Subscribe
Share
Share
Embed
Hosted by Matt Klein, creator of Envoy and co-founder of bitdrift, Beyond the Noise goes inside the minds of the engineers, founders, and technical leaders defining the next era of app-based computing. Forget the buzzwords — this is where the people shaping modern systems share how they really build, debug, and scale.
[00:00:01]
Matt Klein: Alright folks. Welcome to another episode of Beyond the Noise, Signals, Stories, and Spicy Takes, the show where we dig into the stories of the people shaping the future of app-based computing. I'm your host, Matt Klein, co-founder and CTO of bitdrift, as well as the founder of Envoy Proxy. Each episode we'll talk with engineers, founders, and technical leaders who have transformed the way their companies build and understand what's happening inside their systems. We'll dig into the challenges, the breakthroughs, the lessons learned, and we'll wrap it all up with their hottest takes.
So let's dive in.
Today I am thrilled to have Kelsey Hightower with us, who probably needs no introduction, but he is a former distinguished engineer, investor, contributor, and self-described minimalist. Kelsey, welcome.
Kelsey Hightower: I'm happy to be here.
Matt Klein: So thrilled. You and I have known each other for a while through a lot of cloud native endeavors. And I think probably almost everyone who's listening probably knows who you are. But just the way that I like to get started is just give us the quick rundown on who you are and how you got to where you are now. That'd be great.
Kelsey Hightower: Yeah, graduated in 1999 from high school, bought an A-plus certification book for $35, turned those pages as much as I could, got certified and just really started this whole learn as you go kind of career. And the first 15 years -- maybe the first 10 years -- I think I did every job possible. VoIP, enterprise, financial services. And then I landed in the world of open source. So contributed to things like Puppet, learned Ruby, Python, core utilities there. And this whole cloud native thing blew up around that 15-year mark in my career. And I found myself staring at Kubernetes and the whole container movement. That's when our paths crossed. You did Envoy, which became the core of Istio and Service Mesh. And that whole cloud native thing was like my last 10 years in the industry, which was a lot of open source, a lot of industry standards, and really just making a lot of the things that you saw on white papers become real via open source projects and industry standards.
So I retired about two and a half years ago, and I guess -- retired in air quotes. I still advise a lot of startups. I still do some VC and investing work. I still give my opinion on conference stages. So I'm retired, but not tired.
Matt Klein: Ha. Yeah, I mean, one thing that I would love to ask you, obviously prior to getting into the AI chat -- which is why we're here -- is, from my perspective, you are the most skilled person that I have ever met when it comes to relating to other developers. I mean, when I think of a developer evangelist or just someone in this space, I can't think of anyone other than you. You're just so good at it. And one thing that I've always wanted to ask you is where did that skillset come from? How did you get to be as good at that as you are?
Kelsey Hightower: [00:03:21] So, yeah, think about it -- the ability to tell stories, that's like a human thing that has existed for a very long time. In our industry, to be a software developer, you can be terrible with people. You can have the worst communication skills ever, but if you can make a computer dance and the compiler build your code, you have a job. You can even probably become a distinguished engineer with no communication skills, depending on the company.
Matt Klein: Can you these days? Really? Is that still possible?
Kelsey Hightower: I bet you can. I guarantee there are a bunch of principal engineers sitting somewhere who just refuse to talk to anyone. They've outlasted everyone and got that promotion because they've been there 20 years. But communication helps. I always thought it was one of the best human traits you could have. If you're a politician, you need to learn how to communicate. If you're a business owner, you need to communicate. If you're making movies, you've got to tell a story through film.
And software is no different. A lot of times on a team you're trying to communicate direction -- like, do we use Jenkins or Argo CD? So what's the story? Why are we going with one over the other? We can make them both do what we want, so why this one? A lot of the self-learning I had to go through -- the first person I had to convince was myself. I'm very much a skeptic on everything. If a new thing comes out, I'm like, wait a minute, is this deserving of the hype? What came before, what came after? So I go through that on my own, and once I'm convinced, once I'm excited, I found the ability to just translate that back out. And you understand that you have to look at people, you have to be patient. Words matter, so choose them wisely.
I put the same effort into learning technical skills as I did into communication skills. A lot of times we see a lot of senior engineers and a lot of junior humans -- maybe they didn't have to go convince anyone of anything, they just had the best technical skills. I've always tried to balance both, especially when you're talking to a broad audience.
Matt Klein: Has that always come naturally to you? Like when you were a kid, were you convincing people through excellent communication? Or is it something you built up over time?
Kelsey Hightower: [00:05:52] Yeah! Yeah, I went outside. I was born in '81, so we went outside and you had to talk to people -- like, hey, do you want to play? So you're convincing other kids. Then you're on the basketball court trying to prove yourself to get time with the seniors. These people are older than you, why should you get to play? And of course you're talking to girls like, hey, can I get your phone number? There's a million ways to ask for someone's phone number. So learning how to be in those awkward moments and still find the right thing to say -- I was practicing that constantly.
Throughout my whole career you've got friends, you're trying to make each other laugh, you have things you want in life. Learning how to communicate and express yourself in a way that got you those things -- I think just being social, being outside with strangers, learning how to talk to them, learning how to be in uncomfortable spaces and still hold your own, I carried all of that over into my career.
Matt Klein: For sure. Yeah, I think too -- you were saying there are some crusty senior principal distinguished engineers that don't talk to anyone, and that's probably true. It makes me sad. But I think at least in my experience, people have really discounted in our industry how important it is to actually write well, to actually be able to communicate ideas. And anyway, it's fantastic to see what you've done. You are a role model to me for sure.
Alright, so let's get into why I had you come on. Obviously everyone out there knows we are in the major AI hype phase right now.
Kelsey Hightower: Awesome, appreciate that.
Matt Klein: [00:07:48] A lot of people on LinkedIn -- probably with AI writing their LinkedIn posts about how amazing AI is. And from your posts, you are a bit of an AI skeptic when it comes to software development. I have my own opinions on this, which we can talk about. But from someone that's so prominent within the industry, I'd love to hear from you on how you're feeling about the current situation. What excites you, what really concerns you?
Kelsey Hightower: [00:08:32] Just so people understand the level of consistency in my analysis -- any new technology I always approach from a position of skepticism. That skepticism is just the way I form the ability to be objective. So if a new programming language comes out, I'm asking myself, why is this one better than what I currently have? Is the standard library ever going to mature? So I start with skepticism, asking critical questions. And if I can get over that hurdle, then my skepticism turns into curiosity. What can I do with this that I couldn't do before?
And if that proves itself -- let's take Go, for example. I can cross compile without learning everything about GCC. Or in the Ruby world, I have this global interpreter lock preventing multi-threading, and you're telling me Go handles that natively? Great. And it uses way less memory than the Java thing I was working on. So now I'm optimistic. My curiosity turns to optimism, and I start using the tool and replacing things I wrote before. And once I'm in that optimistic state, that's when you typically hear me talk about it.
Now let's get to AI. Two years ago, most of this was just straight up garbage. The things people were excited about was just the fact that a computer could talk. It reminds me of people buying toys that could just talk. You know the toy isn't real, but you can pull the string and it says something -- wow, if I pull it again, I don't know what it'll say. And people got super excited just because someone created a toy that talks. Something they thought was reserved only for humans and super intelligent beings, and now your little teddy bear is talking.
I saw a lot of people behaving that way in the early days of AI. And then I watched people oversell it -- like, "this tax software is dead in six months." Why are you saying that? Do you have any evidence? When I see technologists move like that, it reminds me of the crypto movement where people were just trying to pump their investment.
Zoom out another year and the stuff is getting better. Maybe better training, better results, especially when we get domain-specific, like writing code. A lot of people forget that generating code is much easier than generating natural language because the domain is much smaller and we have lots of code examples. So I got curious. What can an LLM do that I can't do today? And after trying it -- wow, this beats going to Stack Overflow, scavenging through Reddit posts to figure out a direction. That's pretty good. Now I'm optimistic.
But during that discovery phase, I opened the mail and Anthropic is like, hey, you're part of a class action lawsuit, mainly because we've taken your content -- your book, Kubernetes Up and Running -- and used it to train our model. So thanks for the IP. We're just going to give lawyers a bunch of money. Maybe give you two dollars in total. But when people talk to our tool, they're getting your content essentially for free.
And that just reminds me of the center of gravity in terms of the power structure, the way LLMs represent themselves. This is very different than programming languages. We're trying to outsource not just our code and automation -- we're outsourcing the thinking. These things give responses and sometimes there's no credit back to the original author. What we're doing is removing a whole part of what I thought made tech amazing: the community, the people. We're almost erasing that group in favor of giving all the credit to the model. I just don't like that part.
So that's the way I currently think about LLMs. I'm maybe over-indexing on the harms because any technology can be used for helping or harming -- or allowing you to do either at an accelerated rate.
Matt Klein: [00:13:17] Yeah, I mean, we can talk a lot about the trademark issues and we should, because that is something that concerns me too. I think before we do that, though, I've seen you post recently less about the trademark issues and more about what it means to be a software developer. I've been seeing you talk a lot about... I don't want to take words out of your mouth, but it does seem like you're concerned not only about the IP situation, but also about -- is this actually a productivity increase? Or is it less clear than people are making it out to be?
Could you say a bit more about how you think about it purely from an engineering perspective -- what does it mean to be using these tools now?
Kelsey Hightower: [00:14:27] I built things in a social setting. Before you start writing code, there's a bit of setup required. I remember when we went to the Lyft offices -- you were there -- and I was there to learn a lot about Envoy, the reason why Envoy exists. And the Envoy you all built was very different than the vision that Istio had. So before I build anything, I don't mind getting that context from the actual creators of the thing. Hey, when you built Envoy, what did you have in mind? From that conversation I immediately saw the tension between what the community was trying to do and what the original starting point was, and what needed to happen to smooth things out over time.
People with that context, when they get it before they build anything, tend to end up writing less code because they understand the incremental step that needs to occur -- or they just have more empathy. Maybe there's nothing to do at all. Maybe the thing is working as designed. So that whiteboarding session with other people, that context gathering, the skepticism of -- do we even need to do anything? Is there something that already exists that solves this before we write even more liabilities, aka code?
So when people talk about LLMs and say, "oh my god, look how amazing this is, we're writing more code than ever," I'm like... okay. I can see where code could be your biggest bottleneck. It hasn't always been the biggest bottleneck for me. But then I ask -- okay, is your backlog gone? Because if you're getting a 10x productivity boost, your backlog should be disappearing at a rate we've never seen before. And also, how well is the team now functioning together? You don't hear a lot about team cohesiveness getting better. I just feel like the outside world is over-indexing on code generation, not necessarily on building software.
Matt Klein: [00:16:53] Yeah. To set some context that might help frame our conversation -- I've gone from two years ago like you, thinking this is garbage, to now... if I'm being honest, in the last three to six months, 90% of the code I output is written by an LLM. And without sounding like an asshole, I'm one of the better programmers in the world. And I am getting a lot of value out of these tools. This is where I'm so torn, because I feel like we're entering an era where very good engineers find these tools to be superpowers. Because people who know what it should look like get to move faster.
But my biggest fear right now is exactly what you're talking about -- how do we train the next generation? How are they going to know what things are supposed to look like? These tools don't think, they just regurgitate what they've been trained on. Without people that know how to communicate and build systems, it feels like we're setting ourselves up for a doom loop within the next four or five years. We're not hiring junior engineers. I don't know how they're going to learn how to do anything.
And that's the part that really concerns me. The code is trained on all of our work -- all of the open source, all of the books. These tools are real, they are a productivity boost if you know what you're asking for. But if you don't, I am very concerned about the future of our industry.
Kelsey Hightower: [00:19:32] Yeah, you have the benefit of decades of experience. You've built that context. You've done it the hard way. So you can appreciate where it saves you time and you know to avoid the areas where it doesn't. The rest of the people may not get that. What they're looking at is, I don't know if this is good or bad, but it works. And if that becomes the norm over the next 20 years -- I don't know if this is good or bad, but it works -- then we go into a different paradigm. The model vendors decide what the future of software looks like, because the people who could challenge that, there aren't any more of them. You will be the last Jedi. You'll be able to wield these tools and make them do the right thing. The next generation may not.
Now, to be fair, there is an opportunity to say -- maybe syntax and the way programming languages were designed, maybe they were just terrible. Maybe they were never suitable for humans to get things done productively. So maybe we've been looking for a better way to describe intent, and you can look at these new tools as better programming frameworks. You describe what you want, you give it context and structure, and out comes the other thing. Maybe they become super compilers with a much more flexible syntax. I'll give it that.
But when I talk about AI, I'm also talking about what happens socially. I'm asking people -- are you getting a raise because you're more productive? And they're like, oh nah, Kelsey. 30% of my team is gone and we're not getting raises this year. Any benefits are being attributed to the LLM, not me.
Matt Klein: Yeah, but that's total bullshit. We both know that, right? 30% of people's teammates are not gone right now because of AI productivity.
Kelsey Hightower: [00:21:31] No, no -- I guarantee, for a fact. There are people who work on teams, I went to go visit some of them because companies were trying to get rid of more of them. And it could be our fault. If you've gone into mediocre mode -- you hate your job, you're just churning out Jira tickets at the most susceptible pace possible, you didn't do anything to improve anything, you went mediocre for the last 10 years -- when a tool shows up doing just as good as you would have done because you obviously don't care about this job as much as you used to, that tool is good enough. We're not hiring anymore. We're not backfilling anymore. We had 100 developers, we've got 70 now, and they seem to be keeping up with all the issues thanks to these new tools. So we're just going to keep that trend going.
Matt Klein: Yeah, but do you think we know that that's true versus just that people over-hired a lot during COVID? It's very hard for me to tease apart right now what portion of this is genuine AI productivity boosts versus just over-hiring correction.
Kelsey Hightower: [00:23:03] Well, I do listen to the people that say explicitly that's what they're doing. There are people explicitly saying, hey, we are leveraging AI so we're going to reduce headcount. I also actually know managers -- I won't mention their names -- who knew this was coming six months ago and recommended that it happen to their teams because they saw the productivity boost from AI. And on the surface, there isn't a lot wrong with that when you have a plan for what to do with those people.
So for the rest of the people in this craft, what's the conversation for them? You said earlier we're not hiring as many juniors as we used to. So if you're a junior person, what's the conversation for you? Go do something different? Go be a dentist, go be an electrician? And there's going to be some people who get caught up in the "we don't need you anymore." A lot of people feel like -- if you're a little older, that's going to be more challenging to justify your existence, unless you bring something like what Matt Klein brings. Like, you have a lot of context, a lot of expertise. But everybody ain't Matt Klein. So if you're at our age where there's gray in the beard and you're not --
Matt Klein: I was about to say -- you and I both have a lot of gray going on.
Kelsey Hightower: [00:24:48] We have a lot of gray. We also have very strong reputations. I know a lot of people that have a lot of gray without very strong reputations, and they're struggling a little bit. I'm not going to blame AI for that. I'm just saying these are byproducts of any innovation. What do you do with those people? That's just me. I literally think about those other components.
I actually just made a post the other day -- I actually don't care about LLMs or AI. I don't care about Kubernetes. I don't really care about inanimate objects. I do tend to care about people. And that's the core of that answer I gave you earlier about communication skills. The reason people see me as an effective communicator is because I'm actually talking to people, not the tools. So I don't want to be in the LLM bashing camp because I don't care enough about an LLM to bash it. If an LLM gets really, really powerful, I'll find myself using one. Why wouldn't I? But I can't ignore the other part of that equation.
Matt Klein: [00:25:58] No, I mean, I think we're honestly saying the same thing. And I think for me, I'm already over the hump. These tools in the current iteration -- and we can have a whole separate conversation around whether the economics work -- but I have long maintained that right now we have three major companies: OpenAI, Google, Anthropic. They're literally building the same product. In the last three months, I've used Gemini, I've used Codex, I've used Claude -- they all do the same thing. So we can talk about whether this is all going to implode for economic reasons, but I'm over the hump in the sense that these tools are useful at the current price point. I am getting productivity boosts out of them.
But I share a lot of your concerns. I'm petrified about the future and how it's going to play out. And many would say that any line of code, whether it's produced by a human or a machine, is a liability. We're now potentially producing monumental amounts of code that is a liability with no clear owner, no attribution, potential IP violations -- and that's something I think about all the time.
Kelsey Hightower: [00:27:37] Let's also talk about the net win here. I love when I see the DeepSeeks of the world, the open source alternatives. Some people say, Kelsey, they're a year behind. I'm like, that's really good. If you're telling me they're only a year behind and they're focused on free, accessible, and open source -- I can live with that. Because that means in five years, if you're saying it's useful now, they're going to be incredible. The open source thing doing its thing is a beautiful outcome to me. It means while those big three may have the very best because of the capital involved, the community has something great too. There's a future where I can leverage these tools to build what I want to build on my own terms. So when I look at it from that standpoint, I'm actually okay. We found a better way to disseminate information for willing parties. Autocomplete was amazing; this is another dimension, and it's accessible to everyone. I wouldn't throw that away.
Matt Klein: [00:28:49] No, these tools are not going away. I tend to agree that the future may be local open source models running on my computer that I fully control. But I want to come back to something, because I've seen you post a lot recently, less about these technical topics and more about the human impact of what these tools mean. And you talk to a lot more people than me. When you talk to people who are younger and getting into the industry, or people who are older and not as fortunate as you and I are -- what kind of advice are you giving them? How should people be thinking about work in this new world?
Kelsey Hightower: [00:29:59] First thing I have to tell them, given the context -- I am in a privileged situation where I don't have to work. I want to make sure they understand it's easy to talk about these things from that standpoint. But I also remind them that I ground my way up, so I do have some experience that may or may not apply now.
So typically, I do probably two or three phone calls every week from strangers -- from LinkedIn, hey Kelsey, I heard you actually talk to people about this stuff. And usually they start like this. If you're young: hey, I'm at Georgia Tech, I'm a sophomore, I'm looking at all of this talk and I don't know if there's a path for me. They're not hiring juniors as much as they used to. Kelsey, what would you do in my scenario? And I say, let's jump on a call.
They're listening to the industry, from news to financial markets to developers like you and me, and there seems to be a theme -- juniors are not being hired like they once were. A lot of people halfway through college are saying, did I make a bad choice? Should I have gone into another field? And I say, I don't know what other field will be protected from this. Code is one, law is another, there are lots of fields that will be impacted.
What can you do? I say -- look, if I were you, I would be scared like everyone else. But I would pick the winner -- Claude Code or something like that -- and learn everything I can about those tools, because if that's the narrative, that's what will be required of you. When I started, I had to know five years of Linux, Python, Bash or Perl. Those were just the requirements at that time. No one cared if I agreed with them. So step one: go learn the tools of the trade. Bet on the future. But then hedge on the past.
And the hedging on the past is this: even though people say the fundamentals will be outdated by this technology, I would hedge my bets on learning the fundamentals -- how an operating system works, how a database works, how people work. Go to a meetup, but don't just sit there, give a talk. Train your own model. Be able to express these concerns and the solutions you're finding. So that's what I try to give them -- you're going to have to do a little bit more than what everyone else is doing. How many people are actually working on communication skills? How many are actually learning the fundamentals versus saying, maybe I won't need them anymore? Just do what most people aren't going to be willing to do in this new era.
Matt Klein: [00:32:41] Yeah. I think similarly about it. And this comes back to what we were saying -- I actually don't think the fundamentals are going to go away. Maybe in five years the context windows will be so big that they can bring in all information and do amazing things, but at least for right now, knowing what we know now, these tools are powerful when you have a lot of context around how things work and how things are supposed to look.
I do think there's going to continue to be a need for a lot of engineers. But your skillset is changing. Anyone today trying to get a job who doesn't know how to prompt, or at least roughly how the tools work -- as funny as you and I might think that is, it's kind of just the reality. You have to adapt.
I would give people the same advice, which is it's an exciting time to be in engineering. I think these tools are interesting and they do lead to productivity boosts. I just don't know that we as an industry are thinking carefully about all of the repercussions. And further, as these tools encroach on jobs outside of software engineering -- which has really been the first push -- I think that has a lot of larger societal implications that nobody is really thinking about right now. I don't know. Yeah.
Kelsey Hightower: [00:34:26] Look, if people were hyper excited about people having their needs met, I would feel very different about what we're talking about. But a lot of people who sit at the top of these power chains, that's not top of mind for them. It's almost like they're oblivious. They'll just go do something else. And you say, well, what will they do? And it's like, I haven't even thought about it. Because for me, whoever gets control of this technology first will have pricing power, will have the ability to dictate what happens next. If I can create a co-dependency, I don't see a problem with that -- all I hear is revenue growth. Where is that 20% year over year growth going to come from? That co-dependency.
So again, going back to -- I'm really, if I'm going to be supportive of any of this technology, it's probably going to come from the open source side. We saw the same thing with Linux, the same thing with Envoy, the same thing with Kubernetes. That open source thing, I think, is the better path to participate for the average person that missed the initial wave. That's the only thing that gives me a bit of optimism and hope -- that we can collectively decide to build something competitive and operate it under terms we think should govern these things.
Matt Klein: [00:35:59] Yeah. Let me ask you a tough question. Do you think that since you're in a very privileged position -- semi-retired -- do you think that gives you more leeway or ability to push back and ask these tough questions than other people might? Because what we're talking about is giving advice to junior engineers: this is just the way the industry is going, you kind of have to know how these things work. I would argue that unfortunately or fortunately, depending on your perspective, that kind of applies to everyone within the industry. If you're not at least exploring these tools, whatever you think about the societal implications, I think people are going to get left behind. Do you think that's true or not?
Kelsey Hightower: [00:37:03] If you need a job, you're definitely at a disadvantage in what you get to think and believe, because the job will dictate. People who create jobs, small business owners, they may find their ability to compete without using these tools is just fine. Remember, there are a lot of handmade goods that make way more revenue than mass-manufactured goods. But you have to be in that position to decide whether you want to use an LLM or not. If you need a job, that choice may not be yours.
So if you have a company that's all in on LLMs and you need that job, you're not negotiating -- you need the job, you're going to do what's necessary. But that's no different than any other industry. If they say you have to use this tool to do this work or you can't work here, that is normal.
But I do think I have the ability to say -- for example, I get a lot of email saying, Kelsey, if you would just post about our AI product we'll give you $10,000. That's all you gotta do. Just post it. But ethically, I have to say no. I don't believe in what you're building for these various reasons. So I lose the $10,000. And the most recent example was crypto -- someone's like, hey Kelsey, we made a project using your no-code work, there's $40,000 in crypto fees, all you gotta do is endorse this project. And what did I do? I shared the email publicly. Here's what they wanted from me and I just can't do that. Technically that cost me $40,000.
So I'm not saying LLMs are bad -- it's just not something I'm trying to push. And I'm not saying all AI is bad. Google Maps, point A to point B -- amazing use of technology. Spell check and autocorrect -- amazing use of technology. It's not that I'm against it. I'm just making sure there's a bit of balance, and I've chosen to talk, hopefully from an educated position, about the concerns.
Like, I learned a lot about MCP. How? I talked to people who work on MCP. I played with it locally, I questioned it. Why does MCP work this way? What problem were they trying to solve? Is there not a better way? And I learn in public, share what I learn, get feedback. So when I meet that younger person -- Kelsey, what's your thoughts on MCP? I don't think MCP was even necessary, but I get why they did it. They needed a plugin system. They did what they knew. It's kind of clever, but maybe we made a mistake the way we shaped our REST APIs in the past. We built these very small, rigid APIs -- you've got to call nine APIs to get something done. We should have built intent-based APIs. And I test this logic in public. So hopefully I'm just providing a bit more balance.
Matt Klein: [00:40:51] I mean, we've talked a lot about the programming tools of LLMs. But obviously these are tools with wide-ranging potential impacts on lots of adjacent industries and disciplines within our own -- whether that be design or writing or all of those things. And one thing I wanted to briefly touch on is, since you are a creative person, you've spent a lot of time giving talks and developing collateral that goes outside of just programming. Do you feel the same way about LLM usage when it comes to creative disciplines? Is it the same topic with the same answers, or is there nuance depending on whether we're talking about programming versus design versus giving talks?
Kelsey Hightower: [00:41:52] I think it is the same question and the same answer. With the right intent, technology has amazing upside for everybody. That is my base optimistic case. It's only when the actors involved change my opinion.
Think about audio production. Not everyone can afford a fancy microphone and a quiet room to record. Imagine having a $10 headphone with a mic, doing your podcast, and then running it through a tool that cleans up all the artifacts so you sound like you're in a high-end studio. That's an amazing gift to society, allowing people not to be limited by their socioeconomic situation. That's a net win. I don't know how anyone would argue against that.
What makes it bad is if someone takes my recordings and starts imitating me, tricking people into things. Scamming my parents out of their money because they think it's me. And so we say -- okay, there's a downside to that kind of technology, and we have to educate people.
But I think the part of the question that really matters comes down to human responsibility. If I went to court and they said "the LLM said that you're guilty," I wouldn't be satisfied. An LLM is a piece of software with its own weights and biases, tuned by one development team, one company. And now I'm being judged by that versus my peers. I think there remains this distinction between machine and human: lived experience, empathy for what it means to be a human, how hard it is to do things, the challenging decisions we have to make. When I'm judged by my peers, when I'm recommended what to eat, when I'm given advice, I'd like that human element to be a part of it.
So that's where I draw the distinction. And what scares me most is when people say -- is the human element even important? Once we get there, we're in a whole different world where you may not value human endeavors or human life, because the machine is a suitable replacement. That's where I draw the line.
Matt Klein: [00:44:28] Yeah, I think again we think about it very similarly. To me, at least given current technology, these are tools. They're not human thinking. I see behind you right now there are some very nice power tools -- I don't know what you're building, maybe you can tell us -- but it's a tool like an IDE or any other tool I would use. It doesn't think. I mean, it's a tool that I use. And I think that things have recently gone off the rails with people believing these can replace all of the human thinking and all of what we've built together.
That is not getting replaced, at least not currently. And that's what scares me about the current dialogue. I don't see it, and I'm not sure the current technology is capable of it. And I get concerned about what the talking heads say and the impact that has on the people coming up in the industry and all of those things.
Kelsey Hightower: [00:45:52] Well, that's why I feel it's necessary to lend my voice to counter those narratives. The average person may not even know how this stuff works. They're being told AGI is around the corner. Certain disciplines will go away. They even call it "thinking mode" right there in the tool -- they use these words on purpose. These are choices. The way they personify these tools -- look at how we named them. We give them names. So the average person is like, I guess this must be true, because I'm watching the Super Bowl commercials and financial analysts saying this changes everything, that we're going to create super intelligence beyond what any group of PhDs has ever been capable of. And that narrative is dominating.
So when I sneak my little part in there just to say -- listen, guys, let's not forget that we train these things through our lived experiences, our white papers, our blog posts, our source code. We are still part of the training set. It does aggregate technology with a much better memory than we have. So as a tool, it's amazing. But do not reduce yourself to what a computer can do.
Matt Klein: [00:47:10] Yep. Okay. I have two final questions for you. First -- what are you building? I see those power tools behind you.
Kelsey Hightower: [00:47:26] I bought a new house about two years ago when I retired and I was like -- I want to learn how to do everything in this house. Any maintenance, I just want to learn everything. There was wood laminate flooring on the bottom and carpet upstairs, and I prefer wood laminate throughout. So I ordered all the materials, bought all the tools so I could cut and lay it myself. And now I have it throughout the house. It took a really long time, but so does any endeavor you learn for the first time. I've been accumulating tools that enable me to do all of those things. I wanted new electrical outlets behind the toilets for bidets. I learned how to do that, bought all the tooling. For me it's like -- I love the ability to create things. I enjoyed it in the software world. And you can get the same feeling in the real world.
Matt Klein: [00:48:21] It is fun. Absolutely. Yeah.
Kelsey Hightower: And so that's what I'm building. I thought woodworking would be the number one thing I'd care about. I actually don't care as much about creating decorative things. I really care about the things around me -- can I make them do what I want? And it turns out drywall and wood are malleable and you can turn them into whatever you want.
Matt Klein: [00:48:43] Yeah, and now we're going to switch to a home improvement podcast. But doing drywall well is very difficult. It is a real skill. I do a lot of my own DIY stuff. I'm a pretty decent electrician and plumber. But when it comes to drywall -- wow. The people that are good at drywall. What an amazing skill that is.
Kelsey Hightower: It is. It's more art, right? We know engineering-wise you have to tape and mud and let things dry and sand them out. But the art part -- where you shine a light on it to see if you've floated it correctly -- that's why we float things out in a certain way.
Matt Klein: Yep. Okay, last question. Given where you're sitting -- and I'm not going to hold you to this -- where do you think we are as an industry in five years, in 10 years? Where do you think things are going to go with these tools?
Kelsey Hightower: [00:49:43] I think they're definitely going to get way better. And I think they'll get better because of more recall abilities, caching, not just net generation. Generating the same snippet of code over and over again is just silly. At some point we start caching a lot of these prompts and responses -- speed should drop dramatically. The number of pre-calculated use cases are just going to be there. So a lot more imitation learning versus this generative process. It's just one branch of this technology. But we'll get to this amazing place -- 256 gigs of RAM on your computer, maybe a terabyte on certain machines, and you'll be able to run domain-specific models locally. It's going to be really, really good. MCP will probably be gone in favor of something way more pragmatic. Most developers will stop building isolated REST endpoints and start thinking about intent-based APIs.
Matt Klein: [00:51:24] Just so people understand -- what do you mean by intent versus REST? What's an example?
Kelsey Hightower: [00:51:29] So in cloud, we built APIs to create a storage device, a network device, a VPC, a firewall rule, an operating system, a virtual machine. You need eight APIs. If you call them in the right order and reference the other outputs, you would get a running VM. But the intent was -- create a virtual machine I can log into. Not nine different APIs that a genius would compose.
Matt Klein: Right. Yes, makes sense.
Kelsey Hightower: The same thing for banking. How much money do I have? That's it. Can I afford to buy this thing or not? These are intent-based APIs. And so now, if we free ourselves from this rigid RESTful way of thinking, we might take that ambitious leap to say -- let's figure out what people are actually trying to do and give them the real APIs to help them do that. We've always shied away from that because we figured we can never know everything everybody would ever want to do. So we built composable ones. But I think you're going to see side by side -- composable APIs sitting next to the 80% intent-based APIs, overlapping in functionality. And it's going to be fine because you have better tools to help you maintain them.
And then I think the last thing we'll figure out is we're going to be forced to answer the question around the societal impacts of these technologies.
Matt Klein: [00:52:53] I was going to ask -- where do you think from a society perspective we're going to go? Do you think people will have jobs? Do you think there will still be engineers?
Kelsey Hightower: [00:53:02] So I think there are going to be a couple of benefits. If we allow the profits of this innovation to be shared -- your hospital bill should go down, the cost of certain things should go down. If we're sharing the benefits of it, then we should see the cost of certain things come down. And if that's the case, people will say, okay, that's a net win.
But then in the workplace -- if I had to say, in five years, I think developer average salaries or median salaries will be much closer to the firefighters and the teachers. Mainly because people who do good work will now have great tools to help them do that work. And if my theory around caching results holds -- imagine a standard library with a trillion functions in it. The thing you want to do has already probably been done. So now you're really just customizing software instead of making that new software. So why couldn't a person who's really good at listening and translating things and saying "yes, that feels about right" do that work? Salary should come down to meet those expectations.
The other thing I think will be amazing is that finally the people who are without tools will have some incredible tools. There are a lot of industries that don't have much automation at their disposal. Everything is brute force, pen and paper, spreadsheets -- just a lot of duct tape and glue. What happens when they get amazing tools? Will they be able to do amazing work? So I guess I am, in the most part, net optimistic about it. But that optimism is predicated on good people also competing in this space, aligned with the people who are purely profit-seeking.
Matt Klein: [00:54:49] I was going to say -- you actually sound optimistic.
Kelsey Hightower: [00:54:57] ...in this space aligned with the people that are purely profit-seeking.
Matt Klein: [00:55:02] Yeah, it's a good place to end. The part that gives me hope, purely from an economic perspective, is that I don't think there is any monopoly here. There are at least three companies, if not more, plus all the open source providers, all building identical products. And I do think the lack of an obvious monopoly position actually prevents, in the future, some of the terrible behavior we've seen from some companies I won't name right now. But you know who you are.
Kelsey Hightower: So we should clarify one thing -- the monopoly isn't the fact that there's only one company doing it. The monopoly is the fact that it takes almost a nation-state with access to huge amounts of capital and power to even venture down this road. It's not completely like Linux where a single person could compete with AIX or Solaris. But you're right, there are enough competing interests globally --
Matt Klein: [00:56:06] Of course, yeah.
Kelsey Hightower: -- that are all trying to make sure the playing field stays level for a long time. And I appreciate all of that.
Matt Klein: [00:56:19] Yeah. Alright. Well, thank you, Kelsey. That's a fantastic place to end. I think this was a great episode. I think people are really going to enjoy it.
So that's a wrap for this episode of Beyond the Noise, Signals, Stories, and Spicy Takes. Huge thanks to Kelsey for joining and sharing your story. You can find this episode and all past ones on the bitdrift YouTube channel. If you had fun, drop us a review, tell your friends, or yell your favorite hot take into the void -- just make sure to tag us. I'm Matt Klein and I will see you next time. Thank you.