Engineering at Ludicrous Speed: How AI Is Reshaping Infra and Engineering

Well, I can speak a bit about my last startup. So, I had a failed startup for nine months. Uh, so it had been February of 2025. Uh, we raised a seed round and one of the questions we got from our VCs at the time was this is we were like, "How much money are you raising?" We're like, "$8 million. It was going to be a multicloud startup for cloud portability." And they're like, "Oh, this is great. Um, you won't have to hire hardly any engineers because AI will write all the code now." And I thought that was the most ridiculous question. I I I flipped the dummy bit on that VC that asked that question of us or made that statement. I'm like that is so ridiculous. There's of course you need a lot of engineers to write this code. And what I found is our CTO was really cynical about that as well about AI because they had tried AI like maybe six months before that. And now a year later he was right. like now he wasn't right 12 months ago but I couldn't have foreseen at that time um and that whole startup is essentially um you know I left the startup because I don't think it's that but the I think a lot of what we were doing um is automatable now because you can go to an AWS cloud um point your AI model at it and say I want the equivalent infrastructure on GCP or Azure and the models can do that now and it wasn't like that before so I think that's like a an example how fast this stuff is changing. >> So that venture capitalist had the crystal ball. We need to go talk to them. So I I see a lot of questions in open source land about AIS like where where is it learning to code? Uh what what license is the code licensed under when it spits it out? Um I'm getting a bunch of slot pull requests. What do I do with them? Um should I should my project allow people to to use AI or not? Um, how do new contributors get started in in an age where if you ask them to solve an easy bug and they can just generate code in three seconds and submit whatever comes out and they don't have the skills to review it themselves. Um, so I think there's a lot of questions in open source world about how does this fit in. >> Cool. Um, so I think I think the followup for me is just there's a lot of engineers in this room. I'm assuming a lot of folks kind of between CIS admin to DevOps to engineering I guess how should we be thinking about it today like what how should we be approaching it what do you like give me the hard opinions you know whatever they are should we be completely using it should we be not coding anymore I know 90% of the room raised their hand when we asked that question should like what should we be doing >> I I don't have the answer to that but uh I I think a lot of people here are here because they like coding. So, we have to keep that in mind as we figure out what the new jobs of the world are. But I think one thing is interesting is just the terminology that you kind of hit on as you were asking that question. Like they were engineers and then they were developers and now they're coders and like the the terminology even changes as our world changes, which is interesting, >> man. Uh I'm retired so I have less skin in the game. I had to just be honest with myself. My views on this are very different than my daughter's views who's 18 trying to enter the industry for the first time. So, I just have to make sure that I'm being reasonable with myself. I can afford not to care about this. I can afford not to fall into the hype. Throughout this conference, the last couple of days, I've seen a few people walk up to me and their eyes are just lit up. Agents, agents, agent, what are you doing with agents? Our agents will talk to your agents. Your agent will call my agent. And then I'm like, "But why?" Because we have stuff to do and my agent knows what needs to be done and your agent knows what needs to be done and they're going to work it out. And I'm like, "So then what are you here for?" And they're just like stunned. Like what do you mean? I'm here to tell my agent what to do. And and I'm trying to listen. I'm trying to like be patient. And the weird thing about this, when I see another human being, the thing that makes me excited about that human is I am going to climb this mountain. I am going to learn to cook. I'm going to go somewhere with my child. I am going to grow. I'm learning how to play guitar. And one day I'll be on the stage. I'm going to write a song about love because I had this experience and I want to share that experience with the world. And never once have I asked them, "Well, how fast will you write that song? Productivity wasn't part of the equation. There's nine billion of us individuals having experience. We make happy accidents. And when you think about life in general, and maybe some people argue that life may not be necessary in those regards, in those terms, but we created the data set. I always try to remind people, you created the data set. You created all those YAML files. You created all that source code. You created all those bugs and the fixes. You created the questions on Stack Overflow. and you've answered them. And we've taken all of that and we shoved it into a machine that is trying to approximate either sometimes it repeats it back to us. I know because I got the class action lawsuit from Enthropic. Kelsey, we've taken your book. We've used it in our model and we're going to give the lawyers 80% of it and you can split the other 20% with the rest of the authors that we've taken the information from. So these are the actual facts. Like these are like real things. So how should you think about it? Well, if they took your book, you might have a little gripe knowing that your intellectual property was used in a machine like this and it's being rented back to other people for roughly 20 bucks a month. Right? That's one way to look at it. The other part is like you're still important, right? Maybe some of the work that we're having agents do was useless work to begin with. My daughter gave her first talk yesterday >> and they asked her after you've went through this whole learning Kubernetes and Docker and all of this thing. How should you how should it work? And I watched her because I know my daughter, she gave the answer she thought she was supposed to give, which is I'm willing to learn it if that's what's necessary. Good answer for interview, good answer for LinkedIn. And it feels like that's the pressure the industry has put on all of us. this is how you better think about it because if you don't think about it this way then you're going to get left behind. We're not sure by who but left behind you will. And after the talk I sat down with her and I said listen you sent the right answer probably for there but that probably isn't the right answer. I think the right answer is no one would build this thing again this way. This is this is a this is a tragedy. Like we have 50,000 configuration formats. We have a thousand programming languages that are all incompatible. We have a thousand ways of doing things. And I remember in her talk that was so dope and this is what I'm going to wrap. She started learning all this stuff because she learned how to write code and she wrote this really nice burnout timer, right? For someone who loves studying, has all this pressure as an 18-year-old that's about to graduate college. She built this timer to just tell her like you're studying too long, take a break. And it was a cool exercise, something that she can actually use. She learned Docker. She learned how YAML files. She learned a little AWS. Gave up on AWS. Went to Digger Ocean. And she did all of that and she showed us that her app was deployed. And she said, "I went through all of that to have the exact same thing that I had before. Why are you all doing all these intermediate steps?" And there's no way in the world that we're like, "This is the best we could come up with. This is just what we have." And that's the way I think about Loom. LM are a product of what we built. That's what's in the train set. These things didn't fall out of the sky and just bestow knowledge on us. This thing has been trained by the best of us and the worst of us. And that thing in the middle, I think that's the decision that you got to figure out that we still have to make. That's how I hope people are thinking about this. You're important. The models are great, but don't forget you have one up here, too. And that one needs to be trained because I don't think we can checkpoint society in 2026. Thank you. >> I do want to add another chapter to the story that I told. So I had a failed startup with with people that didn't want to use AI. They weren't using the AI models and our engineering was relatively slow for what we wanted to accomplish. And here I work now with Ron at Flocks and we're about 25 people and the we have several engineers that are really trying to learn the new tools using a lot of cloud. We have people using Gemini. We have people using codecs. And I found that the people that are really curious and good at sharing and evangelizing what they learned in a daily basis, they kind of like this one engineer, he keeps a diary of what he's learned that week from AI. And it's an experience diary. It's not like these are all the great parts of AI. It's like this is the really frustrating part. I had to try to rewrite these tests and it didn't work and it failed and it sucked at it. But he's sharing his experiences with the team throughout week after week after week. And I saw the other engineers in the team then start sharing their experiences. So I think that that learning mindset and having someone that's curious and sharing is helping the rest of our team like navigate this because it it is dizzying, right? I feel like right now if I turn away from like two weeks and come back that like the bleeding edge is like some new gas town or uh superpowers or there's all these like you know catchy names of the new framework that people are using but our engineers are learning from each other and that gives them I think the confidence to to keep going. >> So so I think you mentioned engineers learning from each other like where where else are you guys learning what's latest? Are you actually trying things out? Are you just reading? Like trusting someone that you read from or how how do you keep up? >> I'd say I hear about new things from reading. Um, and then I try them out and I actually learn a lot from AI chatbots. I'm like, you can ask all those stupid questions that you would never ask maybe people in the room and you can say, you know, someone mentioned this and I didn't understand it. Can you explain it? And then you can say, how does it relate to that? And it I think it's a really good learning tool just in and of itself. AIS are llm. >> I'm starting to feel like the last Jedi here. And when I go to the grocery store, I do try to find the organic things like because like the selling point is like this is just this fruit and nothing else. Like that's the selling point. Like it's just food and nothing else. Like yeah, you got to pay extra though. The organic thing costs extra. this thing uh you know it's half price right we made some trade-offs for it when it comes to like learning things luckily for me the sources of information are I do a lot of VC work so due diligence before we cut that check and we're about to spend our LP's money hey why did you give this company $1 million what did the founder say what are they promising to do and then I get to do due diligence so anytime I meet a founder like we're going to use AI for this thing I say just do me the privilege of treat me like a smart person just for just for this one call. Don't say AI. Don't personify it. Don't say work in a box. Don't say digital workforce. Any buzzword that works well on LinkedIn. Please spare me. Just talk about what it does. That's it. That's all I want to say. Don't don't say AI. Don't say rag. Don't mention any of the models. Just show me what you do. The worst founders, they don't know what to say. All right, Kelsey man. We import spreadsheets and we put them in another database. Sorry, just be honest. That's fine. Some say, "Hey, I used to do this work for a very long time. I've watched my colleagues struggle putting information between three and four systems, literally calling gas stations to get an inventory check on the inventory in the gas station. It may or may not be accurate. They get that information and they put it in two or three systems and the team gets in a circle and they have to decide, do we do buy one get one free at the local grocery store or not?" Right? Do we have enough inventory and enough profit margin to do that? And today, Kelsey, the way that works is lots of paper, a lot of inaccuracies and a lot of lost product and a lot of lost revenue. So, we built the system to try to streamline that. And we do that by allowing people to take what they have. Sometimes it's handwritten notes, sometimes it's actual receipts from a big vendor, and we allow them just to put it in one place. And we do leverage some of these newer technologies because they can't say AI to read those notes and turn them into bits and bites that we can consume in a certain format. So we normalize on the front end and then we just try to mimic the workflow that they were using before and we show it to them. It's like that's exactly right and we're getting the results we want. This is amazing because only three or four of us at this company could do it and they didn't want to do it. So that's what our product does. Our product takes a thing that needs to be done that no one wants to do. And we think we have a few new primitives to finally get us over the hump. You could have done a lot of this with bash if you knew what you were doing. But there are a lot of things that were just so hard to rig up because you would have infinite if statements to parse all these docs and libraries just weren't there. And I've been working on this for 15 years. And after these new things came about, I finally got much more predictable results. I've been able to kind of see my vision come to life. That's why I'm excited about adding these two pieces to the 30 we already have. And so when I see a founder do those things, then I start d driving into the technical details. All right, show me what pieces of the stack you're actually using. And I'm like, oh, those are the winners. Those are the ones that are allowing people to build real products and solve real problems. And then when I'm hearing someone's like, hey, I haven't wrote code in eight months. Hell, I don't even look at the code, right? I see those posts on LinkedIn. Hey, I'm a professional engineer. I just let this thing do its thing. and I look at it and be like, you know what? Does it matter? It's probably right. If not, we'll find out later. That to me seems highly irresponsible. That sounds insane. But then I check myself because I know what we tend to do as developers. Import big ass library from written by who knows who, call function that you need, and then we deploy it straight to production. >> And so, are we even any better? So, now I'm just trying to be a little bit more pragmatic. show me what you're doing and I work backwards from there. So, that's the way I'm getting my information, but I'm trying to dig into the details that way. >> I love that. And and I think you you touched on the subject now, right? You said um there are parts of the life cycle that we do certain ways and there are parts that we're doing differently. So, what is actually changing? >> Oh, oh, I love So, look, there have been people that's like, Kelsey, I've been doing this for like two years. I was like, "All right, I'm going to believe you and I'm going to believe that the technology is really, really good now. At least better than what it was before. So, I don't want to have that debate. Show me the results, please." They're like, "Well, we sell insurance." Like, great. I'm going to go to the website. Hey, this looks the same as before. So, what are you doing with all this 10x productivity? And then people like Freze is like, "Uh, we're getting through way more issues." Like, great. Show me the byproduct of doing all of this productivity. Are you getting paid more? Like, oh no, we're they didn't give out raises with all this productivity. Okay. Um, did anything get better that I can actually touch? And I think those examples are far in between. I haven't seen a lot of things just get dramatically better or noticeably better. So, right now, I don't know how much is changing because when I was writing software, figuring out what to do was hard. And even if you did it, you could have made a mistake. Like Microsoft moved the start menu. You guys remember like what was it? Longhorn or one of these? They're like, "Oh, we're gonna idea. We're gonna do this metro theme and who needs the start menu?" And if they did that really really fast and they put it out, people like, "Hey, put it back. Lower left is where it belongs." So even if you want to go 10x faster, you're not moving that start menu. And so now the question I have for people now is like, even if this all changed, where are the results? I need to see the results before I'm all in on like this is real. Do insurance companies go into 10,000 verticals now because they can just write all the software they want or do they just stay in their own lane and we just end up with the same site managed by fewer people. That's the conversation I would like to progress to assuming that these things would just infinitely get better. >> I mean I I I maybe you have a follow-up question James product guy. Um I think one of the things that we've been talking about in the last few weeks right has always been like from the product sense and engineering sense has been focus right like figure out what's the main customer problem that you have and build towards that ICP and all but then >> what is ICP >> sorry ideal customer profile right like focus on who you're trying to selling to their problem walk backwards from that but and there's always been the notion of like don't listen to the noise, don't follow all the shiny things. Um but now with I don't know agent AI like should we be taking our central product and oh that person wants a little bit of a different variation. So just run it through the agent machine and make it compatible to that unique use case that no one else is going to use but now it's just such a low lift >> until it's not >> expand. What do you mean by that? I mean, maybe one day the agents will also support it super easily for you, but you're going to end up with a hundred a thousand one-offs and anyone who's had to support software, including yourself, probably wins this. >> Yeah. I mean, things can change really fast and I think if you want to know, go ask Stack Overflow. It's essentially out of business. I'm serious because like it was all good until it wasn't and now no one's using Stack Overflow. I mean not no one, almost no one. >> So it changes really really fast. One of the things and I know Ron you don't already know this but we started thinking about um defensibility uh for you know we're a software company if software is free well maybe they don't need our products. So there are some things that I think are defensible and so we started like writing those things down. So I'll give you an example. Uh network effects are pretty defensible. So LinkedIn is really popular, right? Employers go there and say I want to hire employees. Employees go there to look for jobs and they share their stuff. That is a network effect. And because they're there, there's a flywheel going. I can ask claude to build me a business social network and it'll do it really fast but if no one's there it's not valuable right so there are some things that are defensible and that we are trying to think about those kinds of things systems of record like workday or salesforce and these kinds of things those are valuable too and just because cloud can do some of that stuff doesn't mean people aren't going to do those things so I think those are the kinds of things we're trying to think about in terms of what new things that we can build now Should we build all those things and which which things make sense? We definitely think about the things that we are building that are going to have sustaining value and which things maybe aren't as differentiated as they used to be. >> I don't think we know yet what those things are. Like I think some of them will really surprise us. So I read a stat today that said that readers um visits to like tech journals and tech online sites is down like 40 something%. Um, and I think all small creators, except maybe video makers at the moment, are suffering from this because now I go type my search in and if I have Gemini turned on, it gives me the answer. It re it references the website, but often I never need to go click on that website to learn what I wanted to learn. So, I think we don't yet know what what's going to sur be. >> So, so someone someone actually told me something fascinating today. Um, and it seems so obvious, but I was like, "Wow." Um, someone told me like, "Most of the websites right now are really wrong." And I what do you mean by that? And I say, most websites today assume that the person coming to your website wants to learn about your product. Okay. Okay. What's wrong about that? Right? Because like that's why we have a website. It's kind of like our billboard, right? There's like the use cases all like she told me no. She said by now most of the time that a person actually gets to your friend page it means that there's a high likelihood that some model or some answer agent engine already gave it all the information they needed and now they're in a I want to either try it or dig deeper right so I I think there is a lot of paradigms that are that are pretty aggressively shifting right you know right now um and I love the fact that someone actually raised a hand uh over there. No, no, it's okay. I'll repeat it. Don't worry about it. I'll repeat it. Don't need >> my question for you is Thank you. So, my question is hearing both of you talk about how, you know, the decline in media being generated and media being consumed that's being written and you talk about the decline of Stack Overflow that it's basically a graveyard, but the models trained on that content and that's how they got good at it. So, what's the next generation of models going to look like if people aren't generating human content? So, um I don't know if everybody heard the question, but you know, it's the it's like kind of like the AI inbreeding question, right? Uh what happens when we run out of human uh like actual human content? Uh right, Stack Overflow is dead. We're not necessarily contributing to it as as we used to. Um what what are models going to be trained on in 2027? >> I don't have a great answer for that. I think it is absolutely a risk because and I I worry about this for open source right there one of the the hot takes I've seen out there is like oh uh if let's say like some of these libraries that are JavaScript or you import like you know some 200k size node library to change some color on a font or rendering or something like that and you could just have the the AI model instead just gives you a 10-line snippet instead and then all of a sudden what used to be a really popular, well-maintained JavaScript library doesn't have as many people using it anymore. And what's going to happen to that library, right? So, I I don't think we we actually know whether that's, you know, content that's going up, whether that's a well-maintained open source library. I don't know. I I do know that there's going to be big changes. Uh I just don't know how bad they're going to be. Like one of my the theories is that like the really proven low-level libraries that have hardened um like compatibility matrices like that's really important stuff. You probably can't vibe that and just expect it to work. But if it's just like a simple font thing on a on a library renderer that's probably going to get replaced with you know 10 lines of you know model generated code. So I I keep trying to look at history and I'm I'm not a historian at all. Um and usually I look at like the internet or cell phones or something in technology to understand technology, but in terms of content, I think it will just shift. I think humans are creators and we're going to continue to create at the edge of what technology allows us to create. So I'm sure like we're we're missing some beautiful things in the past like some of the woodworking, some of the the quilt work, some of the knitting. Like there used to be beautiful things created that we don't create very often anymore. Like it's a hobby now. But we create some amazing things that weren't possible like a hundred years ago. So I hope that we just shift up and we create things in technology whether it's solutions or art that we can't even imagine right now. >> Yeah. I know we have a few questions and and I I don't think I can just repeat them as well. I don't want to make you guys run. I think while you're asking that one, I think there's part of me that says, "Man, I loved all the skills we obtained and grew. I love the curiosity that that brought about. I love being able to ask those questions and get the answers." But there's also a part of me that's realizing that maybe the UX that we put out there wasn't good to begin with. A lot of these REST interfaces weren't good to begin with. A lot of these system designs weren't good to begin with. And the demand for backwards compatibility means we got stuck with a lot of these systems for way longer than we should have. We've indust we've industrialized this stuff where we're training people on those things and giving them certifications and we're kind of slowing the industry down. And so as someone who's like bit of an AI skeptic, the one thing that I think is correct is that there is a challenge that should these have ever been the interfaces >> and if you can ask a thing, forget the implementation detail, but if you can ask a thing and we get the other thing, that's a really good API. That's an intentbased API. We spent our entire industries or at least my career building these little Lego bricks of REST APIs that don't really go together and then I watch this write documentation with just hints. I don't know why we do this. You build a standard library and you put hints on how to use it. This function creates an SSL certificate. Good luck on learning how to use it. And you're like, but but you wrote the library like you know everything right now. You know everything. Could you just give me an example of how to use it? Maybe give me an example of how not to use it. Like no, no. Go ask on Stack Overflow. And we did this for decades. We come up with new technology, no manual. and the community jumps in and we all like, hey, even though there's no docs, here's how I think you should use it. And then I've extended it and here's how you should use it. And then we go on this discovery mission every time. It was never good. It was never good. So now we have this machine that's like there are no docs for this, but I've seen the corpus of usage and I can spit out an example for the first time. Even if that example isn't good, it's still better than nothing. And so this is where I kind of give this technology grace. Search engines suck right now. Ads, ads, ads, ads. Maybe right or wrong blog post or outdated. Just give me the example of what I'm looking for. So this intentbased system that we're developing, maybe that is the way we should have been doing it. RPC versus REST is a dead discussion. I don't want any of those. I just want an intentbased thing. Give me a VM with some storage. I don't want to call 7,000 endpoints to coales what a VM looks like through Terraform. Is that really a good design? No. And so I'm actually looking forward to people maybe rethinking the way we build systems to be more intentbased than a collection of components that you glue together if you know how. >> Hey. Um so you know 150 years ago master weaver wo fabric made a cloth make garment out of it. Great. Obviously industrial revolution gets cheaper faster. Um fast forward now fast fashion. If you wear a hole in your sock none of you are probably going to darn it, right? You're going to throw it away because it's now industrialized. It's disposable. If what people are talking about the last couple months in particular, if that sort of progression is true, then in the future, we'll have a future of software being disposable. Not all software. There's going to still be artisans creating crafted software that does a thing, but there's going to be a lot of cheap good enough to use until it you get a hole in it and then you throw it away. Um, so I guess my question is, um, if you could talk about if that's true that we, if we were in that future where software is now disposable, um, some much of the software, maybe not everything, but some of it, um, what does that look like? >> I think I think my tolerance for holes in my socks is much higher than holes in my banking software. But >> so so I I think u I have some thoughts on that. I think I approach it from here's here's an opinion that will probably completely be blown out of the water in the next uh cloud version, right? Because who knows? Um my latest opinion on this is that we're going to have a instance for I don't know how long where we're going to have I'm coming from the Nyx ecosystem so I use pure and impure, right? We're going to have software that needs to be pure because of what it does and what it's reliant on and how critical it is in whatever we're trying to do. And there we're going to have a lot less tolerance for any holes, right? And and and then there's going to be software that's going to be impure. We saw that a decade ago. Who knows what Wix is or or website, you know? It's like I I don't know. I want to put on a website for my gender reveal. Like, do I care that it has 50 holes in it? I don't care. Whatever. Um, so I think I think the pure and impure pieces of software in the modern society are just going to be different. And the impure ones, it's going to be okay to build it and then rebuild it again if you ever need to and throw it away. Um, but I think this touches on a very open-sourcy point to me that James was starting to allude to. um where I think we're still going to be human physics inside of software where we're going to try and I hope we're going to try to figure out recycling uh or reuse which is I think one of the baselines of open source right it's like hey we all need this library let's put it out there work on it together now we don't have to redundantly recreate it across millions of machines every time that it's needed so I think we're going to start finding that there's going to be core pieces of our infrastruct that hopefully we can again package and maybe put it somewhere out there that other people can use. Maybe we'll give it a name. Uh maybe we'll call it, you know, open something. But I I do hope that we're going to actually lean there. So there's going to be pieces of our software architecture that maybe in the future agents will contribute back to open source with those pieces. Um but that's just some thoughts. I know. I mean, I love this question, by the way, because I'm sitting here thinking about what would happen. Like, we have a huge spoken language library. Like, not a lot of new words come up all the time. Dictionaries are pretty thick across multiple languages. Alphabet has been pretty stagnant. Like, I don't know what it would take to add another letter at this point. >> Yeah, you Yeah, you would kill the song. And so you have all of this vocabulary and I'm pretty sure the people who are imagine there's probably a group of people like making words all the time. Orange. Oh man, you got another one. And I could imagine that if that's not what people are doing, creating alphabets and new words, then they're creating movies and books, right? They're they're finally doing something with the words, right? That's where the work is. And if so software maybe shouldn't have been this hard for this long. Like it's still surprising to me. Like I hear people devops, s platform engineering and I started my career in 1999. I'm like we were still trying to copy an application to a server and run it. All of this energy going into running software. It's insane. So what happens if we don't need to do that anymore? It doesn't mean that it means maybe software and all the things we did to make software are no longer important. And if we could do that, would you allow it? I think I would. Now, it's unfortunate for all the people who've made a living doing that. And I have empathy for those people. It's probably where my mind sits most of the time. But if I put that to the side, should software be this hard? If it wasn't this hard, then I I went to the dentist recently and they gave me this clipboard with a form on it and they gave me a pen and they asked me to fill out all of these. It's 2026. You're still asking people to fill out paper on a clipboard. Then I give it to the person and I know my handwriting is terrible. So you probably think I have every one of these diseases that I said no to. And then they type them back into the computer. Maybe if it was really easy, they would all just have kiosk that rival the same kiosk that the big vendors can use with hundreds of developers. So, I think it's a good psychological question is if we got software to the point where fabric is, you can have fast fashion. You can have someone that wants to make that $5,000 suit. You can have all of those things. It wouldn't be the hard part about creating the material. And maybe we've just getting to that point where software should never be this hard to get. Just an ingredient for the thing that actually matters because I think we spend too much time on the pen and paper versus the movies that people watch. And maybe it's just time for that to go. And I think a lot of us are holding on because that was our profession. That was our hobby. That was our passion. But if we zoom out a little bit, maybe it's time. We don't need to do this for another hundred years. 2024, are we still doing DevOps? Hope not. >> I did want to build on one point around this idea of like let's say there's it's a lot easier to create and it's really cheap. Do software engineers have jobs anymore? Well, there's a couple points that I found really interesting the last couple of weeks. I think it was within one or two weeks ago, like Aaron Levy, who's the CEO of Box on his LinkedIn, he put out a post and he showed a curve upwards for hiring software engineers for for job listings. And that's like completely counter to the narrative of, you know, everyone being scared that they're going to eliminate all these jobs in software engineering. And some of the analysis that came out about that was this this Jebans paradox, right? Where something gets cheaper and then that there's more consumption of it. And they're going to need people um that have an engineering mindset to understand how to drive these models and to make sure that it's with high quality so the model's not doing something crazy. and we're just moving up an abstraction level. And then another data point on this was um you know I saw people like really trolling Anthropic because they had a software engineer listing for something like $500,000 for their I don't know if anyone saw this but they're like if we don't need software engineers because Daario from their CEO was saying we don't you know this this job field's going away why are you hiring them still you know which is another interesting point. So I I don't know how it's going to go, but I do feel like raising the level of of abstraction, the skills that we've been building up will still be valuable, but we'll have to adapt. >> What's the question? I'll I'll repeat it. >> Kelsey said And that was kind of the ide don't have that software already. A it's very hard to create that makes software easy tool or it just isn't that much demand for that kind of software isn't that much for software. there hasn't been a reason for those tools. >> So, so just to repeat the question if I correct me if I if I get it wrong. I think um you're mentioning Hyperard as a almost like a conceptual model of creating bespoke software and I guess you're alluding to the fact that you don't see that yet created in today's modern ecosystem where I can just hack my own bespoke software together in a very basic way. and why that's not there yet, >> right? what 40 years ago >> I you know that's a I think that I think this thing where um I don't know why the industry gave them permission to but every enterprise was like you know what we're all going to do custom everything that's our emote that's our IP that's our secret sauce we're going to figure out how to put things on servers better than the next company and that will be our strategic advantage and if y'all don't believe that that's how most vendors sell their software like you know you buy the software that we sell to you but we also sell to everyone And if you use it better than them, you will get to production faster or something something return on value. And I think that whole customization was just a lack of discipline in our industry, right? Like imagine someone's building you a house like, "Hey, we're going to use a new material today. Not bricks, no drywall. Wood is out of the question. We call this antimatter. Hey, have you used it anywhere else before?" No, but we're going to build your house with it. So, we have no history, no metrics about whether this is safe or not. Well, look, we don't have to live in it, but it'd be really cool if we did it for you. We're going to write a blog post about it and everything, right? And so, our industry has allowed that. This is some of this stuff literally irresponsible. You work with the person like, "Hey, we should just rewrite all this in Rust." And you're like, "Dude, what the hell you talking about? This is a static website. It it doesn't need to be in Rust. No, we're going to make our static rush generator, right? And it's like, oh, okay. But a lot of the things that we do in our industry is completely unnecessary. Like I was a big Kubernetes person. I was like, Kubernetes is great. And then someone's like, hey man, I work at this university and we got three servers and it's really productive. It serves all the students with three servers, but I really want to get into this Kubernetes thing. I say, hey, stand to the side for a second. Listen to me very clearly. You see this book, you see the author is you have to trust me. Stay away from this. You don't have this problem. This is good for you. He's like, I don't get it. I was like, okay, there's some breakthrough like cancer treatments. Cancer is bad. It kills people. You don't have cancer. He's like, no, I don't. You don't need the treatment. He's like, ah, I get it. And I think a lot of the technology we see in some of these companies is completely unnecessary. some people just haven't stopped and said I don't I don't think we need that. And so I think a lot of this complexity that we have the reason why we haven't cryst or or settled on a set of technologies and let's be clear there have some people that have I've seen some people like man I just run my business on Heroku man I wrote this thing 10 years ago it's been running on Heroku ever since man I'm doing about a million a year I go to all these conferences and I see this stuff but I'm like should I rewrite this and I'm like but it just works. So there are some people that have found the constraints, conformed to them, and they're getting a lot of value right now. There's a lot of people that are winning on that regard. But I think a big part of our industry still sees this excitement around making something custom, even if the cost means 10,000 people dealing with that complexity going forward. And maybe it was necessary for the last 20 years, but I think what we're seeing now is like, you know what, how many Uber clones can you have? And I think that's just where we are. >> So this is a weird question. I'm listening to all this and it seems like there's a one-dimensional problem. Go put the promps in get the result and it's so mundane but what is QA? I mean it doesn't write perfect code sometimes maybe it does on small scale but you have you know a giant set of trading operations trading different instruments there's a commonality and so on and so forth but how do you go about dealing with that kind of complexity and expecting not to ever because I'm not hearing anyone saying oh I'm going to go after I've, you know, confed and gotten and, you know, reviewed the code or something, I can probably be sure that I can put it out there and, you know, test and all it'll just go right to production. >> So, so just >> there is there an Oh, yeah. It looked like it ran great one time, maybe two times. only one person effort. >> Yeah. >> Yeah. So, so just to repeat the question, I think um I guess what what is changing in the SDLC, right? Everyone's talking about co-creation and ideally putting that co-creation into production. Where's the testing? Where's the QA? Where's reproducibility? Um, you you just flagged I don't know there's >> Right. >> Yeah. Like a few NYX people probably glowed blue when you said some of that. Uh, so so I think I think you're you're you're spot on, right? Like I I've seen folks say AI developer life cycle or a Gent developer life cycle. I I kind of call there's no such thing. There's a software development life cycle. I um we just ran a conference here called Planet Nicks. It's about Nicks. It's really cool. Check it out. There's for some reason they're still in the hallway even though we don't we finished with the conference. Um and and one of the things that I did in preparation was I started reading some of the original kind of like uh software engineering thesis from from like mid 90s, early 90s just to see kind of where things were when when we were starting to build things out there. And one of the things that really resonated with me was was someone um defining software engineering as all we're trying to do is tell a rock a set of constant inputs and have the same input provide me the same output. >> Yes. And it's a rock. But I just want it to do a thing for me. And and I think I was I was kind of equating it to a light switch, right? When I come into a room and I see a light switch, I hope that when I flick it on, meaning give it an an input. Um, ideally, it's going to turn on lights in the room and not turn on and blow up some microwave in, you know, the mesh hall. Um, so I think we just have a new SDLC with parts of it. And you're right, I think a lot of folks are talking about the most exciting thing, which is the code generation, how fast it is. And I think that's also influenced Kelsey was mentioning VC and due diligence. It's also influencing where the money is going right now. Every model is trying to out compete the other model on being faster and better and faster and better at what at code creation. But I think I'm seeing a lot. So if you dig a bit deeper, you're seeing a lot of different companies and paradigms work on the rest of the SDLC. It's still there. It's 100% still there. I think right now it's still kind of open season to define what that SDLC is going to end up looking like, right? And I'm a biased Nyx person. So like Nyx and Flocks and I was just gave a talk about how reproducibility and determinism should be part of that SDLC at the bottom because that's really important for when we have uh uh those agents kind of running in the middle and sometimes doing things that are not exactly deterministic. Um but I think there's definitely work being done there. >> Is that done by hand? >> So so I I don't think it's being done by hand. I think it's for instance I'll give you a flow that I've seen recently. Um again plug for Nyx. I've seen it done on nyx where they actually have asked the models to pre-generate the tests that are going to validate the output that they expect right so so they actually started it's almost like if you remember TDDM right test driven development uh from from a little while ago it kind of came back now it's like no start with the test make sure that I'm getting what I'm trying to get now go do whatever you want to do in the middle and then at least my output would be somewhat consistent but um I don't know if you guys have more thoughts >> yeah I mean one of the things you you know's law which is basically like we can speed up the creation of software, but if I have to then go get it certified, the security team has to bless it, I got to get it through the production team. So all the software is coming into the top of the funnel really fast, but it's getting stuck in the thinner parts of the funnel that that take just as long as they took before. Now there's just a big backlog. We're going to have to figure out how to do those other things at at scale. Otherwise, the other stuff's just going to wait to deploy because it's behind. What's that look like? >> I I think if it looks anything like this in 10 years, this whole thing was a failure. >> Just think through this. We didn't built apparently this super technology to do the same thing we were doing before. That's a failure mode, right? When I go buy new tires, this person doesn't sit there and QA each tire. Why? Because they're dealing with known entities, right? Like this tire has the right tread. You can put a tread gauge in it. They fill it to a particular pressure with the standard stem. They put it on the car and they balance it and I drive away. >> Where are we in that time? >> Not in there. Oh, in the timeline. I think right now for 10 years, the business model that we've chosen to accept, you're going to be stuck in this timeline for another decade. Kubernetes, we're still doing the same things as before. We're deploying apps on VMs with a bunch of YAML files. Some of people, oh, this is a new way of doing it. I'm like, I was there in the beginning. It's the same thing we were doing before. And right now people are very frustrated with Kubernetes. Like dude, I'm doing the same thing as before and I have a lot more YAML files and I created a whole another industry on top of Kubernetes. >> But I think it's the same thing. So if you're going to keep generating software we're unsure of, yeah, you're going to keep testing it. Whether you ask the AI to do it or you do it, at the end of the day, your company's going to put their name on it and then your company's going to figure out what's required to wear my badge. If you think you can just let the AIS do everything, remember, and it's better now. What was that? Canadian Airlines, they unleashed the bot early and the bot created a refund policy and people were like, "Yo, look at this thing. You just get free money from it." And then they put their name on it. And so at that point, your name's on it. So whether the agents are good or bad, your name's on it now. So at some point, you're going to ask, "What will it take to put your name on it?" That's the QA process. Whether like if I use a very mature framework, you don't test all your open source libraries. Most people do not test their compilers. Guarantee it. You don't test your CPU architecture. You don't even look at the assembly it generates. >> That's your belief, but you're probably right. It's that stable tire. Now, right now, the software we're generating is reflecting the type of crazy, untested, unreliable, probably not even correct software we're used to. For some reason, we've accepted that as the checkpoint. So this thing is like I can do what you've been doing way faster. So of course to me this is just a fast way to do the same thing. So this is why I'm not as excited. Oh, it's going to take over everything. No, it just means you're going to be doing the same thing forever for $20 a month, maybe 2,000 depending on how many tokens you need. That would be a failure. So here's what I hope happens. I hope someone gets wise and says, "Why are we generating the same snippets of code over and over again? This is insane. This is silly. irresponsible. There was a musician, I believe, before AI. He just generated every melody and went to court. The ones that were spoken for, those are copyrights. The rest of them are public domain now. We're done with this debate. Every melody is generated once. So, feel free musicians to make the music. You won't be sued anymore for the majority of this, and it's only a matter of time before the other copyrights expire. And no more melody lawsuits. At some point, when do the models start contributing back? Hey, there's a million people using this tool. All of you for some reason are trying to do SAML integration. When do you just create the perfect library that just does that correct? And then it's well tested. It's bulletproof. You can test it again if you want. You're going to get the same results. So, I think we hopefully end up there. And if we don't, then we all got taken for a ride to start paying a toll fee to write software versus being able to do it for free. So, I know we have a little bit of time, so I'm gonna do a lightning round for folks that had their hands up. You go first. >> I won't repeat that part. He's very excited. Go ahead. >> Worries. might repeat that really quickly just for a sec. Um, he was mentioning if folks are not familiar, Mitchell Hashimoto was built a system called Vouch. Um he most recently was building Ghosty and and you know he was getting AI slop uh well AI generated uh poll requests and um and slop and and he wanted to do some verification of who's doing it who's not and and that you know and what was said here is that we've had a model especially in the open source domain has been working right humans coming together getting together to work on a common goal on a common piece of software that we all want to make better and it's just worked because of the human aspect of it. What if we just what if that AI model just comes in and breaks out something that's been working for 30 years or or longer? >> Yeah. >> So my first comment to that was always like these these LLMs learned from something. So we've had a lot of bad behavior out there that they've also learned in addition to the good behavior. Um, but a plug, I did a talk yesterday about this. I tried to make it super interactive like this one with everyone contributing what was working in their projects and what wasn't working. Um, so there were a lot of great ideas that that came out of that. Um, but I think it's it's a it's the other my my second point is always that that agent, we talk about the agents and in the news it's like the agent did this and the agent did that, but somebody created that agent and somebody paid for those tokens. Um, so that agent, maybe it went rogue, but somebody was responsible for it and that person didn't fess up for quite a while. Um, so I think I think we have a a human problem, a process problem, and something we all need to work out together. >> Yeah. Um, I'm going to I'm just lightning questions. Okay, I go there. >> Basically, first you said the content would disappear. No trusted would not the links I always go and check because it's hallucinate other fever of my son he goes at home maybe go to hospital your son that makes no sense so there's always a reason and sometime in my office senior given me these are the solutions you know this is copy paste from >> the human interaction where you have to tell it's up to you stop I don't want to so that that is no there's no plan for that and it is up to you oh I don't want to use it at this place because it's important for you because your son is sick you don't want to go to that doctor who gives you >> I mean I think to summarize and I know we're out of time here like reputation is a human thing. Accountability is a human thing. And if you cause harm to another human, whether it's issuing a bunch of pull requests causing me more work and pain, then I'm going to hold you accountable and just close them all. And I think we tend to work things out by saying you can't tell me the agent did it to me. That's unacceptable. We won't take that as justice. Humans will tend to govern ourselves in the way where there will be accountability. So, right now, maybe we're all excited and we're taking shortcuts right now. And maybe to your point around LLM's contributing code, here's the thing. If you're doing it right, I don't care if it's an LLM. When you issue a pull request, read the contribution guide, put your name at the top, follow the commit message, understand and have empathy for the reviewer, and then put the code in there. Whether you use an LLM or not, why would I know? It's the fact that I know that the LLM did it is where I have a problem. You didn't even take the time to look at this. >> This is a Golang project. Why are you issuing PRs in Rust? That's just laziness. >> Yeah. So, I'm gonna I'm going to just kind of wrap up. We're some of us I'm not going to speak for everyone. We'll have some time. Do we have a talk after this? Is there a talk? >> Yeah. So, we we'll be outside to not bother the talk. Um but I think just to wrap up, a lot is happening. A lot is changing. Um, I think folks here in this panel have had their job titles changed probably 10, 15 times. And I know those that are kind of at the later stage of their career are less concerned or not envious of those at the beginning of their careers. Uh, but wherever you are, I think I think we're going to figure it out. I'm I'm a naive optimist on where things are going. I'm hopeful that we're going to figure out reuse. I'm hopeful that these models are going to help us contribute more to the open domain. Um, but I can totally see and I can totally resonate where these things are kind of scary or spooky. The only things that I can like the thing that I would emphasize leaving this is go test it out. Experiment with things at your own pace, but don't stand on the side because I think right now the way that I view it is a lot of folks had had got a NASCAR uh vehicle delivered to their, you know, uh, front yard and no one knows how to drive one, right? But it's better to try it out a little bit. So if ever you need that or ever we get to a point where we need to know how to use that you know machine you at least have some experience. So again I want to say huge thank you to our balance for for coming down today and I want to thank thank you to scale uh and thank you all for joining us and making this a very live energetic discussion. So then you guys