Will Wilson: Antithesis uses deterministic simulation to find rare bugs — Jane Street went from customer to strategic investor

testing standard. They have an incredibly stacked list of investors. Dores Patel, Shelto Douglas, and Patrick Collison.

Whoa. Two out of the two out of the three roommates.

Yeah. What happened? What happened? You didn't do a road show where you go from bedroom to bedroom showing off what you built and they're like, "Do you want to invest sharks?" I imagine you sit down and you say, "Shark. Shark. Here's what I'm building."

You know, it's it's funny. I think I think they all ended up coming in for very different reasons, but we're certainly very

You got You got Dylan. He just wasn't he wasn't big enough to get into the press release.

Get to the press release.

Brutal. Brutal.

I cannot comment on investors who may or may not be in our press release.

Okay. Well,

uh well, thank you so much. Uh please uh explain this to me like I'm not Dwarash and I'm not a researcher at Anthropic. Uh what what are what are you building uh at at like a very very basic level? Who's the customer? What's the problem you're solving? Uh, and then I do want you to walk me through some of the technology as much as we can, but I'd like to start with just like the the highlevel problem description if possible.

Sure. Okay. So, the super basic problem description is like Mark Andre once said that software is eating the world. I think that's true. I think AI has made that more true.

And the thing is that's actually kind of a bad thing in some ways because software doesn't work very well. Like like we I feel like we all have Stockholm syndrome about this. Like when I when I walk into a building, I don't wonder in the back of my head, is this building going to collapse on me? And yet, every time I use a software product, I know there's some random chance that it's just going to totally fail.

And if you think about it, that's that's really not okay. And that's what we're that's what we're trying to fix. We don't think it has to be that way. Um, we think the reason it is that way is because people have been building and testing software wrong for the last 70 years, which is like maybe a slightly hubistic and crazy thing to say, but we think we've got a better way to do it.

I have an ongoing war with Rune about whether or not the United app has gotten better or worse over the past few years because I use it as a heristic for like AI diffusion because he'll he'll tell he'll say, "Hey, the AI models are amazing. Like you you just you just yell at the codeex model. It'll do exactly what you want." I'm like, "Oh, really? Well, then why did I have a bug in my United app when I was trying to fly across the country?

Well, maybe they haven't adopted

and then he'll tell me it's actually way better. It's actually they actually are improving their software. But, uh, walk me through deterministic simulation. Like, what does that look like in practice of I have a piece of software. I have some some vibecoded sloth app that I'm worried about completely blowing up and leaking all my users data and, you know, and failing to work in a bunch of different scenarios. Like, like how am I actually using your product? Am I plugging in? What's going on? What's the workflow?

Yeah. Okay. So, so, so here's the basic idea. Um, so like the traditional way people test software is you have your app and you sit down and you think really hard and you use your big brain and you're like, what are all the situations that this thing could run in? What are all the things that people could do? What are all the situations it has to handle? And then I sit down, I the developer, right, and I write a bunch of manual test cases for each of those things and make sure that they run in my CI system or whatever. And it's like, hooray, I've tested my software. except you've totally not tested your software, right? Because the thing that's actually going to blow up and totally screw you and and be a disaster is not something that you thought of. It's something that you didn't think of, right? It's like an unknown unknown.

And so we have the exact opposite approach. We say instead, we're just going to take your actual software

and we're going to run it in a simulation

of the real world and we're going to throw the kitchen sink at it. We're going to like have the worst users ever doing the worst things ever and it's going to be running in the worst data center ever and like computers are going to be crashing and like hackers going to be trying to break it, whatever. Like it's just going to be really awful. And then we're going to see what actually happens empirically. And you can tell us, hey, here's the stuff I promise my software will never do and we're going to see if we can find a way to make it do that.

Why Why are you framing this as deterministic simulation? That sounds like a whole lot of nondeterministic stuff going on.

Yeah. So the key is we have the the slightly crazy computer science thing we've done well no one of one of the slightly crazy things we've done is we've basically built a way to run any computer program fully deterministically which is which is a little nuts but it's really really important because it gets rid of the like works on my machine thing forever. So if we find some really weird rare software problem which only manifests one a billion times but like you know it launches the nuclear missiles and so you really don't want it to happen even one a billion times.

Um we can perfectly reproduce that. So if we find it once you can just replay it over and over and over again until you're able to fix it.

Interesting. Uh what's the

what is the what is the like sci-fi scenario that you're building against? like is this is part of this tie-in. Like I'm trying to understand why like specifically why you got the roommate three Pete because software testing is cool but I imagine there's like kind of this like you know 10year vision like AGI vision where we need to test for that you know one out of a billion uh uh times that the nuclear missile codes get launched or the whole surface of the planet is blanketed with solar.

Yeah. I mean, you know, I think I think if you believe that we're about to go into like a very exciting AI future, then I think there's probably going to be a whole lot more software written.

And probably some of it's going to be written by entities that are not like you and me. They're like some kind of AI agent. Maybe we don't fully understand them. Maybe we don't fully understand their incentives or their motivations or their failure modes or whatever. Seems like having a really really good way of telling if software does what you think it does would be handy in that scenario.

That makes sense. So, somebody makes an an AI agent that uh that like processes like things that were faxed in and there's like some, you know, faxed into a doctor's office and there's some edge case where it it hires a private army and like tries to kill the fax machine company and you're like, well, an agent wrote this agent, so we need to actually test to make sure it's not like secretly like completely insane,

right? You've got that and then you've also just got more prosaic and boring situations, right? Like I hired an AI agent to like write the next version of the United app and like oops. Like if I'm if I'm in the middle of booking a ticket when my subway train goes into a tunnel, it books like 10 million tickets. Like I don't I don't want that. United doesn't want that, you know.

No, I I like I like this opportunity. You can sell you can sell the product against like very real problems today and also get the and still like, you know, have the this like bigger very like real kind of like sci-fi element to it.

Okay. So I I I have the interpretation of Roommate Ventures coming in. Uh I don't have the interpretation of the implications of Jane Street coming in. Uh is this is this relevant to highfrequency trading software or are they just purely financial investors? Like how strategic are they as a partner?

Oh yeah. So the the way we got Jane Street in this actually is just that they it started out as them being a customer. Okay. um they've been

one of our best and largest customers for you know a little under a year

and um have really put the system through its paces and are using it to do some really really cool stuff.

Yeah. And basically what initially happened was we were raising around. We planned to go the standard Silicon Valley route, you know, and um and we were just going to have Jane Street put in like a small amount of money as a strategic.

And then they offered to lead the round, which we were not expecting. And you know, we sort of thought about it and we were like, "Wait a minute, that's actually the coolest thing ever." Like, you know, what a what an endorsement. And you know, from one of the greatest names in software, like how can we not do this?

Have you considered writing your codebase in Okamel? I think I think it's likely that there's going to be excellent Okamel support in our platform.

Fantastic. That's the best news I've heard all day.

Hit the gong for that.

Uh very very cool.

That's great news.

One of one of the one of the more one of the more unique uh fundraising announcements we've had this year.

Yeah, this is a fun fun project. I really like it. Uh well well congratulations and thank you so much for taking the time to come chat with us. I'm sure you'll be back on soon. I'm sure you're already getting unwanted term sheets, you know, hurled at you.

I believe it.

Enjoy enjoy processing those.

Yeah.

Awesome.

Uh

have a good one, guys. Thank you.

We'll talk to you soon. Great to meet

uh and that brings us towards the end of our show. Perfectly timed 2 p.m. on the dot. We are in fighting form. Uh we're going to play you out with some Christmas music, I suppose. Uh, in other news, markets are predicting that the time person of the year will either be the pope who's concerned about AI, the biggest promoter of AI, Sam Alman, the supplier of the hardware behind AI, Jensen, or just AI. Uh, it seems almost certain it's a lock that time 2025 person of the year will have something to do with AI. Um, in other news, where where else is Oh. Oh, uh, we didn't get to SpaceX news, but, uh, it was rumored that SpaceX uh, told or was reported actually in the information by Katie Roof that SpaceX told their investors, "We're aiming for APO, IPO, late 2026." Late 2026, we're going to be a public company. Uh, but a lot of people are interpreting this, reading the tea leaves. You have Ken Kirtland saying, I'm not exaggerating. It's not doom posting to say this, but this is the end. They are not serious. They are not serious about any part of Mars. If this is done, there's no world where shareholders will let you spend tens of billions of dollars with the express intent of no return, which is what Mars settlement is. In fact, it would be illegal, literally illegal to do so. Uh, not true at all. Look at the metaverse. Look at what Meta did. uh if you have control over the company you can you can spend what he said $86 billion is what Z

I'm just expecting SpaceX's telecom business to do tens of billions of profit.

Yes. Yes. If you get big enough I mean SpaceX I think to date has burnt something like 10 billion across all the fundraising and a lot of that's been like secondary stuff. So like they haven't burnt that much. Uh and of course people are dunking on Zuck right now for pulling back on the metaverse. Austin Allred had a good post here. He said fire Mark Zuckerberg uh fired kind of botch this but said uh Mark Zuckerberg fired because he lost the company 77 billion should probably consider that he also made his company 1.6 trillion and so uh the same thing could definitely be true. So I I believe if SpaceX went public that would not necessarily take the foot off the gas of the Mars mission. I don't think that's necessarily what would happen. Um but Dan Primac has a little bit more nuance of a take here talking about what might be happening behind the scenes. says SpaceX prepping late 2026 IPO per the information, not just Starlink, which is what people were rumoring they'd spin out Starlink. I don't think that makes any sense. Um, but this comes just after just comes just days after reports that OpenAI was weighing an entry into space. So maybe it's a bit of PR maneuvering given that Elon Musk doesn't seem to like having a public company. He took Twitter private and then he wanted to take Tesla private, too. Um, and then uh and then Katie Roof says, "Does he even have PR people, lol?" And Dan Primac says, "Don't need PR people to have a PR strategy." And uh yeah, he might have he might have he might have directly emailed all the investors that he knew would leak, you know, and that doesn't require a PR person. And so Elon actually chimed in. He said there's been a lot of press claiming that SpaceX is raising money at $800 billion, which is not accurate. SpaceX has been cash flow positive for many years and does periodic stock buybacks twice a year to provide liquidity for employees and investors. Valuation increments are a function of progress with Starship and Starlink and securing global directtoell spectrum greatly increases our addressable market. And one thing that one other thing that is arguably most significant thus far, blah blah blah blah blah. Anyway, ridiculous.

John, I got to say I'm bummed that we got to end the show right now.

We do. We do.

You got to get on with Pasadena.

I do.

You got some celebrating to do over in that in that area.

Neck of the woods. Um, I can't wait for tomorrow. I'm so happy to be back in the Ultradoo. Glad we don't have to travel for the rest of the year.

No travel for the rest of the year.

We're hanging out.

We're podcasting.

The team is thrilled. We're thrilled. Thank you for hanging out with us today. We love you all dearly. We'll talk to you soon. Have an amazing evening. We will see you tomorrow.

Goodbye.

Cheers.

Bye.