Factory AI launches Droids — autonomous software agents for enterprise codebases — after founder dropped out of Berkeley physics PhD

you kick us off with a little introduction on yourself and also uh the company and the news? I know that there's big news coming up. Yeah, absolutely. So, uh I'm Maton, CEO at Factory. Uh here to share a bit about our latest launch. Uh we released Droids. Um droids are not just your regular everyday coding agent.

They are full endtoend software development life cycle agents built for the enterprise. Big enterprise, not just not just startups. Um although you know we've uh in the last 24 hours seen a pretty big explosion of usage with um startups which has been really cool. Um yeah that's uh it's been it's been it's been fun.

Talk to me about the evolution of AI agents and coding agents. I remember there was a company a couple years ago that was training their own foundation model. Then the regime seemed to to seem to shift to okay we're not going to do any pre-training on code specifically.

the foundation model companies got that covered, but we will do a bunch of post- training. Then it became more maybe there's some RL fine-tuning. Maybe there's just some some reasoning that we're doing.

Now it feels like a lot of the coding agent companies are just kind of like we're we're rappers, but we're still printing money and rappers and unnecessary porative. It's actually the best way to build this business and it's helping us.

So So where do you fall and and and what is your take on the different I I just got here. Here I got to compliment Matana. I know. I know. I know. Looking fantastic. Thank you for when you when you come to the temple. When you come to the temple, you must be dressed appropriately. Yes. Yes. Thank you. Anyway, but yeah.

Uh yeah. No, great question. I think uh it's been interesting how in the last two years there have been so many trends within uh the kind of like sub sector of uh AI for software development. Y u you're spot on.

there was a really big trend especially if you measure it by the capital that was put into it in terms of uh companies coming out there saying that they're going to train models in particular for code or fine-tune models for code there's been I mean there's been quite a few who raised to the tune of like half a billion dollars to train their own models and I think something that um uh a lot of people were saying at the time or maybe whispering because they didn't want to offend those who put half a billion dollars in was uh code is a core competency for for the foundation models they will not allow a startup to you know fine-tune their way to having the best models as you see in in general with code it's like alternating between open AI anthropic Google um XAI in terms of the number one spot in code and so I think it I guess it took a lot more money than it should have but I think people come to the realization that fine-tuning an RL for code specific models will be won by the foundation models um is that just a function of like of Code is the internet.

The internet is code. Code is on the internet. And so if you train an LLM on the internet, like you're going to learn like like out of the box, if you do a robust largecale training run on web text, you're going to get Reddit answers and you're also going to get a pretty good model that can code.

And so unless OpenAI says, "Let's not train it on code," it just you're going to lose to them because they're putting the most, you know, energy, the the most cycles into training the big model. And so you just get that as a byproduct. Yeah, that's that's spot on.

It's actually I would go even further to say that um and I know uh there were papers on this in around like 2022.

I'm sure there have been since but yeah there is a direct correlation between how good a model is at code and how good it is at other general purpose tasks that makes sense so it is just it is like it is you know table stakes for the foundation models to be the best at like raw coding um so exactly right there yeah so how do you how do you build a business without getting rolled yeah how how have you how have you guys navigated just partnerships with the labs in general where it's you're kind of fremies you know it's like we can work together, but this is a good question.

At some point, somebody's going to try to kill the other one. Yeah. No, this this is this is a really good question and it's very top of mind.

I mean, I think the reality is the closer you are to the zero to one or like the vibe coding or the like less technical your audience and the smaller the company, um that's really where uh the foundation models are like about to sweep it up or or already have.

So it's like you know building apps from scratch, building websites from scratch that's something where the model providers will basically get it for free like plus or minus some deployment details in terms of you remember with with uh prehat GBT there were a bunch of players that would generate copy for you based on the open openai API and they had just rock their revenue rocketed.

Yep. And I remember it was the same sort of like people like the same behavior of like posting screenshots of like the revenue dashboards and then those screenshot shots stopped being shared postg cuz all that revenue you know just went away as as Shakespeare said these violent delights have violent ends.

Um there was a really there was a really great great tweet I saw the other day. I can't remember who posted it but it was uh you know a lot of people are talking about these like record-breaking like runs to like you know large ARRS. Yep.

Um and what's going to be coming soon is some record-breaking churn for a lot of this like monthly these monthly subscriptions. Um so that'll certainly be interesting to see.

Um but yeah, so that's the part of the spectrum where uh the foundation models are best poised to tackle is that like 0ero to one um less technical use case. For us, we've been focused on the enterprise from day one because one, it's really not sexy and so it's not like conducive to like viral demos and all that hype.

But that's where a majority of developers are getting a majority of their pain. Like dealing with like cobalt, dealing with these like 20-year-old code bases, migrations, refactors. And in order to handle that well, you can't just like oneot with like a chat GPT or a codeex or a cloud.

You need to have deep integrations with their codebase. Oftentimes it'll be like multi-reo integrations. Um you'll need to understand not just where the code is now. Nice. Love the little thumbs up guy. uh not just uh not just where the code is now but how it got there, what the best practices of that org are.

Um integrating with tools like Jira, Google Drive, Century, Data Dog.

Um the kind of first principles thinking there is that in order to produce the quality that an engineer who's been at that company for like 10 years would produce, you need to have access to the same information and that information sometimes requires some really ugly integrations.

um you need to uh kind of get the the workflows that are a little bit more specific to these large archaic orbs which is a little bit further a field um from what the foundation models are are are working on right now.

Talk to me about synchronicity asynchronous coding agents versus synchronous uh co-pilots versus autonomous agents. Where do you think it's going? Do these lines blur? Are they are they distinct product? It feels like OpenAI has three coding products now.

I can go to 03 and I don't even ask it to write code and it just does. Uh I imagine that the amount of code that's being written for non-technical people who don't even know that code would be useful in giving them an answer to a question is just skyrocketing right now uh in that product. Then there's Kodax.

They also have Windsurf now. And so you can imagine a few different bites of the apple there. Are is that uh are they are they indexing the market or are these distinct different markets? Will there be one power law outcome in the coding space in terms of time? Yeah. Yeah. Great question.

I think the the first order like bifurcation that'll happen is between the non-technical audience and technical audience. So like for nontechnical being able to spin up apps on the fly is like cool and fun and that'll continue to be something that's useful. They're never going to Yeah. Exactly right.

and they're never going to really be that concerned with like going too deep into the code itself. They'll just like I want a to-do list app. Make it for me. Okay, great. Um for the technical user, I think that's where it gets a lot more interesting.

Um basically, and maybe maybe this is just a quick like kind of philosophy that we have at factory. Every transformation shift has a very clear behavior change associated with it, right?

So like internet had people uh getting most of their information from like TV, newspapers, books to then like going on this console on their desk and like you know getting all their information there.

Um mobile had people you know walking around like heads up to now like you know walking around on Tik Tok subway servers all the time, right? These are very visceral obvious behavior changes. And yet everyone talks about AI as it's you know the the largest one to come. It's going to put these other ones to shame.

And yet if you look at the mostus product, the most used AI products right now, there is no new behavior. Like chat GPT perplexity, that's just Google with better results. Um the behav like the silhouette of that behavior doesn't actually look that different.

Similarly with a tool like copilot or cursor, the way that software development is is looking, the behavior is still the same. You're still in this IDE which was built for the world where humans wrote 100% of their code.

We're quickly going to a world where humans will write 0% of their code and our take is that their behavior will fundamentally have to shift then and not like in some iterative approach where you like iterate your way from an IDE to whatever this new thing is.

But our take is instead you need to build that from the ground up. And this is kind of a long-winded way of answering your question about the like async versus sync. The reality is in this future, developers are going to be natively working with agents.

And so they'll need to like dynamically adjust between if they're collaborating with an agent to like look through their codebase and understand how they should plan some new feature and then having a good plan for it and now like firing it off, delegating it to these to these agents.

Um, and at the same time it's going to put more emphasis on the testing because if you just shoot from the hip a ton of agents or in our case droids, um, then you're going to have a ton of code to review before you release it to production because ideally you're going to review it, right?

But that's kind of a depressing world where okay, we don't write any of our code, but now you just need to read like thousands and thousands of lines before you can ship anything.

So if you as the developer have better tests and you know, hey, if it passed these tests, I don't even need to review it because I like expressed all of the constraints that I had through these. So if it passes, great. Let's ship it.

There's kind of this new emphasis on testing um to kind of enable that more delegative workflow. What's the secret to avoiding churn in the enterprise? Are you trying to ink multi-year contracts upfront?

Are you just playing the same game as everyone else and getting a bunch of experimental budgets because money is money and you know that seems to be the meta right now.

But I imagine that you're you're you're thinking about this or at least like messaging to the community when a competitor gets to a you know a customer first.

Are you kind of hovering you know with the understanding that like hey you know maybe they're you know if we can really show that we're significantly better we could we could win. And is there is there like a little bit of a price war here?

Because you imagine that you come into an enterprise and you say, "Hey, this would cost you so much to do with like Accenture and we're going to do this this replatforming of your cobalt application to . NET or something or Python and uh and and you would spend 10 million on that. We'll do it for 8 million.

" Uh and that's all of a sudden all of that's all of a sudden that's like 90% margin for you instead of like 50% margin. Um, but then your competitor comes in and says like, "Well, we'll do it for seven million. We'll we'll do it for six million. We'll do it for five million.

" And so there's there it seems like there's some sort of price war dynamic that might happen. Walk me through all of that of like winning in the enterprise financially. Yeah. Yeah. That's that's a great question.

So, first of all, yes, we do year-long contracts just because it's important to have that mutual commitment and in particular because adopting the tools is not enough.

like there I cannot tell you how many like CIOS and CTO's I've spoken to who have adopted like the hot new AI IDE and then you ask them the question that they don't want to answer which is how many people are actually using it and it ends up being like 10 20% and of those 10 to 20% a lot of them are just using it like the idees of old totally so they're kind of like adopting these new tools patting themselves on the back being like look CEO we did it we adopted AI job's done give me the big bonus but the reality is is like they're actually not getting any productivity improvements.

And so part of why we do these longer term uh commitments is because one of our core competencies is not just having the best agents in the game, but also helping them. Look, I love it. I love it. There we go. Um we're also helping them adopt their behavior patterns.

Um because that's the thing is like you could have a tool that's 50% as good, but if you have twice the adoption, then you're now at parody.

And so I think it's just a lot less sexy because all the like you know great engineers who are coming out of Stanford want to work on spinning GPUs and all that stuff talking about like behavior change in the enterprise that's like yo like you know they don't want to think about that but that is where the ROI is going to come.

Um and also John to your question like the way we actually get in and do these deals is focusing on those deliverables about this was scoped out to be four months and we did it in two months or one month or two weeks that's ROI that actually matters to like the seauite when you talk about like we shipped tests 10% faster it's just so like or we 20% more lines of code it's just so amorphous and so not tied to real business outcomes that if we can come in and actually um tie things to things that are shipped per quarter or um you know pulling in dates for certain deliverables.

That's where it's just like it's so frictionless because it's not really existent elsewhere in the market. Uh I was texting with Adam and Ben from Genius, one of your investors, and they said to ask you about a fateful walk that you had with Shawn Maguire. Does that ring a bell? Yes, it does.

Um so this is uh this is in like the founding history of factory basically. Uh two years ago or two and a half years ago I was doing a PhD in theoretical physics at Berkeley um which is what I was doing prior to factory and uh you just phoned it in. You want to kind of take the easy path in school or Yeah. Yeah.

Exactly. Yeah. you know, just uh some fun string theory, which to be fair, it is really fun uh and beautiful, but uh decided it wasn't a path for me in particular because uh you know, to to be a good physicist, you kind of need to thrive in isolation, just like in your room alone like reading papers.

Um did that for like 10 years, was really stubborn.

It's kind of a long story, but ended up uh reaching out to this Sequoia partner, Shawn Magcguire, because he also used to be a string theorist and uh he ended up, you know, going into entrepreneurship, sold a company for a billion dollars, joined Sequoia, saw like a random podcast with him, and I was just like, I've never seen like another physicist who has somewhat social skills, like let me hit him up and get some get some life advice from him.

Ended up going on a walk together. on this walk. He said, "Matan, you need to drop out of your PhD and you should either join one of my portfolio companies uh and just like work on glue. Not a thumbs up. Let's go.

" Or you should join you should join X because Elon just took over and you'd have to be a badass to voluntarily go there. Or you should start a company. Um and so that was uh eight days later dropped out of the PhD and started factory. So Well, congratulations. Very cool. And uh yeah, thanks for stopping by.

This was fantastic. And uh thank you for giving agents a cooler name. I love droids. We got to get ourselves some droids. We got to get you guys some droids. Thank you guys. Great jam. Have a good one, Matan. Uh speaking of automation, let's tell you about Vanta.

Automate compliance, manage risk, prove trust continuously. Vanta's trust management platform takes the manual work out of your security and compliance process and replaces it with continuous automation whether you're pursuing your first framework or managing a complex program.

And next up, we have Johnny from Muan Space. We're talking about data centers with Chase Lock Miller in Abalene, Texas. Now we're going to space and we're putting the data centers in space. Very excited to talk to Johnny. So, welcome to the studio. Johnny, break it down for us.

And I'm going to let Jordy do the intro on this. And if you're on your phone, would you mind rotating it 90 degrees so it's widescreen? Thank you. There we go. You mind kicking us off? What's going on? Great to have you. Where are you calling in from?

Uh, randomly, I'm actually down in y'all's neck of the woods in Long Beach. I'm uh hanging out with a friend at BAS today. Nice. Nice. Um so yeah, why don't why don't you give a quick intro, background on the company, yourself, all that good stuff. Yeah, sure. So, Muon Space, we're a 150 person startup in the Bay Area.

Uh we're building a platform to deploy large numbers of satellites and constellation format uh for a lot of different mission types. Um we were founded in 2021. Um it's been a pretty pretty big rocket ship ride so far. It's been really fun.

Um, my background going back a ways, I mean, I I' I've been in space for, you know, well, longer than I care to admit, including I was an intern back at SpaceX in the very early days, like 2003 and four.

I was part of, uh, I was the chief engineer at Skybox Imaging, which was the kind of first venturebacked satellite company, kind of space before space got cool. Um, very cool. And so, uh, you know, taken a lot of the lessons learned and kind of things I've seen, uh, from those experiences to the new company. Awesome.

and then break down break down the new company. Yeah. So, I mean, I think the way to think about this is um you know, a lot's changed in the last decade of space. Um, traditionally, if you wanted to go do something in space, it required a lot of sophistication.

Um, you had to have it literally a team of rocket scientists. Um, you obvious um a lot of deep expertise in everything from, you know, ground stations to to launch to avionics software. um you know