AI is an accelerant, not a shortcut
I don't see any indication that the layoffs are really because of AI. They're because we've had thirty years of malinvestment: mass-market software so profitable that the hard problem became how to spend billions sensibly, and mostly we just wasted it.
Scott Meyer describes what he does as blood and guts system software development, the deep, unglamorous plumbing that most software quietly runs on. After four decades building big systems, he now works heads-down on a stealth startup, writing the kind of low-level code he says has barely changed in 40 years. His most visible creation is LIquid, the relational graph database that powers LinkedIn's connections at massive scale. He treats the current AI wave with a historian's calm, reaching past the hype for a homelier comparison: the pocket calculators of the mid-1970s, which panicked teachers about cheating long before spreadsheets a decade later quietly rewired how the world does business.
He arrived in Silicon Valley in 1995 from Colorado, the Silicon Slope, and has been programming for roughly 40 years. His first customers were oil-company geologists who happily paid ten thousand dollars for Fortran code that drew what was happening underground, a directness he still misses. From there he worked across computer graphics, networking, databases, and language implementation, eventually leading the storage team that built Freebase, the open knowledge graph Google absorbed. In 2014 he joined LinkedIn as a distinguished engineer to build LIquid, and more recently served as chief technology officer at ID.me before striking out to build his own company. Along the way he has, in his words, done pretty much everything once.
Meyer's thinking circles a single worry: complexity, and who pays to control it. He sees AI as an accelerant that cuts both ways, capable of collapsing a 4,000-line program to 1,500 lines in a few days, yet powerless against the industry's deeper incentive to ship the complicated thing that barely works. Tech debt, to him, is simply people shipping the first thing that functions. He is unconvinced the models are improving exponentially, asking why anyone would build an agentic harness if next year's model made it redundant. And he blames tech layoffs not on AI but on thirty years of malinvestment, the wasted billions of businesses too profitable to spend their cash wisely.
What energizes him now is narrower and more personal: programming feels fun again, because he can chase an idea without drowning in the complexity of deploying it. He sees a real, and possibly brief, window for anyone with something to build. His most hopeful image comes from chess, the first game AI mastered: rather than killing the game, it gave every curious child an infinitely patient grandmaster teacher, and human players got dramatically better. He wonders whether coding, writing, and video editing might follow. He refuses easy pessimism about people who coast, noting that couch potatoes sometimes decide, at fifty, to run a marathon. The best you can do, he says, is be there when they finally want more.
Read full transcript of interview
In this conversation: Josh Rubin (Host, CTO Studio) and Scott Meyer (Founder, Stealth Startup).
I start all these interviews the same way, and it's always the hardest question that I go with first. If you could just tell me your name, and what you do.
I'm Scott Meyer.
And what do you do, Scott?
Right now I'm working on a stealth startup. I've been a computer programmer for a long time, CTO at the most recent gig, and distinguished engineer at LinkedIn and Microsoft.
So what kind of roles in your career? What has been a specialty or an area of focus for you?
I really do blood-and-guts systems software development. So it's very old school, hasn't changed very much in the, I guess, 40 years that I've been doing it. Same sort of problems, basic computer science, controlling complexity, that sort of thing. So I've built a whole bunch of big systems. The most recent one was a graph database called Liquid that is the thing that powers LinkedIn.
And how long have you been in the Valley?
Since '95.
Where'd you come from?
Colorado, Silicon Slope.
Yeah, we're Silicon Hills in Austin. We're all derivative. All right, so you've been doing this for 40 years, you've been programming, working the space, so you've seen boom and bust cycles before. Have you ever seen anything quite like what's going on right now, though?
Well, I was thinking about this. History doesn't repeat, but it sometimes rhymes. And actually, when I was a kid, watching pocket calculators come out is very analogous to what's gone on. If you recall, it would have been the mid-70s. Suddenly, pocket calculators were an affordable thing and everybody could do a calculation. And the first people to freak out about this were students and teachers, right? There was a huge amount of concern over cheating, and would students lose numeracy, and so forth. Which sort of parallels exactly what's going on today with AI and writing and coding. The interesting thing is, for all of that, the market in calculation was actually not that great. The professionals had all moved on to Fortran. So it wasn't until a decade later, when you had spreadsheets show up, that the mass market for computation actually had huge impact on the world. For example, it probably enabled leveraged buyouts, mergers, all that kind of stuff, because computation of data was democratized.
So the opportunities presented themselves when you could finally do the math.
Right. People said, we can do the valuation of a company in an afternoon that used to take three weeks. So that was exactly the same sort of productivity gain, and it had a huge impact. But it was like a decade after the introduction of the thing.
So what you're talking about is both unintended consequences and downstream effects from something that you can't necessarily...
Just finding product-market fit, right? So if you look at AI, it's very clear that you can very rapidly get something that works, if you can define what "works" is. So for example, just working on a little project with Claude, in one day I have a working program, passes the tests, kind of does what I want to do. It's 4,000 lines of code. In another four days, I have that program reduced down to 1,500 lines of code. Now, the question is, who's going to do the extra work, and why? The pattern in the industry is: don't do the extra work, just add to the complexity.
So why reduce the code at that point? Are you just refactoring your own code to make it...?
Yeah, you just make it simple so you can understand it, right?
And the AI makes that simpler too. I have this conversation about tech debt. 40 years in, you've seen a massive amount of tech debt created over the years. Do you think AI leads to more or potentially less tech debt?
Well, it's an accelerant, right? So it's going to do both things. For me to clean this up went much faster with AI. I could just tell it to do something, and once we communicated, it would do things systematically, and it never gets tired, and so forth. So there's a whole bunch of refactoring that you probably wouldn't have done by hand that now you're able to do with AI. On the other hand, you still have every incentive to ship the very complex thing that just barely works. That dynamic already exists, right? Tech debt is basically people shipping the first thing that works. And if you're going to do that, then probably your software is going to get more and more complex and slower and slower and more difficult to work with.
People have always shipped bad MVPs and bad product. Now it's just, everyone can. So somebody's going to produce a better thing, so your thing is never going to actually get traction.
Well, it's actually kind of unclear. The other thing that's going on is the relationship to customers is very, very different. My first job, our customers were oil company geologists who were trying to draw what was going on underground. And they were happy to pay $10,000 to get some Fortran code to make those pictures easier to produce. So you worked directly for somebody who paid you money. Now those days are long gone. That hasn't been the case in Silicon Valley for 20 years. Nearly all programmers are producing code for somebody that they have no natural customer relationship with. You don't pay for Google Maps, so Google doesn't really care what you think about the API, or if you want this change or not. It just shows up.
Well, anybody who uses the back end of Amazon can attest that they don't really care what the individual user feels like.
Well, the question is, are you paying for a service? If you're paying money for a service, then the natural expectation is the company you're paying money to would be interested in your opinion, whether you want this thing or not. That dynamic... for example, Amazon tried to recapture that by saying, we're not going to have internal infrastructure that isn't also an external product, right? Pretty cool idea. But that's pretty unusual.
Do you see that changing at all with AI? A lot of people are talking about...
Actually, yes. That's the thing. Product is becoming more important. I pay money to Anthropic for a Claude subscription because I find it useful. And that is something pretty unprecedented in the last 20 years. Everything has been free, either ad-subsidized or just given away.
Well, except in the creative space. The one place where that hasn't been the case has been like Adobe Suite, in the creative front of product creation.
And they still have a paying relationship with customers.
Yeah, but they have to. So you're paying Anthropic, you're paying for a platform to use. You have multiple ways of engaging with it. Are you in Claude Code? Are you in the CLI? How are you engaging with it?
Both, mostly Claude Code. And the web seems to have gotten much better, because now it does the agentic stuff. I think about a year ago, the state of the art was you say some problem and then it spits out some code, and then you see if the code compiles, right? And now even on the web, you give a prompt, the agentic loop will come up with a solution, see if it compiles, and if it doesn't compile it will fix it. It will keep churning until that's happened. And then similarly, it'll try and pass all the tests. So that's been a huge change just in the last year.
And now, if you're operating in stealth mode, you've gone from running a sizable organization to produce Liquid to now working in stealth. How have tools like Claude changed your approach to software development?
Oh, the ability to move into a new environment very quickly and get something that is on the beaten path is just huge. I think, having a lot of experience, I've done pretty much everything once, so I mostly know how things ought to work. But there's a huge amount of complexity in our environment which has nothing to do with that. If you've never built a React application, there's a huge get-up-to-speed time while you have to learn all of its foibles and idiosyncrasies. And Claude completely short-circuits that. It's as if you had a quick-start tutorial written personally for you. That's fantastic. Now, other places, that's where correctness and what is desired has a reasonable definition. Other places, for example: is the functional decomposition sensible? There's no good statement about what is "done," what's best. And so typically you find Claude can't really make progress, right?
Suss that out for me a little bit more. What do you mean by that?
So, for example, in the example I gave, the 4,000-line program that just worked in a day, the entire program was written basically in the parser for the command line, right? So there's no modularity at all. And many of the operations where you would think, oh, this operation should call that operation, they should be the same by definition, it just had duplicate code. And so a basic transformation that had to be suggested to it is: hey, maybe only do the argument parsing in the main routine, and then have commands that handle the I/O that you'd use in a CLI, and then a bottom layer that actually does the heavy lifting. So that's not a thing that I think it is ever going to do easily.
So it's also the difference between architecting and doing actual development with these tools, and vibe coding something, it feels like.
I think the act of coding is always kind of vibe coding. It's: when do you stop? I'm not sure that people code any differently than Claude does. You think of something that plausibly could work, and you code it up and you experiment with it. And I think that's what Claude does.
Well, that makes a lot of sense. It's trained on people, so it's going to act like people.
Right, right.
Okay, so now you're operating in stealth, you're building product. What is the most frustrating part of this for you? Because I often hear from people, it was great yesterday, it's not good today, and I'm not sure why.
Well, there's a dependency on a thing with opaque economics that, of course, is a little bit nerve-making. But if you're forcing it to get the code out to human level... basically, when you can't read and understand the code, then you're just kind of hoping that it's going to do the right thing, right? So if you're forcing things out to human level of comprehension, hopefully that's not too big of a dependency, I don't know. I think the bigger problem is that the market is just crazy, right? You have huge expectations of, well, you should be able to do this instantly with AI, because I had a long discussion with AI and it thinks this thing is plausible. So people have strange expectations for what can and can't be done, and how fast it can be done.
Is that the simple problem of humans, that people are dumb, they don't necessarily know what something can do? Or is it a fact that this changes every two weeks, and something that was impossible last week, yeah, you can do that now?
That's interesting. I'm not sure how much it's changing, actually. For example, if the model was really getting better exponentially, why bother with a harness? Next year's model just does everything the harness would do automatically.
But that feels like it might actually be the question. Why did Mythos get pulled? Why has the current GPT release been restricted by the government to just a core group of people? Is that politics, or is that because of the models?
Or is it marketing overreach, like this? Please stop me before I sin again. No, who can figure that out?
I don't know, that's why I'm asking you guys. Ultimately, these new tools are allowing individuals, especially on greenfield projects, to produce at unprecedented rates and push incredible stuff to production, often. At the same time, people are getting laid off, insane amounts of money are getting poured into it, and the ROI of it is very unclear from a pure economic basis. But you're in the Valley, you've got to do what you've got to do here.
I don't see any indication that layoffs are really because of AI. I think layoffs are because we've had 30 years of malinvestment. We've had 30 years of mass-market software service that has been phenomenally profitable. So if you have a business in that field, you're generating cash at an insane rate. So the problem of Google, or the problem of Facebook or Amazon, is: how do you invest billions and billions of dollars sensibly? And the answer is, mostly you don't, we just waste it. And we build things that are insanely complex. So when people say, oh, AI, one of the exciting, fun things that people like me talk about is, hey, programming's fun again. What does that mean? It means I can actually do stuff with a program and not drown in the complexity of deploying this thing in a cloud environment.
So that's a really important and fascinating point that hadn't occurred to me. I came from the media world and the journalism world, and there's always the talk about social media destroying the business model around those things. Well, actually, SaaS and social made us all stupid. These were money-printing machines that were extractive to the greater culture and the greater ecosystem of innovation, arguably. And AI is actually the natural result of that. You know what actually destroyed journalism? It wasn't social. It was the fact that every website you went to on this free internet was suddenly loaded up with the most virulent ad tech and tracking you could possibly imagine, with a giant video screen down here. Of course no one wanted to use it.
Or, just maybe, you get what you pay for. Yes, it was free, exactly. People talk about doom-scrolling. There's a great way to fix it: don't read anything you didn't pay for. I read The Economist, I pay for it, and it gives me high-quality stuff that's interesting.
It's ruined healthcare too. You don't pay for your healthcare, you pay for the insurance that pays for your healthcare, which has caused everything to balloon up out of control.
Oh yeah, absolutely.
If you disintermediate the customer from the product, this is what happened. The middleman will extract and inflate and bloat. And AI, although frankly it's the most extractive piece of technology that's probably ever been created, can it at the same time disintermediate us from the source?
Well, we'll see. There are a bunch of creative, capable people who are feeling energized momentarily, and that's pretty exciting. A surprising thing about 40 years in software is how little is new. You know, x86 assembly language, learned 40 years ago, still a valuable skill. Operating systems, databases, programming languages, all still pretty much as they were 40 years ago. In fact, I would say the only surprising thing to a well-educated programmer from, say, 1982 is this notion of distributed version control. So things like Git, that you can collaborate with 10,000 people on something as complex as an operating system using Merkle trees. That's crazy, nobody thought of that. But almost all the bread-and-butter programming that we do is completely unsurprising.
Are you more excited or more anxious about where things are going in your space?
A little bit of both, I think. On the one hand, there's an opportunity while you're able to use AI to automate your way through complexity. There's real opportunity. If you have some idea that you'd like to execute on, there's a window of opportunity to do that. On the other hand, all kinds of crazy stuff, right?
So I guess that's the thing. Is the opportunity for the entrepreneur, or is the opportunity for the rising tide raising all ships? There are different periods in our history as a society where not both of those things are true at the same time. Traditionally, when it is the opportunity for the solo entrepreneur, the idea-getter, eventually it becomes that uplift, but it doesn't happen all at once. And that's how you end up with a railroad age before an industrial age.
I think the rising-tide effect is going to come when people start using computers to do computation, right? Which we've kind of forgotten about.
What do you mean?
Well, in modern software, you get an app, the app does a thing. If that's what you want, that's fantastic. If that's not what you want, well, you're done, there's really nothing you can do. So I would say the whole point of having a user interface is to teach you about the underlying structure of computer science that you're working on, to teach you as much as you want. If you don't want to learn, fine, we can automate it. But someday, maybe you want to do something, and your ability to learn what this application has actually been doing for you is extremely precious, right? Maybe a way to say this is: sometime in the future, when you learn how some application that you've been paying for works, you shouldn't feel ripped off. You'd be like, oh yeah, okay, I guess if you build a machine and it does this thing, then that's a reasonable product and I'm happy to have paid for that.
But it feels like there are people that passively exist and consume the world, and there are those who want to understand it.
Yeah, and those are going to sort themselves out in the market, right? You can't make a horse drink, and all that. I think the question is... and people convert from one camp to the other somewhat surprisingly, right? A lot of people are busy being a doctor or a lawyer or whatever it is they're doing. They don't want to learn how to use a computer. I mean, they would love it in the abstract, just like everybody would love to be a concert violinist. When you see one, or hear one, it's like, wow, that is really cool, I'd love to be able to do that. I don't have 20 years of my life to practice, right?
Well, it's also, when you're using a tool, are you using it to make your life easier? The tools exist to make life easier. Or are you using it to distract yourself from boredom? A lot of the consumption that happens, and a lot of the tools and technology that's been developed, yes, it makes things easier, but often what it's easing is boredom.
Yeah, I think I'd be careful of "kids these days" sorts of arguments, I don't trust them. So for example, Instagram: lots and lots of junior high school and high school age children are learning to be fluent in the language of video and film in a way that never happened in my generation, right? Where you had to have thousands of dollars worth of equipment, and a teacher. So that's pretty significant. Now, yeah, maybe they spend a lot of time distracting themselves, but that's a choice that everybody can make.
I spent my career as an editor for a decade, at a very high level, and some of the stuff I'm seeing on TikTok and Instagram, just from a pure editing perspective, blows me away, what they're able to accomplish. New mediums being developed is fascinating.
Yeah, and so it's addictive and cynical and advertising and all that, and people are learning to express themselves. So I think that's pretty hopeful. Maybe a hopeful point to go out on is: if you look at chess, which was one of the first games that AI conquered, could play better than people, the result is that now, if you're interested in chess, any child anywhere has an infinitely patient grandmaster chess teacher. And you can have as many problems as you want. White to mate in two: oh, you try that move, well, that's actually the second-best move you could have made, here's the first-best one because of this, and then there's the one you chose. So you learn instantly, right? And chess players have gotten fantastically better. So imagine you have that for lots of other domains, like coding, maybe we can hope for that, writing, video editing. People can get a lot better with this sort of instant feedback and instruction and practice.
It either makes us better or it makes us dumber. That's kind of the demarcation point that we have.
And the question is, of the people who are nominally dumber, and I think you have to be kind of careful there, were they ever going to do something different? I mean, maybe they're just too comfortable. Maybe the best you can do is be there for them when they don't want to do that. Some people are couch potatoes for 30 years and then decide to go run a marathon. But never want to rule that out.
Thank you so much.
Be part of the
conversation.
Whether you're a CTO who wants to be featured, a company looking to sponsor, or an engineering leader wanting a seat in the room — there's a place for you here.