seniors who have to train them? And so like it's both getting them ROI positive and taking less time from your other resources as well. The junior problem is really like the compression of that timeline, right? So like it's not am I going to get rid of juniors? You're not juniors. If you'd like to get rid of juniors, what you're effectively saying is like I'd like to collapse my talent development pipeline and you're at that point you're all in on AI being able to replace humans because if you're not replenishing your talent pipeline at a certain point all your seniors are gone and what are you left with? You know, you're left with the AI, right? So like no one's doing that by the way. I mean like IBM's leaning into creating more junior roles, not less, right? And

there's a lot of data to support this. Software engineering roles are generally growing at 11 to 15% depending on who you ask which is outpacing the general labor market anyway. So to me the junior problem is like how do I make them ROI positive faster? How do I encode into the platforms they're using engineering expertise so they can just ask the AI why do we do things this way? How do I solve this problem? I'm getting this error on deployment. My local environment won't function correctly. Like all the things that you would normally be on Slack and be like guys, can you help me? Let's put that into the platform. Let's solve it once. Let's make it retrievable so anyone can access it and let's train the AI on how to get people unstuck. At that point, that ROI that maybe used to be one to three years out depending on where you were, compresses to weeks. Yeah. And at that point, I'm going to go on a junior hiring spree. I'm going to hire as many freaking juniors as I can and I'm going to smash my competition because I'm going to have sort of labor economics that they wish they had. Yeah. Yeah, I think that's like that's generally what makes sense I think in the long term is like if they if eventually it'll make sense like I guess it's from an economic standpoint. I also like the fact that you referenced like setting up your environment, right? Like how many of us have had to do spend so much time doing that and then like had to eventually ask someone like how am I supposed to do this because something's not working and then now you, you know, you like you said you've taken a senior's time or someone who's been there who's actually doing something else. Um overall. It works on my machine. I thought Docker was going to fix all these issues. Oh god. Oh my god. Tell me you haven't had to do development in Docker. I was at a company one time and they were forcing us to do development inside Docker containers to solve this problem. Yeah. Holy crap was that bad. Yeah. Terrible development experience. The new it works on my machine problem now seems to be around like usage of the AI god. You know, it's like how do you pray to the AI god? Well, I pray this way and I get good output but like you're over here you're praying this way. I don't think you're getting the quality output. I feel like I had that experience the other day where I was like why is this output not coming out right? And then the person I was talking to was like oh yeah, you just you got to do it this way. And I was like this is dumb. I hate it because like at least before like there could be a rationale but now it's just like whimsy I guess or whatever and it just happens to work sometimes and not other times. The whole like it times is always I think just a crazy idea. Like I just love that we're all cargo cults now like we're just like okay, I think it's here we go. Will this prompt work? And then like they ship a new version of the model and everyone's stack just disintegrates. It's just like it's awesome. — Oh, it's that's so terrible. Like it it's funny cuz I was thinking they're just doing the same thing at least what it feels like to me was when like new iPhones would be released, right? Like they just like in shitify the old one by like releasing new versions of the OS so that your old one like drains battery and doesn't work as well anymore. And that's what it feels like I was like using Claude for like a while and then suddenly 4. 7 came out and now I can't use like 4. 6 is unusable in comparison. I'm like what how? How can you go from like being usable to unusable? That makes no sense. Anthropic will find a way. Don't worry. It's like it's Yeah, I really taught us early on like too much dependency on the lab or the model is probably a bad idea. Like if you're doing something and you're so locked in to this one model or this one, you know, lab's model or version of it, you really do have to question yourself like am I doing my part as an engineer to put the systems around the model such that I don't have this level of lock-in? And I think that's where if I'm a software manager today, I'm putting a lot of my time, I'm going to start looking at things like this because if it's if I just see my team burning tokens, right? And they can only use Anthropic's models and the output quality is not going up and to the right. The release cycles aren't right or the things that matter as an engineering manager, I got to start questioning like are we just doing this? You know, like we're doing it because everyone else is doing it? Like what is the thing here? You know? I mean I think a big part of this is we're human and we're lazy and you give us a tool and a system that lets us hit a button and create the thing we think we're supposed to get, like yeah, we're going to hit it cuz maybe the third time it'll work or the 10th time, great, right? Like I didn't have to process anything. And if we don't have to process anything, we will do it. And it's just it's who we are. Like and it's crazy how fast it happens, right? Like how fast you're like no, I don't want to think about this anymore. Code, do it for me. You know. The the developer behavior has changed in like a year. Yeah. It was that fast. I mean that is just mental when you think about it like how quickly the industry has flipped. I you see it in the lab's growth as well. I just I think long term I think people are going to find out that like the platform engineering is more important than the model, you know? And that also there's just a lot of

advantages competitive advantages on the business side if you do reduce these models to a commodity effectively and make them interchangeable and not only is that going to show up in your ability to compete, but you can compete on margin if you effectively like if the new metric the new thing that software companies are now bottleneck by isn't labor, it's tokens. Um I don't think that's true, but if it is true, uh one of the things that's going to be a real challenge is who can use those tokens more efficiently cuz there's just not an infinite supply, you know? — Yeah. And so like the engineering around the models making them commodities making them interchangeable will be this thing that businesses compete on if it's like oh tokens are now the new labor pool kind of thing. Yeah. Yeah. No, that that sounds about right. And well I definitely have some questions on platform, but a few more I think on the junior side of things. First of all, do you feel like expectations are shifting in terms of what juniors need to be doing, you know? I don't know. I mean at this stage I don't think so. I think the expectation is the same. I think where I see the biggest shift is like um maybe it's the the soft skills that juniors need are becoming more important especially as we look to the platform to be the enabling function, you know? We kind of used to look more like um I guess it would be like are you do you do we think you have a predisposition to system design? Um how well do you get along with the other engineers? That's so important. But like the thing that's most important right now like that I'm seeing is like can you get unstuck on a problem without having to go and involve someone because you've mastered the tools you've been given, you know? That is becoming really important because if like let's just say hypothetically we're all investing in platforms to help make juniors more ROI positive faster. If you're a person that then still wants to go out and talk to people that might be a problem. Like to not for social reasons. I mean talk to people like because I have a problem and I need someone to solve it for me. Yeah. That is going to really be something that works against you in the hiring process and like I think it's like also this thing like are you 996, you know? Like junior engineers, if you're not willing to work 60 hours a week, you're not getting the job, you know? Like it's just the new reality in tech. Tech has not happened fat you know, fat and happy anymore. They're cutting their workforces down. They're saying it's because of AI. We all know it's because they over hired. They probably created a culture of complacency to a large degree. This has been cited by people like Eric Schmidt and others. Um So I think that like as a junior what is changing is like the expectation from employers I think is that you will work extremely hard. Um you will get unstuck on your own. You will be a self-learner, self-starter. And you won't bother senior engineers just to go and say like hey, I'm stuck on this problem. It's like okay, you're stuck on the problem, but did you go through all of did you use all the tools? Do you know how to Could you have solved this on your own had you leveraged everything at your disposal. I think that's becoming something we're looking for more and more. We're trying to screen for more and more as we look to hire the juniors. And the old things are still important like being like you're predisposed to system design. You think in terms of systems. I think that's important, but also are you predisposed to learn lots of things? Cuz like the future shape of the senior engineer is more T-shaped. So like you have this like really broad breadth of things you're good at that you know how to do. You may not be an expert in any one of them, but then you are an expert in at least one category. Right? That is becoming the shape of the senior. Um So like this as a junior if you are willing to commit to learn lots of new things it really shows like and you're willing to master tools it really shows that you have a path to a senior kind of an engineering role. If you only want to do one thing, you know, like I just want to do this one thing. There is a high probability that AI will close the skills gap. Like we've seen this in studies. Uh like with I think there was one with Harvard that came out where they showed that like through using AI systems and tools people with less skill can compete with people with more skill. So if you're only good in this one area and that's all you want to do, there is a high probability that AI is going to close that skills gap pretty quick. Yeah. — So like I think like breadth of knowledge is super important. Being deep on one area is really important because that helps you build AI systems that can be effectively gated and trained in that one area, but you do need to have a really broad breadth of knowledge these days. Yeah. Yeah, I mean are there any like areas you think are worth investing in like depth-wise? That's a good question. Um Yeah, it's uh I would say like the areas that are that have the like the deepest stacks still seem to be the ones that are most widely adopted. So like let's just take React, right? React is used in what is it? 80% of all web and mobile web applications. It's like the dominant player on the internet. Um [snorts] it has massive depth to that platform. I mean like you could go to the point where you could be part of the Facebook core team. Like if you have that level of expertise, you actually

are super valuable cuz you can put that expertise into coding tools. And it has a massive market, right? So like I would say like the things that are most widely adopted are probably the things worth going deep on and spending a lot of time on just from an economics point of view. Just from like the point of view of like I want to invest in things that will have the broadest set of job opportunities out there. Now you make the reverse argument that's like well, I want to go after the thing that AI is least likely to replace. AI ain't replacing I still I advise to be like don't get into the habit of thinking like I must go after the thing that AI can't replace. It it's not replacing anyone. The job data doesn't support that. Really what you want to think about is what is the skill I should go deepest on such that I can build AI systems that have the broadest market appeal, you know? That's what I would do. Yeah. Yeah, I think it makes that makes sense. I think it's a similar advice I like maybe even before AI it's just like what are people actually using? That's probably the skills you're going to want to learn. And then eventually will they get replaced? Yes, and that's the world you live in as an engineer is every 3 5 10 years things change significantly and then suddenly you got to redo things. Um sometimes you can stick with some certain skills like I mean SQL is still SQL to degree, but you know, things change. That's part of I think part of being an engineer, you know? You got to keep up with all of it. Yeah. Speaking of things shifting and changing, what do you think in terms of like AI and where does it shift the bottleneck? Or sorry, more specifically I think in one of your articles you reference it shifting. Like why do you think that AI should kind of shift the bottleneck from engineers to the platform? Uh because again back to our like earlier discussion like the AI is an amplifier. And so the thing that people don't realize um is that like in amplifying that dysfunction, dysfunction occurs with scale, right? And so as you pass between number of engineers from a from 10 to 100 to 1,000 to 10,000 to 100,000 whatever it is, these are new scale problems you have to tackle. At a certain point, you have to have really good platforms that engineers build and ship product on. And so at Google scale, Google has a mono repo. There's billions of lines of code. There's something like 45,000 engineers who use it. They had to come up with a lot of platforms and systems engineering to solve those scale problems and it was very tough for them as they grew, you know? But they learned a lot and now they can do it. And by doing that, they reduce the surface area they have to manage. They can effectively build and ship product faster. Everyone now has that scale problem. If you have 2,000 engineers that now output, you know, 100X their capability you got a real problem, you know, in terms of platform engineering. And so it won't be the outputs that are the bottleneck. It'll be the outcomes. And the outcomes are going to be driven by who's got the best platforms. And that's just fact. I mean like you can't even argue it because any company that's ever had to deal with scale like Facebook has had to invest in platform engineering making good platforms that product teams can use to build and ship product. So the bottlenecks are clearly going to move to platforms because that's how you're going to effectively be able to manage the number of lines of code growing, the complexity growth as a result of that and trying to orchestrate all this stuff so you can effectively deliver. Like a lot of people don't even know that like at a certain point the pull requests become the bottleneck. Like just getting the PRs merged becomes the bottleneck in shipping product, you know? And so a lot of people haven't had to deal with those problems, but now everyone has those problems. A 10-person team in theory, you know, if you believe what Gary Tan is writing like 35,000 lines of code a day. That's a lot of code. How you going to effectively manage it, you know? Like platforms are the only way we know how to effectively do that. How's it all going to get merged correctly? How are we, you know, going to deal with every conflict? Actually push it. Make sure it doesn't take 3 days to actually get pushed once you make all the changes etc. Yeah. Did you see GitHub's merge queue had a bug in it today where it was like reverting changes? Oh. What? It was pretty interesting. Maybe that was the Yeah. Why I saw some tweet about us creating a single dependency on all repos or something. Uh cuz they're like why are we all on GitHub again? It's like we're all on GitHub. That's why. What in the world It's just how it is. It's just Yeah, why change it? Um not that I'm promoting any other solution. It's just like it's just how it is, you know? They're all they're all some level of good and some level of bad. What do you think or why do you like why do you think most companies get platform engineering wrong? Just not a lot of people have done it, I think, you know? Like in the like let's just say you're 200-person company, you probably don't have a need for it at that stage, you know? You like you're maybe you got 30 or 40 repos. Some of them are legacy, you know? But everyone's humming along and you haven't really created a situation in which you have to have a platform dedicated platform team. You do need standards and practices like those teams to effectively not hit a wall of complexity growth and tank the company's products, they're having to deal with software entropy. They're having to reduce it. They're having to implement abstractions. They're having to find ways to work together. But the idea that you would need platform engineering is not really something most people actually truly need. Uh so I

think that like there's just not a lot of people who've had to do it because it's something that only a handful of companies have really had to deal with, you know, effectively. So you're talking about Fang companies or really large, you know, software development teams. And so there's just not a lot of people that know how to do it. And then there's probably fewer that know how to do it well. Yeah, the ones that come to mind for me are like Hyrum Wrights and Titus Winters from Google. They're incredible. I learned a I've learned a lot just from watching their YouTube videos, reading their books. They're an example, but there's like, okay, there's two people. — There's not probably not 2,000, you know? And so like that's why. Yeah. Yeah, that makes sense. And then like you said, it's like hard to get to that scale. Um maybe not anymore. Uh but everyone's putting out as much code and everyone's going to be expected to make everything talk to each other um in ways we haven't seen possibly in the past. So, you know, there's going to be a lot more demand for it. Um so you you also kind of referenced in that article like that AI success is kind of a systems problem. What do you feel like you actually mean there? So yeah, um let's getting back to the engineering around the model. So like everyone knows Claude Code super successful, right? Like hugely successful product. And it's not because Opus is the most amazing coding model, dude. It's actually the harness that's around Opus. In fact, the creator of Claude Code didn't even want to release it because he felt that it was Anthropic's secret sauce, its secret weapon to go out and compete. And they wanted to use it more in the enterprise space to go out and capture that market. And so it's like that's an example of where the engineering around the model was actually more valuable than the model itself. And we see that because we're building our own harness, Ben, right? And we can swap we're using Flash Gemini 3 light to write 80% of our code, Yeah. So I can say with confidence it's not the model, you know? Like we've proven I think beyond a reasonable doubt that it's not the model. It's the engineering that goes around the model that really makes for um stunning AI systems that are predictable, repeatable, cheap, right? These are things that I think organizations are becoming more and more aware of that like yeah, your model looks great on benchmarks, but when I go to deploy it in an actual use case, I it's a highly unreliable. Like it doesn't produce the same outputs every time. It doesn't, you know, behave in a predictable way. And it costs a fortune, you know? And the insurance industry is the biggest leading indicator of this. Like we've seen three articles come out this week for major insurers pulling general liability coverage for AI systems, Yeah. It's like that's why the AI system is the big deal. Yeah, it's a weird space to consider like insurance, uh you know, people not wanting to be involved. They're like, I we think there's high risk here. because no one's going to be able to take like who's going to take accountability when things go wrong. And who gets sued when a system blows up if it was all AI. It's kind of the same I guess situation when you think about cars and car insurance. Like who's — But it's where the value is accruing. That's kind of what I'm looking at it as. It's like it's actually the models, you know, if you look at the performance characteristics and stuff, they are actually all converging towards the same ceiling. And I think like the solve the problems the insurance companies are pointing to is the AI system. It's not the model. You know, that's where the value is going to accrue. And the insurance companies are just on the they're just the leading indicator, you know? Like they're the ones that have to deal with risk early on. You know, and that's why it's interesting to look at them is like they've identified a clear risk here when you're just looking at the generative AI models. And so like the value is now pointing towards the people who can create the systems that effectively gate those things and resolve this problem that the insurance company's kind of the leading indicator of, which is the lack of reliability, lack of accountability, like all the things that organizations kind of need to show a real positive outcome. It's one thing to post tokens on a token board, it's another to show like revenue moving in the right direction and your balance sheet improving. Like these are the things that companies are going to actually look for when we start to looking at outcomes and not like outputs. Yeah. Yeah, I think that that's always the an area I think about too where when it comes down to like looking at the outcomes we're currently getting, I'm like, are we in a different state? Have we gotten more features from the companies that are using AI? Is it like shocking uh how fast some of these companies are moving? I feel like there was like someone saying like, oh, I now get I don't know what the thing was. Like I now do what used to take like a year and two months. So then you're like, well, by that logic, that means by the end of the year, I should see five years of progress. I don't think I have. Maybe I'm crazy, you know, maybe I'm not seeing behind some curtain, you know, maybe but I don't feel like I've gotten five times more things from the company — Where is it all going, you know? Where is it? This Facebook thing? Like 30 days, 60 trillion tokens? Where is that showing up in their product? Like I it's just it's confusing, you know? It just vanishes in the ether. It all goes into Sam Altman's pocket. I don't know. Like I mean, where are the tokens going? That part is I feel like very Silicon

Valley-esque, right? Like it's like for a while there, um it was all about uh the economics around like uh social and uh other forms of like paid ads online, right? It was like, oh, do your tokenomics work? You know, if you got some product that you're selling, can you get more money than you spend on ads and your other cost of overhead? Great, you have a business. Uh now it feels like and then all the like well, I think they probably had so much statistic where it was like 80% of all dollars of ad dollars or something just went to like Google, Facebook, and a couple other places. It feels the same now. It's like, oh, all money just now goes to this I mean, you said the magic word, ads. It's ads all the way down. That's where like most of the tokens are going, I think, at the end of the day. Like Facebook especially. They you know, what's interesting about Facebook is they actually they released an updated reasoning model. It's the only one I've seen so far that actually worked on tokenomics. Like they basically still are doing a reasoning model, but they're doing it for way cheaper. So it's like they designed a special way of like generating the outputs so that it kind of like stops generating for a little while, then starts generating again. It's like able to reduce the cost. Uh but they actually have a business successful business that's implementing GenAI with advertising, you know? So it's interesting that if you look at the companies that are actually really focused on doing well with GenAI as part of the product line, cost is the major thing they're looking at now. How do we bring the cost down such that we can make do this crazy thing called making money, you know? So it's kind of interesting. Especially as like the AI buffet, all you can eat buffet is kind of feeling like — It's gone. — Everyone wants to pull the rug out and you're like, all right, time to make money. Um It was like a game of chicken. Like they're all just like, okay, are they going to pull their sub model before I pull my sub model? Like what do I do here, you know? Yeah. And I think like Anthropic might have been the last to blink. I don't know. Uh what's weird is like Anthropic doesn't lose money at the rate that OpenAI doesn't. I can't figure out why. It's like OpenAI is like incinerating cash and then Anthropic's — Well, I think I just saw some quote and I don't tell don't uh blame me if I'm wrong. I think somewhere it's like they have like 5% of the users, but make more money or it was something crazy like that. I mean, they focused on B2B. Right. — is where the money is, right? Actually, that was a whole scene from uh Silicon Valley where um it's like I want to say Kevin Newsom, but that's the governor. Uh but I think it's still Gavin uh is like uh all these engineers, all I can think about is like consumer. Uh they missed like the B2B essentially, the enterprise sales. And it's like, yeah, it's sell the enterprise. You know, you get your million, 10 million dollar deals. You don't have to sell — Yeah, Anthropic crushed enterprise. They went like in there and just completely took over. Yeah. And so yeah, maybe that's part of the strategy that's working for them. Um so maybe they're not investing in same level of pre-training, who knows. Yeah. But it's it is kind of interesting like looking at it and like Anthropic looks like the more sustainable business right now. I'm just wondering if that scales, you know? Cuz like OpenAI has the 800 million users on the platform. They seem to be operating on a much larger scale than Anthropic, but like does Anthropic's business model shift and start hemorrhaging money in the same way OpenAI does at a certain mass adoption scale? Or do they find a way to like actually sell this stuff in a way that makes sense? And then they're all their business models are constrained in this token reseller market, too. So it's like the people who could even if they're successful and they get mass adoption and everyone loves it, well, here comes Google and Amazon and Microsoft with margin compression and the next thing you know, they're screwed. Yeah. Uh I'm kind of going on from that and thinking 2026 onward, what do you think it kind of means to be and you kind of referenced this I think a little bit when we're talking about junior engineers, but what do you think it means to be like a top engineer in like 2026 and on? I mean, again, like I think it's the T-shaped developer, broad depth or lots of breadth, you know, broad breadth, depth in one area. Also, you have to have a really good AI intuition. Like you need to know like um what is AI good at doing? Where do I apply it and where do I not apply it? And you have to learn that I think from experience. Like it's not something you're just going to like wake up one day and know how to do. You got to like kind of trial and error this thing and get that scar tissue from applying and reapplying again. So like the AI intuition is probably the most valuable thing because there aren't many people with it right now. And I I honestly think using a tool isn't the same thing, you know? It's like if you can use AI successfully in an outcome and get a production use case deployed and it's driving, you know, it's moving the needle for the business, that experience you gained is irreplaceable, you know? And so that AI intuition and successful use of AI to drive positive outcomes is the thing I think that is going to define it along with like how do you orchestrate whole teams together with this stuff? I mean like we can all show a great demo of us working alone, but like no one works alone, you know? And so if you want to be a senior engineer, it's like how do you empower the team to work together with these tools is also going to be like a really important skill because that's scalable. I can

plug you into an org and take 30 people, make them work together and they act like 3,000, you know? So I think that's also going to be super important. Yeah. Um and then kind of on the lines of like skills, what which ones do you think matter more and better less moving forward? Um yeah, I mean definitely not your use of Vim and Vi, you know? I don't think anyone's going to miss Well, some people are going to miss that a lot, uh but like I think, you know, people are priding themselves a lot on their setups. — uh the Primeagen very upset. Yeah, I think so. I think I pissed him off. Uh don't come Don't Do you ever see that pizza wrestler that looks like just like him? This dude who has like has pizza in the wrestling Oh my god, it's hilarious. I swear to god this is like his double. Uh but yeah, like I don't No one's going to care about the typing, you know? Like how fast can you spit code out? I don't think is the bottleneck anymore. The models Like you're John Henry, you know? Like you know you brought this up in the pod the other day. I was like no, dude. Like you're trying to beat the machine that drives railroad spikes. You aren't going to compete on the like oh my god, look at my setup and it's amazing and I can spit code and look at my keyboard. It's incredible. Like no one cares anymore. I think I think at the same time you still have to edit code. be like involved in the generation process, but like it is no longer relevant how fast you type or how quickly you can get the code out. The system design aspects are super important. Your ability to orchestrate the stuff together super important. Experience building platforms that brings all this together super important, but I don't think the ability to like type quickly or have good ergonomics on your keyboard setup is as important as it used to be. Yeah. Yeah, I think that makes sense. Let's see. We might have some questions here before we go on. Um some people are asking about what skills to save. I think we've kind of answered some of those. I don't know. Let's see. Yeah, I don't know. It's kind of strategic skills. Everyone same same feeling about skills. Uh let's try about this one. What do you think about this progress? What will be What will the finished programming system look like in 5 years? Will it be uh challenge Will be a challenge for model groups? I don't know if I can get a good question out of there. Unless you do. If you don't, then I might just be picking the wrong one. Um like the finished system if you mean like what is the finished like future coding platform look like? I have no idea. I don't think anyone can answer that. No way. 5 days from now I couldn't answer that question. I think like uh 5 years from now we may have a fundamentally different technology that isn't even based on LLMs given the amount of research going in the space. I would just focus on like what can I do today to leverage the tools that are there to produce a good outcome recognizing that it's going to evolve quickly and so not like hold strong opinions that are loosely held and adapt to new information as it comes in because like we're all changing. I think Ben, you can attest to this too. We are changing daily based on how this thing changes, you know? Yeah. Yeah, everything just changes so quickly. Um again, you suddenly have to rip out a model because something is deprecated or you know, something that was working is not um because you and 12 months ago no one was using Anthropic. 12 months ago. I think that's that is kind of the wild thing because now suddenly I'm already forgetting the French company because that's what it was. It was um it was opening eye and then now I'm — Mistral. Yeah, Mistral. Like it was like that was the thing and then suddenly it was like does anyone reference Mistral anymore? Like I just don't like it was like now it's like opening eye and Anthropic and you're like when did we switch? When did we make the switch to this choice? I'm not saying that people probably still using Mistral, but like in the zeitgeist it's just like it happens so fast where you're like suddenly everyone's talking about Claude as if it were always there, the option um compared to open eyes. Like what what? When did this happen? What did I What person went back in time and did I miss and like made the switch and like it's like things never happened. Once like it's like that feeling you get when you've been driving for a long time and you don't realize how you even got there, you know? — Yeah, actually that's a great Yeah, that's how I feel right now. So yeah. I was just like I don't remember this happening, but I do remember it not being this way. Um I remember where I started where I got to, but I somehow the middle is a little fuzzy. Um I'm trying to think if this isn't so much Is this a question? Uh juniors used to learn by sitting next to senior engineers. Uh that apprenticeship is being broke by LLMs. What replaces that apprenticeship? I would say it's not breaking the apprenticeship, it's scaling the apprenticeship because it's like um that leads to a long ROI curve. So like yeah, that is how we've been doing it, but like the ROI curve is really long. You're looking at like a year to 3 years depending on, you know, the people involved and how well they manage that relationship. Um so right

now it's not breaking that. It's how do we scale it and get it compress faster to like weeks not years. And so how do we take the senior knowledge and code it into the platform so that then it could be shared and used by juniors in a sort of more democratized way instead of having like one junior and one senior one five juniors and one senior. I can do 50 juniors, you know, to a senior. That that's how we have And those juniors become ROI positive because the platform is guiding them through common mistakes much faster. So like that I think that's what changes in the apprenticeship model is way more important now because to me like the companies that are effectively do that are the ones that had the best apprenticeship programs and now they can work on encoding them and scaling them and now you got this massive labor arbitrage opportunity to take advantage of the junior pool and go out there and like kind of smash your competition with it. Yeah. It's a good question here. Uh maybe just for everyone to ask themselves after months or years of using those tools, what if you be suddenly lost access to them for some reason? This is good. This is like a skit or an episode or like a meme or something, dude. We talked about this though too like what if the future is like the AI labs are trying to push on us is like you put out so much code that you can't even work in the code base anymore without AI because there's just so much code in it. You can never go back and effectively digest it. People should be really aware of this. Like this is actually a really good question just to bring up to your engineering manager or to your board or to like whoever is making these decisions in your company about we need to token max is like just hypothetically you go on the journey with me. Like what happens one day if we wake up and we can't use this tool anymore? — Yeah. Or what are we going to do? Or they 10x the price overnight and you've again built a 10 million line code thing. How? Like you're going to hire 500 engineers suddenly and tell them to fix it? It's not that's not going to happen. And then it breaks and then no one knows where the problem is because there's 10 million lines and like one poor engineer sitting there trying to draw out where and how the flow is and what's calling what and what's abstracted away. And there's eight different ways that people coded and abstracted things away. So you you're like right. You've got you think you understand how the code base works and you get to the next section and it's a totally different code base and you're just like oh. What am I doing here? Uh and that's assuming you can read through it all. So no, I think that's the concern. I I think there's like an article out there about um not I mean they're doctors, but I'm trying to what doctors specifically that looks at x-rays. Um that x-ray imaging team radiology. Yeah, yeah. And like apparently after like using these tools for even a short period of time, like their skills degrade. It just like happens. Like they Skill atrophy is a real thing. Yeah. And what's weird with reading and writing is like they're not they're correlated. Like your ability to write well is tied to your ability to read well and vice versa, you know? So like if you just stop doing one and you only do the other, you are going to get a skill atrophy problem. And then there's also the context switching problem because like one of the things you're constantly doing with AI is you're not going deep on a feature for months at a time. You know, you might like produce the whole thing in a day and then context switch over to some other thing. And that is actually proven to be the thing that for long-term memory is the biggest killer is context switching, you know, or short-term memory. Sorry, for short-term memory the context switching is the biggest killer, right? So that you stacked all these things in like perfect storm, you know? Yeah. It's unreal. It it's it is strange because in a weird way like you are more productive and other way like you like come back and you're like what was I doing here? Like I don't — There's no review tax. Like you know, they call that the there's like the generation thing that goes faster, but then there's the review tax takes longer to review. But then there's this other tax that question is pointing to which is like the LLM tax of just having it in your stack, you know? Skill atrophy, vendor lock, like I can't modify the code base now if I ever lose access to it. These are all real big problems that a lot of companies aren't even thinking about right now at all. I will keep I'm going to continue saying you will know nothing and be happy uh is their goal because then they can make as much money as they want. Like they can charge you what they want. — Exactly. You will own nothing and be happy and you will know nothing and be Exactly. It's a perfect world. Uh you know, you just sit there and be happy with your AI generated like scrolling through Instagram with AI generated content and it's all good. You'll be happy. We're going to live in that world of AI slop and you're all going to love it and — Yeah. AI slop, bull slop from Kava and Yeah, yeah. all of just Yeah, awesome. We're getting close to the end of the end. I think it's all I think I haven't missed any questions. Uh, I don't know if you have any final thoughts. Don't want to keep you too long, but uh yeah, this was fun. Yeah, I think um as a junior you should be bullish. Like if you are a junior, like I think the future is going to actually be really, really bright for junior

engineers. And not to buy into the AI doomerism, like the actual data does not say there's any job loss. If anything, it says jobs are going up and to the right for software engineers. So like I wouldn't buy into that. Then there's the Jevons paradox thing, like you know, this thing becomes more accessible that used to be precious. Like you're just going to use more of it. I totally think that's where we're headed. I do think that software itself may become more disposable in the long run. Uh, so like it'd be if it's much cheaper to get a unit of software deployed into production, we may produce a lot more of it and we may dispose of it a lot faster. Uh, which is actually kind of good for the economic picture because like maintaining software is actually a lot more expensive than the freaking uh, you know, producing the V0, you know, so Uh, it may change the economics a little bit. Uh, but I would be bullish as a junior. I would definitely focus on how good of an apprentice can I be, how fast can I learn things, how hard am I willing to work. Build those soft skills so that you know, you're a self-starter, you can solve problems. And just like don't be negative, you know, cuz like one of the worst things you do is to sit there and be like, "Well, what's the point? I'm not going to have a job. " Well, you're definitely job if that's your attitude. Yeah. — [snorts] — Yeah, it makes me think of uh my parents used to say they're both from Romania and they would both say like there's some sort of saying along the lines of like, "Why why work? Because God's coming back soon. " — That men- that mentality exists forever, right? — What? So, yeah. Why put effort in? It's over anyways. It's like, "Well, you got to be a little more Don't give up. Don't crawl in the corner, you know, continue to do good. Yeah, continue to do your best as you learn like Yeah. We'll figure it out. Awesome. Well, thank you so much. Thank you everyone who showed up. Uh, yeah, this was great. Thank you everyone. Thank you. Enjoy your weekends. Bye.