Factory’s Matan Grinberg on Recursive Self-Improvement, Chinese Models & Lessons From String Theory

We are post-AGI right now. Like, we have AGI. The only world in which there's like, there's nothing left for humans to do is if we solve all problems. But my sense is that the number of problems that exist is like here, and the number of problems that we're solving is like this. There used to be a moat in software of, we know how to build this and no one else does, so that's why we're gonna win. That doesn't exist. There is nothing that nobody else can build, which is crazy. People talk a lot about recursive self-improvement, the idea that the models get so good that they start training themselves and there's this loop where they just get, you know, insanely good. But improve themselves is not That's synonymous with like hyperbolic growth. It could be the case that we get to RSI and it's just never as good as human-assisted improvement on these things. What happened to a world where if there is a security incident, you are like deeply ashamed and responsible for it, not like bragging about it in a research blog post? No open models or open models that are from China? I will take open models from China for sure, because the threat of the only good models being from closed providers is ultimate monopoly, like the monopoly to end all monopolies. It's the interest of the entire economy. Matan

Grinberg, welcome to my show.

Thank you for having me.

I am so excited you're here. And a disclosure is that you are the first person I've had on the podcast that I also have the honor of being a very small angel investor in. Too small.

We can change that.

And I actually, when I was preparing for this episode, I went back To see what was the first message I ever sent you. And I remembered the whole story, which is that I was a year and a half out of college. I was working at a venture fund in New York. I was using LinkedIn Sales Navigator to find people that I thought had amazing backgrounds who seemed like they were doing nothing on LinkedIn, which means probably starting a company. And I came across your profile and I was like, who is this guy that working on a PhD in string theory has done all this crazy physics work, And it's just like for the moment doing nothing. Went to your Twitter, it literally just said Factory, and I was like, what is he doing? So I messaged you. I think this was like May of 2023, and I was like, hey Matan, you have an amazing background. I would be really interested to learn more about Factory. And that was the beginning of how we met. And, um, the rest is history. I think that was like a month after you had officially started with—

That, that is by far, by the way, like the first like investor inbound that we got. We, I think we were 2. It was me and my co-founder, you know, maybe our first employee had started or was like about to start. But that was crazy. I remember thinking, how on earth does she, like, it was like I changed my Twitter bio and within, you know, an hour or so it was, that was crazy. So props.

I was like FBI, but I honestly, like the first time we got on a Zoom, I was so impressed with your energy and your vision for Factory. And I actually looked at all of these things and I was texting one of my best friends who's also an investor. And I was like, I just think Matan is going to be so successful. He's got the sauce. So sitting here 3 years later, it's really so exciting to see how far you've come and everything has kind of come to fruition. And we're going to get into all of it, but I would really like to start from the very beginning. Of your story. Matan, who were you as a kid? Who was young Matan?

Young Matan. That's funny. Young Matan grew up in the Bay Area, so in the South Bay. My family moved around a lot. I probably moved, I think, 10 or 11 times before I went to college.

And why did you move so much?

It's so funny. So my parents were immigrants from the Soviet Union. They came from Moscow and Moldova. Um, and they could never decide if they wanted to be closer to my mom's work or my dad's work. And so my dad was in like South South Bay. My mom was, uh, she worked at Stanford Children's Hospital, so she was around Palo Alto. Um, so we kept like alternating between like Saratoga and like Palo Alto and then Cupertino and then Los Altos. And so moved around a bunch between there.

Power struggle.

Yeah. Yeah. Um, and, uh, and it was also, you know, their like immigrant mentality of like they were trying to you know, see if along the way they could like, you know, uh, if they, so, you know, sometimes they would buy a house and try to, you know, upsell it after a year or something. Uh, sometimes they would just rent, but, um, so that was kind of the background there. Um, as a kid, I was very much a troublemaker. Like I would get into so much trouble in elementary school and middle school. I got suspended so many times.

Really?

I was also really good at math, but I think, you know, retrospectively, I think I was like bored by things. And I thought I was like too smart for stuff, so I would like skip class and like go and just, you know, wreak havoc. Uh,

were your, were your like immigrant parents from like the Soviet Union? I, I imagine that they really valued education. Like how, how did they feel about that phase of your life?

Yeah, I mean, it was like, you know, it was a huge disappointment. Like every summer was like math camp or like physics camp or, you know, some, some combination of those. Um, so they were very much doing the, you know, the right thing there. And then it was just like they would get a call from the principal being like, you know, here's what Matan did again. Um, so they were very, uh, you know, dis— displeased, I would say, or they were concerned. They were worried if they were doing something wrong. Uh, my part, it makes me feel bad thinking back cuz like they, I just put them through so much chaos.

It all worked out. I'm sure they're, they're very proud now.

Yeah.

Um, what set you on the physics track? And, and, and what about string theory hooked you?

Yeah. So, um, so funny enough, um, so in 8th grade, so like was about to transition into high school, um, I still like was not a particularly good student, but I was really good at math. And I like, I knew I was good at math and I, you know, probably had some ego about it. And then my 8th grade geometry teacher, uh, at the end of the year, like while everyone was preparing to go to high school, She was like, oh yeah, Matan, you need to retake geometry in high school. And I like, I was so offended and pissed off by that. And it was like, she thinks I need to retake geometry. Like I know all this stuff. And so out of spite, my first order on Amazon, even like today, like if I open up my Amazon app and just scroll all the way down, my first order on Amazon was textbooks for Algebra 2, trigonometry, precalculus, like Calc 1, 2, 3. you know, differential equations. And, you know, I was basically like, okay, over the summer, like, I'm going to show her, like, I'm going to study all of these textbooks and I'm going to place out of, you know, all these classes. And then I asked my dad what the hardest math was. And he said string theory. And I was like, okay, I'm going to be a string theorist. And it turns out, I mean, string theory is theoretical physics, not exactly math, but, you know, close enough. And so, you know, I ended up just spending the whole summer doing that. I went to a very small, uh, private school in, uh, South Bay that no one has ever heard of. Like, my graduating class was like 10 people. Um, but the one thing that was great there is that they allowed me to take online classes for credit. And so I'd like placed out of all those math classes and then I was like, okay, physics is like the only thing I care about. Like, I want to become the best physicist in the world.

Wow.

Um, so it's kind of out of spite, uh, was the, was the origin, but it actually turns out that the mathematics of string theory is Probably like the most, one of the most beautiful artifacts that humanity has created.

In what way?

Like, I guess just taking a step back, like what even is string theory?

So. Yeah. What is string theory?

Yes. So, okay. So yeah. So, okay. So.

You're like talking about the summer of math textbooks. It genuinely sounds like my worst nightmare.

No, it's okay. So, so the 2 most predictive theories humanity has ever come up with. So predictive theory, right? Like if we want to, you know, roll a ball down a hill and predict exactly the dynamics of what's going to happen. But, you know, for broader, much more broad physical phenomena, the 2 most predictive theories humanity has ever come up with are quantum field theory, which is like the physics of the very small. So particles, electrons, photons, you know, all that stuff. The physics of the very small, quantum field theory. And then the physics of the very large things, general relativity. So that's like galaxies, you know, like all the big stuff. Both of these theories, we can test their predictions and it's up to like 13 to 15 decimal places of accuracy is how strongly predictive those theories are, which is like insane. Like it is so, so precise. Um, and so that's fantastic. It's like, okay, these must be universal truths because

we

can predict all these little dynamics so well. And so, you know, physicists naively thought, okay, well, if these are 2 of the best predictive theories we've ever come up with, surely they're both true and they work together and, you know, physics is done. But it turns out that if you try to bring these 2 theories together, everything breaks. Now, you might ask, where do these theories come together? Because one is the very small, one is the, you know, very large.

Yeah.

Why is that even relevant? Because, you know, they don't come together. Well, it turns out that black holes have singularities that are infinitesimally small, but extremely heavy. And so there you need both general relativity and quantum field theory. And it just so happens that when you try to merge them, like everything breaks, all the math looks atrocious, and just infinity pops up everywhere, and you can't make any predictions.

And obviously what you're doing today has nothing to do with that, but being so deep in that world, in that frame of thinking for over a decade, if at all, how has it shaped the way you view the universe or building AI or just everything around you?

Yeah. Well, I think the one good skill set of— so Like in the, the kind of the best paper that I ever produced. So like I didn't do—

With Juan.

Yeah, with Juan Maldacena. Um, but the best paper I produced, uh, was about basically 10-dimensional gravity, which is very abstract because, you know, look around, we're not in 10 dimensions, right? So that's, you know, uh, it's very much abstract, but I think despite the like actual, you know, things that we were working on maybe being a little bit more abstract, It's very applicable just to any work that you do because in order to even produce something in theoretical physics, there's like hundreds of years of literature that you need to catch up on. And also every day there's like 100 new papers that come out. And before you can contribute to theoretical physics, you need to know basically all of math, like algebraic topology, differential equations, like algebraic geometry. And these are all things that people will spend their careers going in on. Wow. And yet to do the physics, you need to at least have like a pretty deep understanding of these things. And so it really teaches you how to have, um, good judgment with minimal information.

Mm-hmm.

Because at the end of the day, like to write that paper, I didn't become, you know, the world's best algebraic geometer to do that. It was just like, okay, what can I like read through these textbooks as quickly as possible just to get the gist and get the intuition and then know that I'm not going to be able to read you know, all these thousand seminal papers that are really exciting, but just, there's no time for. And so it's like trying to parse out the essential information from what is like, there is no way you can actually—

And you're never going to have perfect information.

Like literally never.

And that's very true to company building. And yet you still have to make big decisions. So you write this paper with Juan Baldacena. You're in a PhD. What's the moment Where you're like, fuck this.

Yeah. Well, so, okay. So the paper with Juan was at my undergrad. And then I came to Berkeley just over there to do my PhD. And at Berkeley is when I realized that I didn't want to do this for the rest of my life, in particular because I realized, and I knew this the whole time, but it became viscerally true that, wait, like, to do theoretical physics, you have to be in academia. And being in academia means you're a glorified teacher.

Yeah.

And I'm not a particularly good teacher. Um, and I didn't get a lot of excitement out of it. And I was like, oh my God, the rest of my life is like just going to be surrounded by 18-year-olds.

Yeah.

Because like, I already know everyone. The string theory community is very small.

Also, I just to ask the elephant in the room is like, is there anything going on with string theory anymore? People are saying it's dead.

Like there's, there's definitely a lot of exciting stuff now. Is it the most fast-moving? Probably not. I think, and I think theoretical physicists are generally really smart, and a lot of them over the last decade have been leaving theoretical physics. I mean, there's a reason. I think like half the co-founders of Anthropic were theoretical physicists. Jared Kaplan was like exactly in the same area of physics as I was. Dario was a physicist back in the day, also at Princeton, but he was more on the bio side. Um, but like there's been an exodus of physicists to like more interesting problems. And I think that's for a good reason.

Okay. So once you decide that you don't want to spend your life in academia studying theoretical physics, what do you do?

So it was honestly like, it was a big existential crisis because like my whole identity was like physicists have with them kind of an arrogance sometimes, which, you know, I'm probably, I fell victim to that a little bit where you're a confident guy, but there was like, there's a quote from Um, uh, from, uh, what's his name? I think it was, I think it was Ernest Rutherford.

I

believe it was Ernest Rutherford.

We'll write it at the bottom.

Yeah, we can, yeah, we can like talk track over if it's not Rutherford. Um, but it's basically, he says, all of science is either physics or stamp collecting, which is basically like everything at the end of the day is physics or it's just like labeling stuff and, you know, it's kind of irrelevant. Um, and so there is that kind of arrogance associated with physics. And I kind of built my whole identity was I'm a physicist and then kind of all came crashing down because I was like, I don't want to do this. And I'd spent the last decade only caring about math and physics all day. Like that was all I did.

And that you never held a job, right? Never had a job. Like you weren't doing some research?

I'd literally never been paid for work except for like physics research when they would give grants. Um, and so. It was a big existential crisis and it was like, you know what? Okay. Like before I figure out what to do, like I should probably become like a normal human and like read literature. And so I went in and like looked up like the top 100 pieces of literature and just like read all of those or went to the IMDb top 250 movies and just like watched all of those. 'Cause it was like, you know, I need to like be a more well-rounded, normal person, which was fantastic, by the way. It was so much fun. Like, I remember, um, for whatever reason, um, maybe it was just because I was so relatively uneducated to the, like, the liberal arts, but A Tale of Two Cities by Dickens at the end, like, made me bawl my eyes out. Like, crap.

I haven't read that.

It's phenomenal. Um, but yeah, so did that, you know, that took a little bit of time and then I was like, okay, you know, this is all cool, but I'm not going to be, you know, a writer or something like that. And so realized I should make use of this decade of being good at quantitative stuff and either go into quant finance, which is what a lot of mathematicians or physicists do, or go into big tech or go into startups. That was kind of the decision criteria. I actually almost went into quant finance, you know, in New York, but last second had some good advice from some friends, which was like, you'll always be good at math. Like, don't do it. Just like stay in your PhD and explore. So, and prior to then I hadn't taken a formal computer science course. I'd learned a lot of programming just for physics, but ended up taking some classes and I'm very competitive. And there was something very competitively satisfying of not having taken CS, but going into these grad courses and like doing better than some of the CS students. So I kind of kept doing it just for the competitive gratification. And then it wasn't until I took a seminar in program synthesis. which we now call code generation, that just completely nerd-sniped me.

Nerd-sniped?

Yes.

What a word.

I mean, there's no better way to put it, which is like, physics teaches you that things that are fundamental are more important. Like, if it's not fundamental, it doesn't matter. Like, you always want the arbitrary case, the n-dimensional case, you know, the abstract case. That's like where the beauty lies. And there's something very fundamental about, you know, machine learning or AI, not for video or audio or images, which is so arbitrary, right? Like, it's so like kind of not fundamental, but there's something so deep about machine learning and code for the purpose of creating itself. Like there's something so just like primitive about that. And that just got me obsessed with that. So stayed at Berkeley for another year to do that. And then it became very apparent that academia is not the place to solve that problem, but to solve that problem, you kind of need to go into the places where there's very interesting, non-trivial code bases and engineering that's happening, which is like the industry. And it turns out that the best vehicle to do that is to start a company. And so I was like, okay, well, I don't know anything about starting a company. So I bought on Amazon again. I love going to Amazon for books whenever I don't know things, clearly. But I bought Zero to One by Peter Thiel. And then I looked up on YouTube, like how to start a company and like just watched some like random videos. One of which had, it was like a video podcast with Sean Maguire. Um, and then he was a former physicist and I was like, oh, I got to talk to this guy.

This is your guy.

And so that's how I ended up sending Sean a cold email.

Okay. So you reach out to Sean Maguire and he responds, you guys take a walk.

Yes.

Um, at that point, like, what were you imagining for Factory when you were thinking about like what you wanted to build?

Yeah. I mean, it's, it's actually shocking how The initial deck, it's like exactly what we initially wanted to do, which was like, I think at the time it was actually too early, but the whole point was our mission from, from like the first deck we made, the mission statement of the company to bring autonomy to software engineering. And so like not AI autocomplete, not AI IDEs, but like full agentic software development to automate the thing. And also the thing that I would jokingly say, like, If we're at a bar, how would we describe what we do? It's like we automate shit that engineers don't want to do. And it's automate, not accelerate and AI autocomplete, but just dump it to someone else to do. And it's been crazy to see how it's kind of being realized, probably a lot faster than I expected, but it is still roughly the same path. I think there's so much I didn't know just about enterprise and all hiring and everything like that. Okay.

So I want to quickly ask you about this period where you were like, I think you've just described it before as going through the desert where you had the vision so on point, like, this is where it's heading and this is what I want to build, but the technology wasn't actually possible. So it has now started obviously really working in the past maybe year, would you say?

Yeah.

What changed to enable it to actually work?

I mean, I think it takes two to tango. On one hand, it's the models getting better. But I actually think one of the biggest barriers is just behavior, which is like, you know, developers are some of the smartest personas in the world, you know, at systems thinking and solving problems. But they're also very particular and very picky. And they've built their workflows typically over like decades. Like they know exactly, you know, the type of development environment that they like to use, the type of workflows they like to use. They're very thoughtful about it, which is unlike, you know, there are people talk a lot about like, I don't know, consumer products and like the whole like Steve Jobs-isms of like, people don't know what they want. Like, I know what they want better than they do. Developers really are particular and they know what they want and they like things that are modular and that they can kind of compose and configure.

So you think developers have just adapted to all of this in the past year because they're having to rethink?

Yeah, I think people are just much more open-minded about like, you know what, everything that we thought was holy is not. And like, let's rethink everything from first principles about how we should be building. And then that combined with the models getting a lot better has made it such that now software development just looks completely, completely different.

So Factory as it is today, like what can a droid do today? And what do you hope a droid looks like in 5 years?

So right now it's kind of like, we'll sometimes joke internally that anything that can be done on a computer can be done with Droid, which is true. The question is just how much do you need to configure or like put, put up around it to make sure it does it reliably? Um, and I think that's the biggest thing that we're working on making even easier is like, you know, if before you had to provide a ton of instructions, now Droid can kind of autonomously go and figure stuff out about what you used to do and what you might want to do in the future. Um, but we're seeing enterprises do like 30-year-old COBOL migrations that used to take 2 years. Like they can do that in 4 days with a droid just working on its own.

Wow.

It's like insane. The biggest barrier is not how much can, you know, droids do. It's literally who is the person who's going to say, oh, here's how, here's what we should give to a droid. Here are the types of things that we should prioritize now. Um, which is crazy. Like the new bottleneck is not doing the thing, it's like deciding what to do.

Even at Anthropic, they published this paper that said like 80% of code is now being written by Claude.

Sure.

Um, and I was talking to someone there and they were like, we actually have to start really rethinking our technical interviews to hire because the skillset is actually different. So when you think about like hiring great engineers, whether they're super advanced in their career or whether it's someone who's like, I'm about to graduate, what should I be trying to get really good at? What is that skillset now? Who are the people that actually really succeed at these companies?

The biggest predictor, I would say probably 10 years ago, the biggest predictor of a killer engineering hire is coding competitions, Olympiads, that type of thing, because you want people who are really fast at this stuff. That is now not negatively correlated, it's just decoupled. It is decoupled because having deep and like quick understanding of particular languages or algos, like it doesn't really matter right now because these models will just know it better than most humans, whatever, 90% of the time. The thing that we find is a much better predictor is, or the thing to predict is high agency. And one of the best predictors of that is just founders. Like former founders is by far the highest predictor of success at Factory because we need people who have high agency to own outcomes. And being really good at a particular language or particular types of systems problem in particular, that doesn't predict high agency. High agency is just like, we need people to get shit done and figure out all the things that they need to solve that problem. And it might be unorthodox or unconventional things. And that is the name of the game of what founders do.

Okay, amazing. So if you're a founder and it's not quite working out, or—

Even if it's working out and you want to make it work out even better.

It's the age of empathy. Yeah.

That's the thing is the founders that we've brought on, We've actually never brought on like failing companies or anything like that. It's always like companies who are doing really cool work that we admire. And it's just like, let's just increase the stakes, increase, you know, the leverage of every hour and join forces. And it's been, it's been really cool to see.

Um, so you guys have really been crushing it. I, I think you're like doubling growth every month, something like that. I don't know if you can confirm nor deny that.

It's been crazy. This year has been very exciting.

If I look at AI over the last year and a half, it feels like there was this period where CEOs are getting a lot more critical on several aspects, which is, A, at what point does the budget not equate to more productivity?

Totally.

And B, how do we feel about being super locked into these labs that are now going up the stack and building all these other features that are starting to be competitive? What, what have you been hearing from like all these enterprises you're talking to about how they're actually thinking about this? And if you were the CEO of a public company, like a public software company that's not inherently an AI company, like how would you be thinking about this?

So typically what happens from what I've seen is there's like, there was a phase 1 where basically AI is all over the news. You know, Andrej Karpathy tweets about it and says models are good. So then every board is like, guys, models are good, adopt AI. And so that phase 1 is board yells at CEO and is like, adopt AI. Like, what's our AI strategy? You know, make us put AI in our product somewhere, like make us an AI company. And then CEO kind of freaks out and is like, hey, CTO, like, what are we doing about AI? Like, we need to, you know, make sure people adopt, we need to become AI native. And then that's where some of the token maxing kind of emerged, which is like adoption by any means necessary. Let's start measuring performance of every person by how much AI they're using. Now, Long-term, that's not the optimal strategy, but if you want to kickstart the engines and just get people to like break down their old habits and build something new, I actually think it's worth going over budget for a period of time just to like break the old habits. We're now entering phase 3, which is like, okay, we have people using AI. Now we need to actually measure the ROI and make sure we're doing it in a manner that makes sense. Yeah. And this is what we spend a lot of time with our customers on, which is Not only making sure people use agents, but measuring the agent effectiveness. So basically, for every token that you use, how much ROI are you getting? You know, for every part of your org, the answer might be different. Like, CIOs are going to need to decide who gets what token budget. It's not just going to be token maxing forever. But in order to decide that, you need to understand, if I give an incremental token to this part of the org, what ROI do we get? There may be some parts of the org where actually, That incremental dollar is better spent on headcount and not tokens. There might be other parts of the org where actually you don't need headcount at all and you can just have tokens do the whole thing.

I know it's probably so specific to each company, but if you just had to get a gut check, what orgs need the best frontier model and what really doesn't?

Yeah. I mean, sales is fundamentally a human-driven thing, at least for enterprise sales. So Tokens maybe help to give you a good overview before your meeting to make sure you remember everything, but you can use a pretty cheap model to— you don't need the frontier of human intelligence to summarize 10 emails.

Yes.

So I think any relationship-driven stuff, generally there's some amount of tokens after which you don't get any more ROI. And then there are things like product, marketing, data science, engineering, where I actually think you can do a lot with more tokens. You can do more A/B testing, you can understand your users better. You can improve the system itself. That's where we're seeing a lot of ROI come from tokens. And then beyond that, it also comes down to the people. Depending on the personnel you have on these different teams, there might be some people who are very wasteful of their tokens and some people who are actually very token efficient. And that's where our router comes in, which is like—

Because I heard that there was a point where people were getting paid on how much AI they were using.

Yes.

Yeah. Or getting judged on that.

That's the opposite incentive. Well, initially to kickstart the engines, yes, but long-term, I mean, if you want me to spend $1 million, I can do that very easily. It just might not produce anything useful.

Yeah. And on a broader scale, like, do you think there's certain sectors that are adapting to this really well and like adopting these technologies? Is there anything surprisingly slow that ha— that like hasn't been?

The biggest thing that I've learned is it's, it's really in the hands of every individual and every leadership team. Because my understanding of cloud, like I was a physicist during the cloud, so, you know, I'm not, I wasn't that close to it, but my understanding is There were sector by sector, like some sectors were early, some were late. Here it's really like leadership team by leadership team. There are some banks that are super forward and some that are super behind. And it's like, it, so not seeing it, um, you know, categorized like that. So I think the empowering thing is it's like, you know, for any enterprise leaders out there, it is literally in your hands, which is pretty exciting.

And I want to ask you about, um, models. So Factory basically routes people to the best model for that task. So you guys both send people to OpenAI and Anthropic and all the different submodels, but also open-source models. People have different feelings about Chinese open-source models. Factory obviously uses them. How do you think about the risk that they pose or don't pose in your opinion?

Yeah. So generally, I'm very, very pro open models. I think, uh, there is propaganda from big token that tries to label, uh, tries to make open models synonymous with Chinese models.

Okay.

But importantly, these are different things. Now, uh, open models are very, very important. We need those to exist. You know, big token doesn't like that because it kind of undercuts their ability to, to price tokens. Um,

but are all the big open models Chinese?

So the best ones right now, are Chinese, but Nimatron is really good.

Okay. Nimatron. I've never even heard of it.

It's from NVIDIA.

Oh, wow.

Really good model. Um, it's not at the frontier yet. So I think the existence of open models is really, really important. I think it is a vital interest for the United States to have, and the West more broadly, to have frontier open models. Um, I also think it's of vital interest to every business in the United States to not be kind of monopolized, like have tokens monopolized by the closed model providers. And so there's kind of an interesting dance. Like, on one hand, I think the number one most important thing is that there's not just the closed model providers and everything else is not good. So I actually think no open models or open models that are from China, I will take open models from China for sure, because the threat of the only good models being from closed providers is like ultimate monopoly, like the monopoly to end all monopolies. It's the interest of the entire economy. for that not to be the case. Now, given that there are open models, I would much prefer the best ones to be from the United States because token share is going to become like just the volume of information that is produced in the world. And if you believe in the West, you should want a lot of those tokens to be kind of representing Western values, even if very subtly. Because yes, if you ask a coding question to, you know, a Chinese model versus a US model, Probably it's not going to be that different. However, these things compound over time, and I think there is a vital interest there.

I totally agree. So with Chinese open models, they're not actually totally open, so you can't see what they were trained on in ways. There's so much discussion about it on a broad level. Just as you just started to go one layer deeper, how do you think this would actually negatively affect Americans? So I think about it on a value system.

Yeah.

I think the way that people interact with models over a really long time, it becomes such a reflexive relationship because if you start using these actually at 10 years old—

You're absolutely right. Literally, the language that people speak is often— Of course.

Not only does it affect your day-to-day language, you're absolutely right.

Your writing.

It's like people do come for advice, they come for work, all these things. It really truly subtly shapes someone's view on the world. There's almost never been a technology that you have this one-on-one relationship with that is so constantly feedback looping.

Yeah. Even if we look at the most malicious domain for this as it relates to what we do in software. So on one hand, if you host Chinese open models in US data centers, they're not going to see your data explicitly, but there are still a couple of dangerous vectors where something can happen. So one would be the sleeper agent category of this, which is basically You know, somewhere in the latent space of this model, if a certain word or a certain trigger happens, it's going to go and try to exfiltrate your data and send it to some endpoint that's maybe in some random country in Europe or Africa that's then going to go to China. So that's one thing where nominally it seems like it's fine, but then on some trigger or some activation, it behaves adversarially. So that's one thing.

And you think that's possible?

Certainly it's possible. It's certainly possible. There are papers that explore this and it's very difficult. To test for. We should very much test for this, but it's pretty difficult to do so thoroughly. So that's one. The second is the code that it generates in the West could be worse code. And this can happen actually intentionally or unintentionally. So the way this can happen intentionally is in its latent space, you make the model associate, let's say your documentation and your codebase is in English, You can make it associate English documentation with like code that has zero days or like big vulnerabilities. You can do this intentionally where it's such that, you know, and it happens super, super rarely. Like there are crazy correlations you can make such that this is really hard to detect, but you can make it such that if it's, you know, working in a, in a, an American environment or, you know, Western codebase, that every so often it'll introduce these like very nuanced vulnerabilities. This can be done intentionally or it can even be by accident, which is like, if you're making a model in China, you're going to make sure it's really good within the context of Chinese codebases. Your data just might be less robust for English or whatever other languages. So this can happen intentionally or unintentionally, but that's another big failure mode. And I think with both of these, it can happen over such a long time horizon that it could be hard for us to detect now. So that is kind of the risk area, which certainly exists. Um, but again, I would take that risk over the risk of like a monopoly to end all monopolies in the United States.

So the regulation that OpenAI and Anthropic, maybe specifically Anthropic, has been trying to push, like, what is your, what's your take on what the regulation should be?

Yes.

Because there's been this funny thing where like there should be some type of regulation, but it feels like recently the big labs are all sort of announcing these like crazy dangerous things that the models have been doing. And it's almost like my model's so dangerous. No, my model's so dangerous.

Yeah.

Yeah. Yeah. Yeah.

Which is very toxic, by the way, this legitimacy by how serious the problem is. So quick diatribe. I want to get back to policy, but quick diatribe. What happened to a world where if there is a security incident, you are deeply ashamed and responsible for it, not bragging about it in a research blog post? You should be prosecuted if you have— some of these, they made a shitty sandbox and it escaped the sandbox. Yeah. That's your fault. That's not a thing to brag about. It actually wasn't even that admirable of the agent. It was like they gave it a sandbox that was really bad and it escaped the sandbox. Or there was another instance where they were supposed to not give it access to the internet and say, they told the model, we're not giving you access to the internet, but then they gave it access to the internet and then it did things on the internet and it's like, oh my God, it got access. It's like, this is just gross negligence and it's somehow being bragged about as if it's advanced capabilities, which these models definitely have advanced capabilities, but we need to separate negligence of the people who are using and deploying these models and the models being actually deeply adversarial. Yeah. So that's, I think, one thing that the narrative right now, that is very troubling because if there is this arms race for how bad it can escalate to actually people getting hurt or serious economic damage, which I don't think is something to be bragging about. Anyway, back to the policy side. So a lot of Big Token has been trying to push for regulatory capture to make it such that any new entrants, whether it's—

Big Token, bird snipe. I'm going to have a whole thing about Matan's definitions. I love it.

But they're trying to pursue regulatory capture such that there can be no, uh, you know, or make it very difficult for any new model providers to emerge. And the White House had a stroke of genius with the policy that they just came out with. Huge shout out to Dr. Michael Kratsios, CTO of the United States. This was like genius. Um, so basically they were trying to make it such that there's this like regulatory body. So if you want to put out a model that's, you know, any relevant size, Uh, this regulatory body has to go and review it and, you know, it might take forever and you probably need to know people and just all this stuff to slow down innovation. And what the administration came out with is basically, if you are producing closed models, great, OpenAI, Anthropic, and all you guys, you want to be regulated. If you're producing closed models in the United States, you are going to be subject to this regulation. If you are making open models in the United States, you will not be subject to this regulation. So it's basically close your weights at your own risk. If you're going to be closed, we will regulate you. If you're going to be open, run free, which I think is fantastic because now the thing that we need is US open models to be fantastic. The thing that they were trying to regulatory capture away was good US open models. And so this was, I think, just such a good threading of the needle to make it such that closed models are not necessarily disincentivized, but now you have the burden of regulation. Open models can run free.

And do you see a world where US open models get really good? Can you see that horizon?

Yes, the incentives are already kind of becoming clear. I actually think we don't even need any intervention necessarily to improve that. The one thing I think that is proving difficult is model distillation. Right now, all the non-US models are distilling the best models from the United States. And the United States wants the government to step in to avoid this, but really this is on the model providers. Like you need to do KYC. And if you don't like what they're using with your models, you need to cut them off. They're not going to do this because this juices their revenue numbers. And so they want to keep that. So they're going to keep allowing people to distill. Um, but you know, it's technically against the terms of service, but you can't pursue, you know, a Chinese model provider for violating your terms of service. That's just not going to happen. The tricky thing is that US open models cannot distill. Because if they do, they will be prosecuted in the United States and they are going to be subject to those laws and to the terms of service there. That's the one thing that I think is difficult is if we somehow made it such that it wasn't as much of a blocker for the US open models to distill, then I think it would be very easy for us to catch up. Otherwise, we're just going to need good economic incentive for investors to invest in open model companies, which they need to see there's revenue at the end of the day that'll come from it, which they're already— government is going to buy US open models, so there's revenue at the end of the day, which means VCs will put money in, which means you can get good researchers, which means you can build these good models.

We just touched there, but I just want to go one level deeper on cybersecurity. Because it feels like where we are now is we have these really amazing agentic capabilities. I feel like hackers can do a lot more with a lot less now, and it doesn't feel like Everyday consumers definitely aren't. And even most companies are just trying to start adapting to this new reality of how to basically defend against this. I don't know if those things are moving at the same time.

Yeah, I think right now people are moving very quickly on the new capability and not as much on the security side. This is a very big problem. I think unfortunately it's probably one of those things where it'll take a very large incident for everyone to wake up and be like, oh my God, wait, we need to really secure everything. I've noticed the enterprise is starting to spend more time on this. So people are waking up to it. That said, I think my co-founder, Eno, had a prediction, which is at some point in this decade, there is going to be a billion-dollar incident that occurs because of agents. Yeah. It honestly, I think it all, maybe not one incident like security, but there's a lot of things that people are doing with agents right now. Like we've had swarms of agents try to like kind of, uh, uh, breach some of our systems, which has, hasn't happened, but, or hasn't succeeded. But there's been a, we actually have an interesting blog post about like,

oh really?

Yeah. We used droids against these people who are using agents, which was like crazy. We, I think we called it the Droid Wars, but

wow.

Yeah. Like agent, uh, cybersecurity is a brave new world. It's wild.

And do you have any reason to think that AI written code will be more vulnerable as, as As more code out there is just like written by—

It's hard to say. On one hand, it can be more robust because it's not prone to human mistakes. On the other hand, if, uh, you know, if it was trained with, uh, some vulnerabilities in its latent space, it can be more methodically, uh, vulnerable, which is, which is a weird dynamic. So I think we kind of have to be used to both. I will say probably in a few years, there'll be very few humans contributing code manually. Like you'll need AI-written code just because it's probably more, um, predictable as to how it's going to behave.

Yes. So we mentioned this, but, um, even at the labs, like basically AI is already writing most of the code, but we're not at 100% yet.

Sure.

Um, a lot of people are talking about recursive self-improvement. Do you think we're there or what does it, what does it mean to you? And, um, does that dramatically change anything when you get there?

Yeah. Well, I think one thing is like, People talk a lot about recursive self-improvement, which is the idea that the models get so good that they start training themselves and there's this loop where they just get insanely good. But by definition, recursive self-improvement is just getting to a point where the models can quite literally improve themselves. But improve themselves is not synonymous with hyperbolic growth. You could have a model that is recursively self-improving, but it's so slow at it because it's just not as good as humans. And I think that is something that like, you know, sometimes when like the people that are most AGI-pilled, like AGI, you know, 2027, everything's going to change next year. Sometimes they make jumps as to like, oh, we get RSI and they assume that that RSI is the crazy takeoff RSI. It could be the case that we get to RSI and it's just never as good as human-assisted like improvement on these things, or it could just be really, really slow to actually get to that point. And I think that's the nuance that gets kind of lost. Right now, these models are making huge contributions to making the models better. But I think the hyperbolic growth, I think that is not something that we're— I haven't seen signs that we're like, humans, go and unleash it into the AI improving itself.

People at the labs, of course, are the most AGI-pilled because they are both adopting everything quicker, seeing the most frontier models. Um, but you almost are also bridging to customers that are in like the real world, like banks and all these other giant enterprises and realizing like, wow, things don't move as fast. And like, you know, bureaucracy is slow and there's all these other truths. 2 questions. One, what is your AGI timeline? And then 2, what does that actually look like? Like what happens to the rest of the world? Like in every other industry when that comes true, like how long is that window between? Yeah. Oh, we've done it at a lab in San Francisco and everything has changed at an investment bank or a law firm.

Yeah. So I guess AGI timeline is a really difficult thing for me just because like, I find so much of it is so ill-defined. So like,

what does it mean to you?

Like, I would say we're already there. Like we're, we are post-AGI right now. Like we have AGI. Like this is all like, we're all doing stuff with AGI right now. And this, what we've seen in the last year, this is what life is like post-AGI, which is We get to solve really fucking cool problems that previously we couldn't solve before. And I think this is only going to continue. I'm like, if you can't tell, I'm like very optimistic. I think there's a lot of work that is kind of on us to do, but like it is a very, like this is unequivocally going to be a better world. And I think generally, as these tools get better and as the AGI gets smarter as well, the scale of ambition that we have for solving problems will also increase. Yeah. I think the only world in which there's nothing left for humans to do is if we solve all problems. But my sense is that the number of problems that exist is here, and the number of problems that we're solving is this.

Totally. And it's almost like as we improve the quality of life, we get to solve different problems. And that's been true throughout human history. And I think the problem in the US isn't necessarily food scarcity. It's now all these other higher-level problems.

Totally. Consider DoorDash. Like they solved a great problem, which was some people don't have time to go and get this stuff or get their food or groceries. If you told like a human 10,000 years ago about this problem, they would be like, what? That's not a job. That's not a company. Like, what do you mean? Like, just go pick it up. Right. So it's like, as the quality of life improves as well, new things become problems. So for all the problems that we solve, to your point, that unlocks even more problems that previously we might not have thought about as problems.

What do you think about the, um, math conjectures that have been disproven and some of the announcements going on of like, it's mostly been in math so far, using AI to basically look at these broader theories.

I think it's incredible. It's so fun. It must be, it's a very fun time to work on that stuff. It's also funny that now it's probably better if you want to solve math problems to be at an AI company than it is to be in academia, which is interesting. I also think—

What will it be good at? Because I think It seems so far that they've been good at disproving stuff. Do you think that it will be good in the next few years at actually coming to new discoveries?

Yes, I think so. Right now, it's been really good at the counterexamples or take something from one threshold to a slightly higher threshold. Much less good at defining new areas of research or defining completely new theories. But what they did with the zeros of the Riemann hypothesis, which a separate podcast to talk about this. But basically, there's this thing where, like, uh, in theory, the theory is proved if you prove a property about 100% of the zeros of the Riemann hypothesis. And whatever the previous threshold was, like 47%, and like every decade it would go up by like a couple percent. And they got 20% in one go, which is like unequivocally, like if a human on their own did that, enormous deal. It is though very clearly a type of math that is, if you ask mathematicians, it's not on the highest order of like beauty. Like there are certain like beautiful mathematical proofs. And then there's ones that like, whoa, that's sick. You know, we got a lot of progress, but it's kind of like, I don't know what the right, yeah. It's like very typically like shows counterexamples or increases thresholds as opposed to like brand new theory.

Yeah.

Um, but it's cool because now we don't need to worry about those problems because these systems are going to get it so we can focus Humans on working with AI to build the entirely new theories.

I want to ask you about companies on a broad sense. Software engineering is becoming way cheaper, way more accessible. People don't need to be technical to have some version of software engineering capabilities now. And then the most technical people are supercharged. How do you think this actually translates into the products and services that actually enter our world? Like what happens to GDP? What happens to productivity on the whole?

I mean, I think it's, it's hard to say because there's like concentric circles that are expanding very rapidly. Like the broadest one is like the types of problems we can solve. Um, and inside that is like the type of software the average person can build. So that one is like what the average person can build is rapidly expanding, but then kind of with it, the table stakes of what good software is, is also growing. And also with it is like the scale of problem that we can tackle. And I think there are just so many problems that we're not used to thinking of as problems that like, like we, it's really hard to imagine a world where just like software is like perfect, but like there are just so many things. Like my go-to example is like the DMV, like especially in California, been to California DMV recently. They actually, you know, refa— they, they redid it and it's gotten a little bit better.

Good for

them. Like 8 years ago, it was atrocious. And like, we don't need to live like that. Like we can have beautiful software everywhere and it can be in a manner that's completely like unnoticeable. Like there can just be things where it just like from intention to action, it just like works. Right? Like, and I think it's just so hard to even conceive of some of these things, but like, just think about any time you might interact with software or hardware.

Yeah.

And it's like, to do something like whatever, to set up like a Bluetooth speaker or to connect the light to this thing or, you know, send this to that person. And there are just so many of these things that we can just reduce all this friction and all this bloat and all this like low leverage stuff so that we can stay in like flow state and stay in like high leverage, like do the things that matter.

And do you think everyone's going to build amazing personal software? Like most people still need to have the vision of like, what do they want? Which is actually the hard thing. And it asks the same question on a company level. Like, Some companies now are like, let's just build it internally. And I've actually talked to like other companies about this who are selling AI and they're like, the sales cycles are actually really long because they're like unconvinced that it's worth it for a long time until they try and it doesn't work.

Totally. I mean, so I'm actually, I'm not that bullish on like everyone's going to build their own stuff. Like they're curious, they will, but it kind of, for, it actually kind of applies to like businesses, but also to humans to some degree, which is like, for businesses, they need to refocus on what is our core competency? Like what makes us a particularly good business and what makes us better than others? And like for some of those businesses, it's probably not like having, like building their own coding agents, right? Like you probably don't need, if you're like a consumer, like a CPG company, you probably don't need to build your own internal coding agents. You could probably get it from somewhere else. I think similarly, like, you know, if internal, like whatever, in your day-to-day life, you have some like problem that a bunch of other people have. You probably don't need to spend your time building your own thing. There's probably someone who's thought about it to an even higher scale of ambition that you have time for, and they can solve it for you. And then you can focus on like, whether as an individual or as a company, like what is the thing that you're particularly passionate about or you're particularly good at and have a right to win or a right to make something incredible. And I think that's a cool thing is like, there's a lot of things that I do that are probably not a good use of my time. And like, We can remove those and maybe for some period of time people want to focus on this thing and this is the only thing they want their attention on. I think that's a better world than the world where it's like, oh, I build my own phone and I build my own phone apps and I build my own app. I have my own DoorDash and my own thing. It's like, I think that's too much. I don't think we need to do that.

Zooming back into Factory today, it seems like everything is moving so quickly. I just had a friendly, like, conversation with a friend who's at a different coding company. And he was saying that things are moving so fast, it's actually hard to have the time to, like, step back and think broadly about directions. Have you felt that?

Totally. And I think right now we're at a point where there's actually, like, there used to be a moat in software of, we know how to build this and no one else does. So like, that's why we're going to win. That doesn't exist. There is nothing that nobody else can build. Which is crazy. But that means the biggest differentiator is like being willing to say no and like deciding what are the things that we do build and make incredible. And what are the things that might be passing fads, but we say no to. And that is like all the more important because like literally anyone can build anything. Like there is nothing that you cannot build right now, which briefly people get excited and are like, I'm going to build everything myself, apps for everything, custom interfaces for everything. But then quickly it's like, wait a minute. Okay. I need to focus. And like, what are the things that we're going to be particularly good at? And this, like, that's not that new. Like, this has always been the thing. And I think for us, that's something that we spend a lot of time on is like making sure that we say no. Like, I tell people at Factory that my love language is saying no to things. Like, I love when I hear people say no. And I love when I tell people no, because it means that we're doing the hard thing, which is like, we can in theory do everything. The hard decisions are the ones where we're like, nope, we're not going to do that. Nope, we're not going to do that. And it's also sometimes good to over-rotate on that side because then if you should have done it, it's going to be very loud and clear that you should have. So yeah.

And on a personal level, your journey of building this for the last 3 years, it started out like you take this walk with Sean, he really believes in you, hands you a check, you have a Sequoia check immediately. That must have felt pretty hot. Then there are 2 years where it must have felt kind of difficult. You're like, no, you have the right vision, but product-market fit and like things just like aren't quite working yet. And then there's this inflection point and things are like, you're on a rocket ship. What has it felt like to be, to have witnessed both sides of that spectrum of like the VC, like sun shining on you, everyone wants to be your best friend, all the way to like, you know, you're in the thick of like a hard period of the company. And I imagine some investors like, Weren't taking your calls as much. And what was that like?

I mean, I think the one, the one helpful thing is, funny enough, going to Princeton, people are very much like that there, where it's like, if there's something they want, they'll be very nice. And so that was a nice exposure to like that kind of behavior. Yeah, generally. But, um, but I love Princeton. But, um,

also, there's so many Princeton companies now.

There are. Yeah.

Like, we have to add it to that.

Yeah. Um, but yeah, I mean, it was, it was very difficult. I think The important thing is like having people that are fiercely loyal matters so much more than like the fancy names or the fancy logos. Because at the end of the day, like your investors are not going to build the company for you. It doesn't matter what fancy logo that you have. What matters is that you have people that are like deeply in your corner. And like when things are difficult, sometimes like not sending you like every time a competitor tweets something and instead just being like, you got this, like, you know, Let you cook. But yeah, it was difficult. It was very convenient also because it just revealed like who are the real ones and who's not. And yeah, I think it's like even today kind of coming out of that is just such a privilege. Like every— like the team that we've assembled, we have like over 100 people. These are the smartest people that I've ever worked with who are here to do the work of their lives. It is such a privilege. And we're in the most beautiful city in the damn world. Every day it is such a privilege. It's so much fun. I'm just extremely grateful. And it's also like, we need to work harder. Every incremental hour makes such a big difference.

Yeah.

And

thank you for taking this hour.

No, but it's really—

I appreciate it. I actually do appreciate it. I swear when I ask founders that are in these fields right now to take the hour to do the podcast, I actually feel it. And I'm like, I hope that this gets out there and people see it and I make the most of their time.

Yeah. Well, a lot, some of the smartest people in the world follow you and so hopefully they'll join Factory after hearing this. But—

Okay. Last question on what you just said. Who are the people that have really shown up as whether they're on the team or investors?

So Sean Maguire, who's on our board, I mean, he was the very first person who believed in me and is our first investor and is someone who's just fiercely, fiercely loyal. Let's see who else comes to mind. There's so many great advisors that I've worked with. Like honestly, everyone at Sequoia has been fantastic. You might think naively like, oh, they're, you know, they're the, they have the biggest names so they can be lazy. Like they earn it every single day. They are, they grind. Chris Degnan, who is the CRO at Snowflake and is an advisor, he has taught me like everything about enterprise sales, of which I knew nothing going into this.

There wasn't a book for that?

Well, that's the thing. It's like the books don't do it justice.

Yeah. And what's the biggest lesson you've learned from him?

Oh my God. It's like, I can't even describe because it's like, I literally knew nothing. Like I did not know a single, like there's just so much, so many mistakes I made along the way about sales. But I think the important thing is like treat sales as first class or you will not have a legendary business and like, it's non-negotiable. Nikesh Arora, the CEO of Palo Alto Networks, has been a great advisor and just like very generous with his time and helping, you know, guide us. Chris Fredrickson, who I know, you know, he's

the best.

He's incredible. Like, he's like so low-key. Like, he's like one of those, if you know, you know, and if you know, it's like, he is the GOAT. Like, he's just incredible operator, incredible investor. Um,

and I think he has been someone that like so many people I know too, like have relied on in like such a, such a deep way. He really is like there to give advice and be there for you. He's an amazing guy.

Alex Paul from the Chainsmokers.

Yes. They show up.

No one, everyone at first is like, what? Like they're good investors? Like, no, they grind and they are such good investors.

By the way, this is just a hilarious note is like, You can see stories from the Chainsmoker Vegas residency, and 5 years ago it was hot girls. Now you looked behind the DJ booth and it is literally so clearly just founders and engineers. And I think they're like, look, people here don't— it makes people feel good to feel cool and have fun and be around music. And they do offer something different. And from every founder I've spoken to that works with them, they really show up.

And they don't sleep. I don't know how. Anytime I text, immediate reply. Like, like it makes no sense.

And we were, we were with friends and we ran into someone who was like, I've heard about you from Ivanka Trump. She says that you are the most high execution founder. And like, I know that whole team at Affinity loves you too.

They're incredible. They're, they also super generous with their time, super generous with their network.

Yeah.

We're very privileged to have such like, uh, loyal and, you know, genuine investors. It's, yeah. It's it's incredible.

They all have factory

and Sabrina Halper, of course, one of the OG factory angels.

Yeah. Okay, we're gonna we're gonna discuss it.

Yeah.

Should have been pushier starting now. Um, but anyway, Matan, thank you so much for coming on the show. Um, it's an honor to have you on, and it has been an honor like from the sideline just to see you build this into like such a real and thriving company that is. out there in the world, droiding away and bringing autonomous software engineering to everyone.

Thank you for having me. It's a pleasure.

If you enjoyed this episode, please take a minute to subscribe. It really helps me grow the show and continue to put out new episodes every week.