He Thinks AI Code May Break Everything - EP 59 Will Wilson

Imagine you're like a detective and you're at the scene of a car crash, right? Car has like veered off the road and crashed and everybody's dead and you're trying to figure out was foul play involved, right? The moment you walk into that crime scene, you're, you're like accidentally destroying evidence, right? Like you could be stepping on some like key trace of something. You're like contaminating with DNA. Like, you know, it's like actually very, very hard and delicate and you have to be really careful about everything you do. And you can't, like, rewind time and watch the crash from a different angle. You can't rewind time and go look and see if the guy's brake lines were cut. You can't rewind time and give him a field sobriety test, right? It's exactly the same way with debugging software. Most of the time we get to the software after the car has already crashed and after the machine has fallen over, the server has exited its memory, its core memory has, like, been dumped, you know, dumped somewhere.

Well played.

Like, you know, it's like It's like the aftermath of a crime scene. And not only is most of the evidence gone, but anything you do might destroy evidence further. Whereas when you are debugging in a simulation, because it's a simulation and because it's deterministic, right? Fully replayable. You can just rewind time.

Welcome to the Core Memory Podcast. This is Ashlee Vance. We are joined today by Will Wilson. Awesome name.

Easy to remember.

Uh, you were getting, obviously get into everything you've done and, and our journey, but you're the CEO and co-founder of, of Antithesis. That's right. Does it have a second part to it? No, it's just, no, no, it's just Antithesis.

Just Antithesis.

Yeah. Yeah. You guys were operating in super secret stealth mode for several years and have emerged?

Yeah, well, so this is actually kind of a fun story in itself. We were in stealth mode for 5 and a half years.

I saw that. It was lengthy stealth. It is.

And that, that requires a sort of an extreme set of people and a slightly extreme, you know, approach to building a company, which is maybe something we could get into if you want. But we could also just start with what it does.

I did kind of, because that jumped out at me. I mean, yeah, just, just to give people context, I mean, the big picture, we're going to talk about a lot of stuff, but clearly we're in the this moment where how people create software is changing dramatically with AI. I've been fascinated by software forever. I was an open source dork for many, many years. But yeah, so we're in this moment. I mean, you know, I think people who are immersed in the San Francisco AI scene know this all too well. I wanted to try to give people who are not, you know, professional software developers or even maybe just like super dialed in to what's been going on with AI and software, a picture of of where we're at, because I think everybody sees these stories about companies are using this to write all this code and I can never actually get a good sense as to like how much of this is actually going on and what's good and what's bad and where these tools are at. And so I wanted to level set people there. You're a perfect person to do this because Antithesis analyzes software and tries to make sure everything's working smoothly. The AI in fact told me that, that this is prescient having you. It could not be more prescient to have you on the podcast.

And then did it tell you you were absolutely right?

It did. It did. So it was telling me I'm picking good guests. But yeah, why don't you set up antithesis in your own words?

Okay, sure. So let's actually start not in the present moment. Let's go back like 75 years. So you use like software programs because everybody does because software is eating the world or it's already eaten the world. And a thing that you might notice about your software programs is that they don't always work super well. Like often, I mean, just literally yesterday, right? Like Verizon had this huge outage that left like most of the US unable to make phone calls. And why is that? Like, we're like very used to this idea for some reason that software just doesn't work properly all the time. You know, you log onto your bank and you can't log into your account and they have a little red error message that's like, hey, sorry, something didn't work. And we just are, we just roll with that. But like, that's actually very strange if you think about it, because we don't just expect buildings to like randomly fall on us.

Right. I mean, it's this essential part of our infrastructure now. Yeah.

And we're just like kind of okay with the fact that it's all really crappy. Like, why is that? So I ha— this is like a thing I've been thinking about for a long time. And I have a theory. Well, there's, I've got a couple theories. One is that software is just actually really, really complicated and really hard to make. These like software products are super complex. If they were machines, they would be like the biggest machines on earth, right? But another thing that I think is really true is that we haven't been writing software for very long, right? Like we've been doing it for like 70 or 80 years depending how you count. And if you go back to like 70 or 80 years after like the wheel was invented or the steam engine was invented, like people sucked at making those things.

Yeah. Yeah. Well, I mean, and on your point about the bugs, I don't know. I think that's my general impression is even though We live in a very tech-filled world. On some level it still feels new-ish and you just kind of like expect the kinks are being worked out, you know? Yeah, I mean, we cut, we cut websites and software services a lot of slack in general.

That's right. And so I think what that means is that we should be open to the idea that there are radically, transformatively better ways of building these things that we just haven't figured out yet.

Okay.

And that's kind of, you know, we basically think that testing in particular and making sure software really, really works right is one of these things where we just, most of the world has not even begun to think about how to do this correctly. And so the way that basically everybody's been testing software for the last 75 or 80 years is basically you follow the following pattern. I'm writing my code and then I try to think of all the ways something could go wrong with my code. Hey, maybe the user's gonna put in a funny name, or maybe the network connection will go down at just right this moment, or maybe whatever. Like we think of all these scenarios and then we write like kind of special purpose example-based tests for each of those situations to like check what happens. And then we try and like, you know, make it actually work in that situation. Right. And that's the basic approach that we've been doing all these decades. And it just like totally doesn't work. Right.

Yeah. I mean, is this— I don't want to set us off on a different course, but I mean, my general assumption would have been that the tools to check all that kind of stuff would have gotten quite good over the years. I remember doing a story on Facebook where I went inside with their team and they were pushing the limits of updating such a massive website and infrastructure every day, right, with multiple changes. And, you know, they had all these sophisticated tools in place to make sure they weren't just wrecking Facebook every time they they pumped out these new changes. And so I guess in my head, companies of like that scale, large software makers, I would've assumed, you know, Microsoft had built incredible— although Microsoft's maybe not the worst, not the best example.

But you might actually be surprised. Okay. I have been inside these companies and seen how the sausage gets made and it's actually still very primitive.

Why? Okay. I saw that 'cause you, you had a previous company called FoundationDB, which Apple acquired, and then you guys actually were working at Apple for a while.

I was at Google for a while too. Yeah. No, I think what these companies have mostly invested in, and it's not crazy given how bad we are at testing software and making sure it works, what they've invested a ton in is what's called observability, which you can think of as like having a fire alarm as opposed to not making your building out of straw.

Okay.

Right. It's like, it's like basically being able to notice something has gone wrong and react really fast and then roll it back. And they've gotten don't get me wrong, they've gotten really good at that, but they are still really bad actually at validating changes before they go out, which I kind of feel like if we could do that, it would just eliminate huge categories of problems.

And like, what happened? I mean, the industry just developed so fast that you were almost having trouble keeping up with the task at hand. So it was just like, we're using the tools we have, or you, you guys, I mean, I know you're a mathematician by background, you, Did you just think about this in a totally different way or?

Right. So, so basically, you know, basically everybody's been doing this thing like example-based testing forever and ever. And if you, if you think about why it shouldn't work, the reason it shouldn't work, well, number one, it's a lot of work, but the reason it fundamentally doesn't work is that most of the situations that break your software or break your hardware or whatever you're building are not situations you can think of ahead of time. Right? They're like unknown unknowns. They're crazy things your users did that you never would have dreamt up, or they're crazy things that happen in a data center somewhere, or they're like a crazy thing some sysadmin did or a crazy thing some hard drive did that you just couldn't foresee, right? Like that's true in most parts of life, actually. Usually, usually the risks that end up destroying you are not the things you carefully calculated. There's a 2% prob— it's like, no, it's like some crazy thing out of left field that you didn't even see coming.

Yeah.

And it's exactly the same way in software development too. And so there's almost like this catch-22 where if you could like think of the problem ahead of time, it's probably not the one that's gonna get you because if you thought of it ahead of time, you probably wrote your software to handle that situation. And so the question then is like, how do you test software if the things you need to test are the things you would never think of testing? That's like the fundamental paradox.

Yeah.

So there's this whole other way of thinking about this, which was like existing in this tiny niche, mainly with academics, wasn't very practical. And I'm going to call that whole world like property-based testing or autonomous testing.

Okay.

And the idea here, and which to be clear, like 0% of the world still does approximately.

0% use these techniques.

I mean, like, you know, rounding error, right? Tiny number of people. The idea there is, hey, we're not gonna do this. We're not gonna try and think ahead of time of everything that could happen. We're just gonna take your software and we're gonna put it in like a simulation of reality and we're gonna just throw stuff at it. We're gonna randomize what happens. We're going to send random inputs, random API requests. The user's gonna be clicking random places, put in random names, and then maybe we could also like randomize the order that packets get delivered by the network and randomize how long they take to get there and randomize whether the hard drive actually stored the data you told to store or whatever. And by doing all this, you know, we're going to really quickly find all these situations that you wouldn't have thought of. So that sounds like a great idea, right?

Yeah. Yeah. I mean, you know, when I was reading about you guys, I mean, you do do techniques similar to this, right? I was sort of like, I mean, I assumed that that's how this actually was done. I mean, why has this been frowned upon? Right. It's just too hard.

So the thing is, it was not practical.

Okay.

For a very, very long time. And there's basically two different reasons it's not practical. And we figured out at our last startup, FoundationDB, which you mentioned, how to fight both of those problems that make it impractical for most people.

Just for developing the products you guys were doing.

Yeah. And then at Antithesis, we have like totally crushed this problem and made it easy for everybody. So basically the two things that are problematic And they're both, like, really fundamental deep facts about the universe, right? The first one is the number of possible combinations of random stuff that could happen is just so huge.

Yeah.

It's, like, infinitely huge. And so if you just, like, are trying stuff randomly, you're probably not gonna try the stuff that really is important that will actually cause something to break in real life. And that's a very, very fundamentally hard problem., you know, you can get like teams of scientists to like look at it and be like, are we covering all the, you know, and it's like most people don't have time for that. Most people can't really think about that. It's like a little unreasonable to ask people to do that. Right. Um, so that's, that's one set of challenges. The other set of challenges is actually kind of even more fun. Um, and it's that computer programs, the vast majority of computer programs are non-deterministic. Meaning if I run the computer program twice and tell it to do the exact same thing both times, it'll actually do slightly different things.

Okay.

Because the hardware has, like, slightly different timings. Like, a CPU is a tiny machine. Like, if I increase the temperature in the room, that machine will behave differently. So it's like a physical thing.

It's just firing at different speeds.

And yeah, and then if I have two computers and they're sending packets to each other, like that's a physical system. Like, you know, and there's a router in the middle and like what else that router is doing could totally affect how long it takes a packet to get from machine A to machine B. Well, why does that matter? Well, because maybe machine B is in the middle of doing something else now when that packet arrives. And as a result, your software just does something totally different.

Okay. Okay. So you can, you just, you just get a problem, but you can't even really replicate what happened and fix it.

Exactly. And so you have, that, that causes problems. I mean, every developer is already fighting with this. It's called like flaky tests or non-determinism, right? This idea that like a user could report a problem, for instance, and you could literally maybe never make that problem happen on your machine. But it's an especially big problem for this, this, you know, this autonomous testing approach because it, um, you know, we're just like constantly finding really, really tricky, weird, subtle edge cases. And you want to be able to just like make it happen again and again and again and again until you can fix it.

Okay.

So at FoundationDB, we, you know, basically before we even built the database, we decided to try to do something totally crazy, which was to build a fully deterministic simulation of a database cluster where we could take all of the different database processes and rather than running them all on like real computers, run them on like fake computers inside of this simulation that we had complete control over and complete control of all the timings and everything. And so if we decided to make a computer like crash at some particular instant in the simulation, we could always reproduce that, always replicate that. And if that turned up some kind of crash, we could just make it happen over and over and over again.

This episode is brought to you by Brex at Core Memory. We spend tons of time with companies that are trying to build fast. And they often run into similar problems. Their corporate spend either spins out of control or slows them down. Brex was built to break that trade-off. It is the intelligent finance platform that helps you move faster and spend smarter. Cards, banking, expenses, all in one place with AI agents working in the background to handle annoying tasks like collecting receipts and enforcing corporate policies. This is a massive time saver for us. More than 35,000 companies use Brex, from early stage startups to giants like Anthropic and Zoom, and soon-to-be giants like CoreMemory. We run on Brex, so should you. Go to brex.com/corememory to learn more. So you're— and you're replicating, like, say I'm, I'm a giant, um, I don't know, I'm a car company and I've got huge computing infrastructure. You're, you're, you're doing this replication of their entire data center system, their computing infrastructure, right?

Well, so back at FoundationDB, we were only replicating— we were using this ourselves, your own stuff to like, to like make our development process go faster. The thing that we're doing at Antithesis is taking that same approach and that same technology and yeah, making it so car companies, you know, financial companies, everybody on earth can use it easily.

Yeah. So I mean, I was trying to get my head around how that even works. I mean, I saw on your website you guys do tick off like Jane Street and crypto companies and then cloud company. I mean, all the sort of big usual suspects. But if I'm walking into a financial services company and you have to replicate their IT infrastructure in this virtual way, what, you're just like sending out a bunch of agents to like go grab this information on what it looks like and then, and then try to build it back up?

Right. So this is something that that when we started the company kind of a long time ago was looking a little bit dangerous and like a big risk, but it's actually just kind of turned out to be really easy because basically the whole world has already moved to like these like standardized cloud deployment things.

My other question.

Yeah, yeah, right. Everybody's using Kubernetes or they're using Docker Compose or using Terraform, whatever, you know, kind of everybody builds against AWS, you know, asterisk. There are a few others. Yeah. Like, so basically if you can just develop a way to like slurp up, you know, those infrastructure as code description languages and stand them up inside of your system, your customer doesn't have to do very much at all.

Yeah.

And you can just like, you know, stand up a simulation of what they're doing very quickly and very easily.

And so why would the, why would the simulation give you these results that can be replicated versus the real world where, where it can't?

Right. So basically in the real world, We're running on real computers and real computers, like I said, they are physical machines. They are, they obey the laws of physics. Like they have to, you know, they have real electrical voltages and real clocks. And, you know, in our simulation, we've basically written what's called a hypervisor, which is like a thing that runs virtual machines, like, you know, VMware or Amazon's got one.

Literally just walked by Pat Gelsinger, the old CEO of VMware.

Yeah, right. Exactly. So, so I, we have a hypervisor too. Except ours is cooler than anybody else's.

Okay.

Because it's fully deterministic, meaning you can run it on a real computer and it is able to simulate a perfectly deterministic computer, which kind of nobody thought was possible before we did it.

Okay. And you have to like totally just pat yourselves on the back. I mean, what, you guys are just— you're just super, super clever dudes who figured this out.

I mean, I don't know. Like, I think nobody Look, I think insofar as we had a key insight, it was not that this was possible. It was that this would be unbelievably valuable.

Okay.

Right. Like, I think probably if you had polled software engineers, probably like 80% of them would be like, that's impossible. And 20% would be like, yeah, you could do that, but why would you? And I think our true insanity was not that we're in the 20%, but that we were like, no, no, if you could do this, it would be worth a lot of money.

Okay.

You know, because you could finally crack this problem.

Just prefacing this with, I know very, you know, I only know software to the point that I could write about it. I've never actually coded anything, but if you are, I, I get the part where you're creating this virtual environment where you can replicate all this stuff and, and now you've taken the nuances and the, the weirdness of the hardware out of it. But then when you come back to that real world, I mean, aren't you just reintroducing all those new, all the problems with the hardware all over again?

Right. So the key is, if you can find all your bugs in the simulation, which at FoundationDB, to be clear, we did. And at Antithesis, we find most of our bugs in the simulation too, actually. Um, and a lot of our customers find pretty much all their bugs in the simulation. Then what you can now do is you can also debug them in the simulation. And the nice thing about that is you're, it's like you're debugging with a time machine.

Okay.

Like, let me give you, let me give you an analogy since this has been a little bit technical so far. Like, imagine— okay, imagine you're like a detective and you're at the scene of a car crash, right? Car has like veered off the road and crashed and everybody's dead, and you're trying to figure out was foul play involved, right? The moment you walk into that crime scene, you're, you're like accidentally destroying evidence, right? Like you could be stepping on some like key trace of something. You're like contaminating with DNA. Like, you know, it's like actually very, very hard and delicate and you have to be really careful about everything you do and you can't like rewind time and watch the crash from a different angle. You can't rewind time and go look and see if the guy's brake lines were cut. Can't rewind time and give him a field sobriety test. Right. It's exactly the same way with debugging software. Most of the time we get to the software after the car has already crashed and after the machine has fallen over, the server has exited its memory, its core memory has like been dumped, you know, dumped somewhere.

Well played.

Like, you know, it's like it's like the aftermath of a crime scene. And not only is most of the evidence gone, but anything you do might destroy evidence further. Whereas when you are debugging in a simulation, because it's a simulation and because it's deterministic, right? Fully replayable. You can just rewind time. You can be like, all right, take me 5 seconds before the car crash. Uh, okay. No, take me 10 seconds before. Okay. No, no, no. I want to be standing on the hill and watching it.

No.

Okay. Now I want to be inside the car and watching it.

Okay.

Right. And like, it's just this, drastically higher productivity environment for like very quickly answering that question of what happened.

When you guys— Apple buy you guys and you bring these tools, I mean, you know, they're buying essentially your database technology on one level, but you get inside the company, you have these tools. I presume you like showed this to Apple, you know, that it could be used. And I think I remember from my reading, I mean, they did not adopt this inside of Apple, right?

Well, so Apple still uses that database very, uh, you know, in a very significant way. Yeah. Um, that's been publicly disclosed by them. I'm not violating my NDA to say that. Um, but you know, the way that we had built the simulation technology for that database was really specialized for that database. So they couldn't just take that and apply it to every other system at Apple. But with Antithesis, they could actually take it and apply it to every single system at Apple. And, you know, I think they should.

And then, okay. And then, you know, when you guys spent all these years in stealth building this thing, I mean, obviously, I guess it's just— it's an incredibly hard problem. I mean, is that why you said it was— it was 4 or 5 years?

5 and a half.

5 and a half years.

Yeah.

You're in Virginia.

Yeah.

Is the whole company in Virginia?

At the time it was. These days we've got— we've just gotten a few people in SF and we actually have a pretty big office in London.

Okay. Are you from— why is it in Virginia?

So that's a kind of funny story. Um, FoundationDB was, was in Virginia and, and, and that went great in a lot of ways. And, you know, that got acquired by Apple and, and Apple actually insisted that we all move out here to the Bay Area. And so we all moved out to the Bay Area and, you know, the Bay Area is amazing in a lot of ways. I think the like superpower of the Bay Area is the extent to which ideas cross-pollinate and to which people just like hop from place to place. And what that means is if somebody has a really good idea, it spreads really fast. And like, there's no non-competes. Like, this is like, this is all good stuff for the ecosystem. But there's a flip side to that. And there's, I think everything comes to the trade-off in life. I think the flip side to that is it's actually really hard to do something different in the Bay Area. Like, I think there's this weird mimesis where everybody's kind of trying to do the same thing and everybody kind of runs their companies the same thing. And we were just like such an outlier. We were so weird, right? We had this like vision for a completely different way that people should be writing code and debugging code. And I just, we just didn't think that it was going to work if we did it in the Bay Area because, because it's like, it's, it's very hard to keep people for the long haul. It's very hard to keep them motivated and focused and quiet, like doing something. And so we kind of had to like sail off into Siberia and then like burn the ships, which is what we did. And, and so, you know, we went back to Virginia where, by the way, you can hire incredibly great engineers. Because they all work at the government or they work at federal agencies or for contractors or whatever. These are very deep people, deep tech, you know, working on super secret stuff. They're very good. But then the difference with SF is you can keep them and you can build a distinctive culture and a culture that's like different from just the average of every tech company.

Yeah, but you got to live in Virginia.

Yeah, you know, the weather's not as good. That's right. But the real estate's a lot cheaper.

It is. It is. I got to support California now. Everyone's leaving. I got it.

I love California. But look, I actually think it is, you know, you wrote a book about Elon.

Yeah.

And I think that say what you like about Elon, he has achieved incredible results with his companies and he runs his companies in a very different way than most people in SF.

Right.

And I don't think it's a coincidence that, you know, SpaceX was started in L.A. and not S.F. Like, it's not an S.F. company. It's got a very different ethos and attitude and a different kind of esprit de corps. I don't think he could have done that in S.F.

Yeah. Where did you grow up?

I grew up in Hong Kong.

Oh, okay.

Yeah.

For how long? I mean, like, your whole—

Yeah, yeah. I was, I was there until I was 14.

Okay. What'd your parents do?

They're just business, overseas business stuff.

Yeah. Yeah. Okay.

Nothing exciting.

Where I was trying to— I was trying to find where people go to school. Just when I'm doing my digging, I didn't find it. I didn't know if this—

are you?

Yeah. Yeah. What's going on?

No. So I— yeah, I went to a— I went to a small, you know, government-run school in Hong Kong called Island School.

Okay. No, no, no. But for, for college.

Oh, I was at Yale.

Yale. Okay.

Yeah.

Right.

And then my, my, my co-founder Dave is the, is the dropout.

Okay.

You know, I feel like I should have done that instead.

I think you might literally be the first person from Yale I've met that's done a startup. Really? Yeah. It's like, it's really weird. Is you obviously Harvard, you got, well, I mean, they're kind of dropouts, but you got dropouts galore that have started mega tech companies. MIT, fair share. Cornell drops off like really quick, man. It's, it's funny to me. I, I was always wondering what happens with the Yale computer science department.

Um, yeah, I wouldn't know cuz I wasn't a computer science major.

Math.

Uh, yeah. Math and philosophy. I, I think like, I think that the, you know, I think the Yale CS department is pretty, is pretty theoretical in its focus.

Yeah.

Although I think that's actually quite common for, you know, for research universities. I don't know what it is about it. Yeah, it's a good question.

Okay. All right. Well, I want to get back, let's get back to software for a second. I mean, I'm doing this research on you and you can see in my notes, it's like planes, ICBM silos, AI is gonna destroy civilization. You give these talks and you're very eloquent on this stuff. And then, you know, because you, the business itself is, it takes some explaining and, you know, it's just not like right in the mainstream. Stream, but I feel like you do a great job of explaining the importance of software in the world and then, and then some perilous situations that we, we may be heading toward. I mean, okay, so let's— you've created this tool. It, you know, analyzes this very complex software. The world runs on software. That's not going to stop anytime soon. And now I feel like you're better at describing this moment than I am. But I mean, I see it everywhere. Obviously, people are— they're vibe coding, they're, they're using AI. To write code. You made an analogy in one of your talks, you know, just like your kids are probably tempted now and again to cut and paste some stuff for their term papers from AI, that naturally coders are doing the same thing. I guess for me, again, I don't code, but like in my head it seems like coding has been one of the things that AI got good at first. It makes tons of sense to me that you can have these systems look at billions upon billions of lines of code and I would think they'd be pretty be good at it because, because they have all these examples. And, and so when you were breaking some of this down, I was like, well, if we're cutting and pasting, you know, how bad actually is that compared to some— what a human is doing anyway? They're probably making mistakes too. So, um, you know far more about this than I do. I mean, give me the lay of the land of— it is clear in this moment, right, that AI coding is just, is just totally mainstream and, and taking off.

Yeah. Yeah. I think pretty much everybody does it. We actually, funny story, our company does relatively little of it. Um, in part that's because the stuff that we're working on is so hard and so different that there's not a lot of it in the AI's training data.

Okay.

And so we really do need human beings who can, who can think. Um, I think, so I think AI coding has really, really made clear something that was a little bit implicit before, which is that software engineering is not really one job. It's actually a lot of different jobs, and there's a lot of different kinds of software that people write and a lot of different tasks even within writing software. And I think right now AI has this like very sort of spiky capabilities frontier where it is really good, perhaps even superhuman at some of those jobs, and really bad and dumb at other ones of those jobs. And so like, if you ask AI, hey, make me a React app, like, you know, make me a little website that lets me, you know, whatever, lets me make bets about who's gonna win the playoffs or something.

Sure.

It'll, I mean, it'll just like, destroy that, right? It'll, it'll spit it out really fast, really cheaply. You know, it will save you a vast amount of time and money. It's great. If you then want to yourself make some change to that application that it has produced, it's actually now going to take you quite a long time to go and read through everything that AI did and understand what it all did and, and make those changes. You know, it's like I've compared it before to like trying to edit somebody else's writing as opposed to editing your own writing. Like, if you write something yourself, it is in your head. It's like part of your heart. Like, you know it, you know it like the back of your hand, right?

No, as you were saying all of that, I was like, yes, this— I see this, this makes, makes perfect sense. Yeah, yeah, right.

And so, and so that's one kind of interesting, one kind of interesting thing. And then if you ask the AI to modify what it has done the results are a little uneven. Basically, basically AI does okay at codebases that are pretty small and that fit within its context window or codebases that are like really well structured and well documented, um, where there's a lot of like reusable abstractions where, you know, where somebody has like architected it thoughtfully with the idea that somebody's going to come make changes to it later. The thing that's kind of paradoxical is AI is actually pretty bad at like keeping it that way. It tends to make a mess. Like it tends to, it tends to add stuff in a way that doesn't really look ahead to the future. Um, you know, the way that, that somebody I know at Anthropic put it is that kind of has poor taste, which is like this like kind of indefinable quality, but it's like very, very true. And so if you are using AI on a project that's not just a throwaway project, that's something that you can you know, you intend to keep working on forever, um, or making better or selling to somebody or maintaining in any way, you really need to stay on top of it. Um, and to the extent you need to stay on top of it, I'm not actually sure how much time it saves people.

Okay.

Like, it, it, it definitely— if you get really good at this, it definitely does save you some time. It definitely makes people more productive. But people who say it makes them like 10x more productive, I think, are exaggerating. Or they're working on a very different kind of problem where it's okay to just kind of one-shot this and then like put it out the door.

Hello, geniuses. Let me tell you about E1 Ventures. They are a venture capital firm in Silicon Valley, a longtime supporter of this podcast and a longtime supporter of big, fantastic, world-changing ideas. If you have such an idea, hit up E1 Ventures or send me a note and I'll put you in touch with E1 Ventures. It's, it's quite a, quite a deal for listening this podcast. Um, thank you again, as always, to U1 Ventures for their support. All right, I have so many questions now. Yeah. Uh, I mean, you— because this just—

sorry, by the way, I know this is like heresy in SF, I'm not supposed to say this or whatever, but yeah, right, we like heresy. Uh, you—

it just raises so many questions. I— okay, so one line of thinking though that you were saying, you know, is, is that people are cutting in pay, they're, they're going crazy with AI code. We're cutting and pasting stuff. We're kind of undermining ourselves, these, these skills that, it's a craft, right? And, and, and people spent decades working on this stuff, just like a lawyer is, is built up on hundreds of years of, of precedence and, and you're, you're a legal scholar and you get a little too dependent on the ease of these tools. And, and over time, because I want to break down your different arguments. So one argument that you, I've seen you be making is that if we get too dependent on this stuff, you're kind of like undermining undermining these base skills that we have. Is that a fair way to put it?

Right. Well, here's one way to put it. Right now, I think that a very, very skilled developer can use AI coding tools in a thoughtful and pragmatic way and can get really great results. And I think like a not super skilled developer often gets pretty horrible results and like creates a mountain of technical debt for themselves. And there was actually a study that just literally came out yesterday about exactly this. Like they looked at when, open source projects adopted Cursor. And, you know, basically when they adopt Cursor, like their productivity goes way up. And then a couple months later, their productivity goes way down.

So you're just kind of creating this mess that nobody really knows what's going on.

And that, that, that's right. But so, but so let's focus on the really skilled developers who actually can use these tools and get really great results from them. How do you become a super skilled developer? You become a super skilled developer by writing code yourself. Like all the time and thinking about it.

It's a craft.

It's a craft.

Yeah.

It's like you can't read a book, you just have to like do it over and over and over again. It's like anything else. And so, you know, how do you do that when everybody's using AI from the moment that they enter the field? Are we just gonna not have super skilled developers in a couple years? I don't know the answer.

Okay. Okay. This leaves me with two questions for that. I mean, sticking on this for one second, I mean, one of the, the scenarios that you lay out is, is we become very dependent on this stuff. You're like, I don't, I don't wanna be flying on a plane in 10 years. There's all this code. Nobody's gonna know what the hell's going on. These, these like systems that we take for granted actually do get better all the time out in the real world will start, will start decaying. I mean, were you doing that to make a point or you really believe that that's gonna happen?

So I really, really don't know the answer to this question. I think there's, I think it depends a lot on the future trajectory of AI. Basically, I think that if AI plateaus at a level that is below human abilities, there is a conceivable sort of terrifying idiocracy world where it just like, where humans become really lazy and dumb and sort of cognitively unfit because they use so much artificial assistance. They never develop these skills. But on the other hand, AI has not gotten smarter, and so it's still pretty dumb. And so we're just all dumb together and civilization falls apart, right? Like, that is, that is one conceivable way that AI could kind of destroy the world, even if it doesn't get very smart. I don't know how risky that is, right? Like, I do feel like eventually civilization would, would evolve better institutions or better cultural norms or whatever for using this stuff. But the, the point I was trying to make is that that evolutionary process is often pretty slow, right? Like, it takes a while actually for civilization to find ways of like digesting and integrating fundamentally new technologies and ways of doing stuff and making them helpful rather than harmful.

So you think we might have this, whatever it is, 5, 10-year period where, say, AI stuff's come out, we just go crazy whole hog on it. Your company's telling you to use it. They're firing your buddy because he's not using it and trying to save money. And then we wake up up one day and like nobody knows the planes, the planes not flying and, and so nobody knows who wrote the code. And we got—

Right. I do think that's very plausible. And that is one reason I'm doing what I'm doing.

Yeah.

Is that, you know, I do think that very powerful autonomous testing or other forms of software verification could help to reduce the odds of this apocalypse. Right. I like, I don't think it's necessarily— You're being more—

when I see you give these speeches, you were, you're fired up, man. I mean, I mean, I, the, I mean, you like— it seemed like you do kind of think this is going to happen.

Well, I, I don't have a crystal ball. I think— look, let me put it this way. I think not enough people are thinking through all of the possible ways this could go.

Doesn't it just get better though, over time? I mean, if code was like, again, the first thing that AI seemed like it got good at, seems like it's just getting better all the time. I don't use I don't know if there's a lot of this, but I see people just raving about everything in Claude and, and, and I guess in my head I see what you're saying on the craft thing because, you know, I will confess like every now and then I will use AI to write like a YouTube description. I would never do it on like a story or something like that. And then in the back of my mind I'm like, oh, this is gross, man. I'm gonna lose, I'm gonna lose my thing that I've worked so hard at refining. And, and you know, you can, it's just like a muscle and And so I get that. But then AI just seems— code just seems to lend itself to these LLMs so well that I would, I guess I would assume, and I'm assuming, so that's dangerous, but I would assume it just gets better and better and better and does what all these people are convincing me of all the time. I mean, it starts to reach superhuman levels where it's, you know, it's improving itself and it's going beyond what we can do. And then you don't have a human who makes mistakes. Putting bugs in the code in the first place.

Yeah, look, I, I would love that world. I, to be clear, I, I'm not a hater here, man. I like, I would, I would, that would be great. Like many of my best friends are futurists. I like, sounds, sounds awesome. Sign me up. Um, I'm a little bit skeptical.

Okay.

For a few reasons. Um, and again, Many people have looked very foolish making predictions about AI.

Yeah, it's not going to stop anytime soon.

That's right. So, so I want to give you that proviso. Um, basically the way that AI has been trained to write code, there's sort of two steps. One is what's called pre-training, which is where the AI is fed the entire contents of the internet and, you know, kind of gets good at producing kind of the average of the internet. And there's a lot of code on the internet. And most of it kind of works. So great, right? It can now write code that kind of works. That was not enough to get AI to the point where it was economically useful for coding. The part that made it economically useful for coding was what's called post-training. And in post-training, basically what you do, it's reinforcement learning. Learning. So the early versions of this were reinforcement learning with human feedback, where you have the AI generate some code, and then some human being looks at it and is like, yes, that's good, or no, that sucks. And then you do that a lot, and eventually it gets better. There's a problem with that, which is, are you training the AI to write correct code? No. You're training it to write code with bugs that the human can't bot, which is basically the most dangerous kind of code imaginable, right?

You're feeding it bad, bad info.

Yeah. So then the next generation of this, which is better to be clear, is, uh, is something called reinforcement learning with verifiable rewards. And the idea here is we're going to take this messy human factor out of the picture. We're just going to take the code that the AI writes and we're just going to run some tests against it and make sure it actually works. We're going to compile Does it compile? Great. It type checks. Okay. It passes these tests. Okay, great. It was good, right? But the whole first part of this conversation, I was just telling you that that whole approach does not work, has not worked for 75 years. Like it actually doesn't work at all. And if you look at a lot of the benchmarks that, you know, people report, there's, you know, how good their models are, et cetera. They basically are using this incredibly primitive, incredibly old-fashioned, I don't know, notion of how you know that software is correct, right? It's like it passed some unit tests. It's like I can write a lot of programs that pass some unit tests and contain subtle timing bugs or concurrency issues or, you know, will totally blow up when somebody provides an input that it wasn't expecting or whatever. So my, my kind of my theory actually is that the reason that Claude actually can write pretty decent code for a pretty long time, actually it's relying more on the pre-training than people think and more than they realize. And, you know, and that the RL that people are doing are not actually making this really a lot better at writing bug-proof code. And I think I think that, you know, as this stuff, like we're still very, very early, right? Like Opus 4.5, which is like the best coding model, literally came out at the end of last year. Like it takes time for code to get written. It takes time for it to ship and to make it into production. And then for like people to gain operational experience with it and like debug it and troubleshoot it and so on. Like I think that we are going to gradually realize that this is actually that this stuff is not flawless, that this stuff can write a lot of code really fast, and some of it has errors, and there are errors that are probably pretty hard for people to spot. And I think that it's gonna become clearer and clearer to people that we need way, way, way stronger forms of software verification. And that might not be us, right? It could be us. It could, there's a lot of people who are now trying to apply what's called formal methods to this AI-generated code, which you can basically think of as like writing mathematical proofs that the code must be correct.

Yeah. There's a startup from Convergent Research, I think, that's doing this, right? Yeah.

There's a bunch of startups doing this, actually. Okay. You know, I'm a little skeptical of that, but I wish them luck. Like, if that worked, it would be amazing. You know, there's actually a lot of different people trying to attack this problem from different ways, because I think people are starting to wake up to like, oh no, this is actually a problem.

Yeah. And I mean, what do you, Do you use everything or what do you use if you are coding with an AI?

So we at our company by and large don't use AI code.

Okay. But like if you're— I mean, just when you're kind of checking the state of—

Oh, I mean, I just use Claude.

Yeah.

Yeah.

And I mean, how does it strike you?

Because it's really good. Yeah, it's really good. It's really impressive. It's actually super good.

And do you think it— do you— is— I mean, you're somebody who's like studying this deeply. Do you feel like— is your intuition that it's already plateauing or is your— because it seems to be getting dramatically better.

Yeah, no, I don't think the— I think if you look at, you know, Opus 4.5 versus Sonnet 4.5 versus, you know, whatever it was before that, I don't even remember anymore. The trend line is still pretty good. Like they are still getting better for sure.

But it seems like you are kind of— you do have this gut feeling that it might plateau. Quickly unless I'm— I—

let me, let me put it this way. There's like, there's like a few different possible worlds here. One is the world in which it plateaus, and I think we should be really ready for that. Another is the world where it does not plateau, and I think we should be really ready for that. And I think that those are both worlds that have tremendous promise, and there are tremendously exciting things about those worlds. And they're also both worlds that contain tremendous danger, different kinds of danger, but both very dangerous in their own way. And I think, like, I just think people should just be having this debate and arguing this out and talking it through and not like, you know, I feel like there's been like a slightly unhealthy dynamic, especially on social media, especially on LinkedIn, where if you have any questions or reticence or whatever at all, it's like, oh, you're a hater. It's like, no, I'm really not a hater. Like, I've thought AI was super cool for like 20 years. I have tons of friends who work in the AI industry at like Frontier Labs. By the way, a lot of them are like trying to figure this stuff out and arguing about this and talking about this. It doesn't mean I'm a Luddite just 'cause I, you know, have questions.

It does, however, seem like, I mean, your company is on some levels fundamentally a bet, you know, against AI on some, against the abilities of AI on some level.

I don't know about that. I think that if AI gets as good at writing code as human beings, let's say, it's going to be a lot faster than human beings. There's going to be a lot more code in the world. I think we're going to make a lot of money in that world, right? Like, there's going to be a lot of code to test. Human beings do not write bug-free code. Code. Really, really, really smart human beings do not write bug-free code. If AI gets so smart that it can write bug-free code, I just like, I think that that is a world in which civilization looks unrecognizably different. And like none of the bets that I am making or that pretty much anybody else is making make any sense at all. So I'm kind of willing to bet against that extreme world. Right. But in any of the kind of intermediate ones, I think we have a business. Okay. Yeah.

And not to give you too big of a chance to talk your book, but like, you know, you were saying when you first started this company, people were skeptical this technology would work. Even if they kind of believed it would work, you know, they weren't really sure of the use. And then, you know, clearly on your website, you guys list customers and things like that. And so, you know, now that you've been at this for— you were in stealth for 5 years. It's been what, like a year, year and a half that you've been—

Yeah, like almost 2 years.

Yeah.

Okay.

So I mean, as you're going around giving this pitch, like, I'm sure you'll tell me it's being very well received. Who, who, who is like more receptive? How are you changing people's minds who were skeptical?

No, no, look, I, I, I, I, when we started the company, I, Dave, my co-founder and I wrote down on like an 8.5 by 11 sheet of lined paper. Paper, like list of risks. And I still remember this piece of paper and it had a lot of risks on it. And I think the first one was like, this probably won't work, you know? And then, and then you get down the list. And I think one of them was like, nobody cares. Like nobody cares. Like the revealed preference of the world seemingly is that like people are okay with software that doesn't work. Can we actually convince people to care? And then the next one after that is like, can you convince them to pay real money for this? And like, I do think I've been kind of pleasantly surprised on those two. Like, we, we came out of stealth, we announced this, and like, yeah, the reception was really good. A lot of people were very, very excited. Um, and a lot of people are paying real money for this because I think they recognize that it's like a really good investment and it speeds their team up and avoids disasters. And it's like, it's good. Um, in the last year, the AI thing has really started to take off. And I think what's really cool is that they're sort of coming, most people are coming out from the opposite of, of, of what you said. They're kind of, they're kind of saying like, look, I have AI writing all my code now. I have no idea what it does. I have no idea if it works or not. I really need something to tell if this works or not. Sounds like Antithesis could help me with that. And so it's actually been very good for business.

It is funny. I, I should have brought this up earlier cuz I had the thought last night. I mean, you guys did start this company, AI was like a thing, but it wasn't like this thing and it definitely wasn't coding like it is now. Now. So you were, you were, you're kind of like making a bet on some of the things you outlined before, that systems were already complicated enough and there's all this important infrastructure and, and people's need, obviously they had code they needed to deal with it, but then, and now AI's added this other wrinkle to all of this. Yeah.

I, I kind of think that most really successful companies have like two pretty important ingredients. One is like having some great insight or foresight, and then another is just like plain stupid simple luck. That you did not deserve in the least, you know? And I think AI coding was like that kind of second thing for us. Even if, like, even if it did not change the facts on the ground at all about how much people need what we're selling, it's suddenly made every single person think about what we're selling. Suddenly everybody's like nervous about it, and that just makes it a lot easier to sell.

I don't want to like nerd people out too much. How do you actually run this? I mean, do you run this in somebody's cloud?

Yeah, yeah, yeah. So we just run it in our own hosted environment and basically—

Like your own data center or—

We run on AWS.

Okay. Okay.

And for, for people who are really, really security conscious and we have some very secure— I mean, you mentioned Palantir and Jane Street. They're both very security conscious people. We have the ability to run it in their cloud account instead, but that's like kind of a newish thing.

Okay. Can I tell a Palantir story? Story for a minute, please? I think I wrote the first big magazine profile about Palantir, and they used to do— like a lot of companies do this pen testing, penetration testing, which like you can do on, uh, software systems. They were doing it on like their physical, on their building. And I remember a story, they, they, you know, they hired some company to like, you know, push the limits of what could be done to break into Palantir because they, they were right here in Palo Alto. And it was so funny because Palo Alto is very sedate and it's just like coffee shops and diners restaurants and, and I worked right next to Palantir and, you know, all of a sudden as the company's getting big outta nowhere, there's all these dudes with earpieces walking around and all this stuff. And so I remember Karp was telling me a story back in the day, they, they put this company, they hired, put a dude in a shipping container, like a big wooden box and like shipped him into Palantir and he had like a breathing apparatus or something in the box, you know? And he's, he's just like staying in there until night comes and then like, break it out of the box and get it to Palantir. Yeah.

Um, yeah, well, we, we, so we have to be pretty security conscious because like often people's code is like their crown jewels and what do we have? We have their code and we have like a list of all the ways their code can break. So that's like pretty, pretty scary. So we actually did have one customer. It was not Palantir, it was somebody else demand exactly that. They wanted to be able to do physical penetration tests in our office and we gave them that.

Okay. Did you, did they, they didn't, they didn't take us out of the box.

They didn't do it. No.

I mean, I know I asked this before,, but so I'm, I'm a, I mean, aren't most of these people running in the cloud anyway for the most part? I mean, I'm sure you're getting like Jane Street or a financial services company. Maybe they have their own data center. Um, I, it's like, I don't fully understand how you replicate like what I have running on AWS in this virtual system. And that is what happens.

Yeah, exactly. Yeah. Yeah. So it, I mean, look, it takes a little bit of work. This is part of why it was hard to to do, but, but the thing is, people have already gotten really good at this. Like AWS and GCP, you know, and Microsoft all have an extremely vested financial interest in making it easier for you to lift and shift your old legacy on-prem thing into their cloud. Right. And so they've built all these tools, which are like very general purpose tools. And they've put their weight behind all these standards, which are open standards. That just kind of make it easy to do that. And the thing about computers is like, they're easy to virtualize because they're computers, you know, they're all Turing complete. And so, and so it's just not that hard for us to do the same trick that all the cloud providers do and take that logical description of your systems and port it into our crazy world.

Okay.

And look, I'm, I don't want to downplay this like if you have a really big, really complicated system, it's still going to be a little bit of a pain in the ass, right? And so we have our version of like Palantir's forward deployed engineers who can like help you do that. Yeah, yeah.

To do all this stuff. Okay.

That's right. But, but, you know, what we're finding is that over time that gets less and less necessary, both as our stuff gets more polished and also just as the world is just converging on like, of course you run it in the cloud.

Yeah. I'm going to ask you a question now, which I hope doesn't come off as badly as it's like playing out in my head. I mean, I had, you know, before you came on, Man, for, for like the last year I've had all these people tell me, gotta talk to Will, he's this fascinating dude. Um, I think we met at Frontier. Yeah, yeah, you said we were chatting about SpaceX.

Yeah, I was a SpaceX fanboy since like before the Falcon 1 made it to orbit.

Oh wow, you go back, man. Yeah, yeah, so you were watching all the early—

I was watching all the explosions.

Yeah, okay. Yeah, yeah. Um, my question was, was kind of this, and I say it with all due respect, this is kind of what I asked the Collison brothers sometimes.

It's like, why are working on this.

You guys are interested in so much stuff, but you know, I mean, you're a mathematician, you mentioned philosophy. I can just tell by the way we're talking and some of the speeches that I saw you giving. I mean, you have this big picture thought of the world. You're a curious human. And so yeah, you're working on an important problem and I'm not trying to trivialize this, but out of all the things you're like debugging software for a living.

Yeah. I'll make, I'll make your question more rude. Okay, who gives a shit about software quality? Why, like, why do you— why am I doing this? Yeah. Okay. I think that there are two things that are both very true. One is it kind of doesn't matter how good at talking you are or how smart you are or how hardworking you are or really anything if you haven't at some point kind of done your own thing. Like, I think most people kind of don't care until you have actually done something real and done something difficult. And so, you know, that's why I decided to start a company instead of just staying at Google where I was very happy in a lot of ways.

Yeah.

I think, you know, I think the second question is like, why this one? And I think it's sort of precisely because it's really boring. Boring and because most people don't care about it and think it's stupid. Like, you know, Peter Thiel has this like pretty good advice, which is like, just don't, don't engage in competition.

Yeah.

Like don't, don't try and be better than everybody else at a thing that everybody's trying to do. The great thing about software testing is it's a thing that nobody wants to do. It's so awful. Like this is like, you know, it's like a, it's, it's like janitorial. Most, most really smart engineers are like, oh my God, take that away from me, right? And that's like, that's amazing as an entrepreneur because it means that you have nowhere close to the number of competitors that you deserve to have. And it means that there's a lot of low-hanging fruit here, actually. You know, that's kind of my whole thesis is that like this, this whole area of the world has just stagnated for decades and decades and decades. And there is a tremendous opportunity here.

Like, I should know this. I would have known this earlier in my life because I covered databases. Was, was FoundationDB open source? There was some MySQL component to it. Yeah.

So FoundationDB was closed source while we were an independent startup.

Okay.

And then Apple eventually open sourced it.

Okay. Yeah. I mean, that was a stupid way to get to my question, which is, I don't know, you know, when I first started out as a tech reporter, I don't want to say software was romantic, but there was, there was a romantic element to it where it was still a bit newish in the grand scheme of things. And, and it was a craft and you're writing and there were these wonderful wars between the proprietary and Open Source camps, and it was this, this battle over ideas, and, um, people got really passionate about this stuff. And it, it didn't quite play out in like the mainstream press, but, but kind of close, you know. And Microsoft was evil, and the other people were the freedom fighters. And, um, yeah, I don't know, like, maybe I don't— I'm not as dialed in as I used to be, or maybe it really has changed. Like, like, has software gotten kind of taken for granted. And, and, you know, I always think of— I especially like open source and Linux. I think of it as like almost these economic marvels that are underappreciated in a lot of ways.

Right. Well, I think there was a— I think there was a real demographic transformation that happened, right? Like, you know, in the '80s and '90s, only weirdos went into software.

Right.

It was like a very— it was looked down upon, right? If you were a, if you were a smart kid from an Ivy League university, you'd obviously become a lawyer or an investment banker.

Yeah.

Or a doctor, right? Like, you wouldn't go work at Google, or like the, you know, back then there was no— you wouldn't go work at Oracle. Like, that would be like a totally weird, like, sort of disreputable thing to do. And then what happened was software got super prestigious. And so when something gets super prestigious, it does just mean that you get a lot of people doing it because it's prestigious. And that is in some ways good and in some ways bad, but it definitely changes the character and what it, what it feels like to do that thing and, and, and who else is doing it. And so I think, yeah, I mean, software has become big business, right? It does not have that. I, I remember the Linux versus Steve Ballmer wars like very well. That was a, that was so much fun. That was so much fun. And like, you know, I was a Linux guy and yes, I did feel like I was part of a guerrilla army.

Yeah.

Like it was, it was totally fun. And yeah, you're right. These days it doesn't feel like that.

Yeah. When you were laying out some of your other scenarios, I was playing things out in my head. I mean, you could see a, you could see a case where some countries, um, more developed economies take the issues you're talking about seriously. They're trying to guard against this giant AI swarm. Slop problem and the undermining of their own infrastructure. You could see other places where it's like maybe they don't have that luxury of time or whatever. And, you know, they've just got these AI code armies and they're doing stuff too. I mean, they're making planes and trains and things that are out in the world.

Yeah.

Like, you know, you made a point in one of your talks, it's just like, this is so cheap to do that the temptation is truly massive.

Um, well, yeah, and that's one reason that I think people, um, I like— I think in some ways I'm more bullish on AI than people, because I think that even if AI doesn't get super smart, it will still cause a profound economic transformation because it's so cheap.

Oh, and wouldn't you—

I mean, it's like, it's okay if it's not smarter than a person, who cares? You just run it 10,000 times in the time that it took that person to take a breath, and you're kind of rising, and it might— Well, it's hard to say. It's, it's really hard to say, right? Like, you know, Dwarkesh had a, had a great tweet that kicked off a firestorm a couple weeks ago that was like, look, if, if AI actually becomes competitive with human labor across a very, very large sector of the economy, and if, you know, all of the returns on that are going to the people who own the GPUs and own the shares of Anthropic stock, Like, you know, what, what happens to inequality? Like, isn't, isn't this just like— and I actually feel like Dwarkesh in some ways didn't go far enough with that tweet. It's not just income inequality to worry about. It's like, will people continue to have rights if you don't need human beings to run your economy anymore, right? Like, I, I think, like, I think people are maybe a little bit, uh, heads in the sand as to how big this could be, even if AI does not get super intelligent.

And just because you brought up the legal profession in one of the other things I was watching, I mean, you know, because we've been talking so much about code and, and tech land and how that, that could be affected by all this, I think lawyers tends to come up pretty quickly as well.

Yep.

I will confess, you know, it's like sometimes just doing like a podcast access agreement, it's like, am I going to pay a lawyer $1,000 for that? Or You know, I'm pretty sure my ChatGPT does an okay job on this one-page document. Do you have a— not to put you on the spot in the legal profession, but yeah, do you have like an intuitive sense when you're talking about this, this kind of dumbing down of a craft, of a profession over time? You know, like, does it— in code, there's these real-world consequences if your plane's crashing. I guess the legal profession would have some too.

There's consequences. And by the way, human lawyers lawyers make mistakes all the time. Like, you know, one of the things I learned from, actually from my co-founder, was always read every document, every single word, even if a really smart lawyer gave it to you, because they make mistakes. Um, so yeah, I mean, I like, I think the legal profession is gonna experience some pretty profound changes. I have a really good friend who's a lawyer who was just like DMing me the other day. He had just tried some new AI legal tool and he was like, oh my God, like it's over. Like, You know, I, I, I don't know exactly what's gonna happen there.

Yeah. It's like not a novel conversation. People, I play this outta my head all the time though. Like if my son was, one of my sons wanted to be a lawyer, I would, I, my dad's a lawyer, my wife's a lawyer. I'd be like, no, for, for sure. But then at some point this has to catch up with you because you have to have somebody who knows, right?

You need, you, you need some associate so they can become a partner, right?

Like to advance the field, right? Yeah. Yeah. I mean, I mean, it's, and I, I don't actually I don't understand how this is going to work, right?

It's possible that there's a similar kind— okay, here's a, here's a potential scenario. Um, there's kind of two kinds of lawyers, right? There's the lawyers who kind of write policies that nobody's ever going to read and that nobody was thinking about while they wrote it, and, and the lawyers who then like write things to like comply with those policies that nobody has ever read, right? And you could imagine a world where that ends up becoming kind of all AI. And it's like, I just have my AI, you know, creating, like, doing some box-checking exercise, you know, writing some compliance materials for your AI that then reads them. And neither of us actually ever cared. And, you know, we've just made everything more efficient now. And that's great. And then you have the other kind of law, which is like, at least in corporate law, which is like really big deals or like really big important agreements that people really, really, really care about. And maybe that will have some degree of AI assistance, or AI like, hey, make sure I didn't, you know, misnumber my sections here. But there's still a human being who's carefully scrutinizing it and thinking about it.

I know, but it feels like you want that human being that suffered through 20, 30 years of experience to get there. And this is like unclear to me on the writing front as well, as like people are turning more and more over. Like, you know, at some point you just don't get anyone who's good at the craft because they haven't sort of done all the basic mistakes you have to do.

Yeah, I, I don't know. I, I, I actually really like writing.

Yeah.

Um, I, I write, you know, every blog post that appears on our company blog under my name, I actually wrote it myself. And, you know, I think AI is making people worse writers, but it's also making the good writers stand out more.

Yeah.

Like, you know, it's actually easier now in some ways to like just— you just have to not produce slop and people will notice you.

Yeah.

And I feel like that maybe is like the hopeful message for, for kind of every field, right? It's like in some— if we do head towards the idiocracy world, like, you know, it's easier to stand out.

Yeah. Yeah. Well, I know that you're going to meet with my arch rival and very good friend, Cade Mattson. I should be respectful of your time. Am I allowed to say that? Yeah.

Yeah.

It's Cade. Be fine. With that. He's— we go way back. We used to work together at The Register. I hired Cade.

No way.

Yeah, back in the day. Um, uh, good guy. Does amazing reporting on AI. Um, well, any like last words of wisdom for Silicon Valley from somebody who's been in the thick of it and now has a little separation? And, and, uh, how are we looking for— from Virginia at this particular moment in, in time with our AI frenzy and our, our bio tech madness?

I would say that the superpower of Silicon Valley is everybody doing the same thing, and that people should not be afraid to do something different from what everybody else is doing. If everybody else is AI crazy, you could try and do something else. If everybody, you know, the thing that's kind of pissing me off right now about Silicon Valley is I feel like everybody's engaging in very short-term behavior, and that a lot of the cultural norms that made Silicon Valley really great are like kind of eroding a a little bit. And, you know, you can just not do that. You know, like, you can actually try and build your company in a different way, try and treat people in a different way. And I think people who do that are going to get rewarded in the long term. And also it's just the right thing to do. Yeah.

Okay. Well, there you go. Will Wilson has given you words of wisdom. Thank you so much for coming on, man. It was awesome getting to chat with you and know you a little bit better. Is there anything like major people you've launched the company? I don't Anything big in the, in the nearish term future people should be—

See, this is like, I'm, I'm so bad at shilling. I, this is like, you know, I don't know, probably, probably our head of marketing's gonna shoot me after this, but yeah, no, just if it sounds interesting, check out our website.

Okay. All right. Thanks. Well, thanks so much for coming on, man.

Thanks for having me.

The Core Memory Podcast is hosted by me, Ashlee Vance, and/or Kylee Robison, or both of us together. Together. It is produced by me and David Nicholson. Our theme song is by James Mercer and John Sortland, and the show is edited always by the John Sortland. Thank you so much to Brex and E1 Ventures for all your support, and thank you most of all to everybody for listening or watching. We love you. Please leave us a like, a review, a subscribe, all those tremendous things. Thank you, and we'll see you again.