Is AI Eating Software?

Speaker 1:
Welcome to TD Cowen Insights, a space that brings leading thinkers together to share insights and ideas shaping the world around us. Join us as we converse with the top minds who are influencing our global sectors.
Oliver Chen:
Investing in AI, welcome to the Retail Visionary Podcast Series, a podcast about visionary ideas and people. My name is Oliver Chen, I'm TD Cowen's new platforms, retail, and luxury analyst. We're excited to host Francois Chaubard. He's currently a partner at Y Combinator. He's an accomplished entrepreneur and technologist with a strong background from Stanford in engineering and AI, and he was the founder and CEO of Focal Systems and the author of AI for Retail. Francois, it's great to have you here.
Francois Chaubard:
Thank you for having me.
Oliver Chen:
So at Y Combinator where you're a partner, what are the biggest investment themes you're focused on from an AI perspective?
Francois Chaubard:
Yeah, we're looking at AI. AI is infiltrating all parts of the economy. And I think there was the software is eating the world. I think AI is eating software. And that's certainly been true over the last three years and probably been true over the last 10, just slower. And everyone has access to these huge, amazing frontier models. And the application layer is really not been built out yet.
So if you think about it, when Oracle Database came out, it was a huge step function, but you probably didn't feel it until there was a Siebel CRM or PeopleSoft or something, other applications to go build out. So there's a lot to be built out in the application layer on top of these frontier models. And I think that's where we have been really focused and I'm the most excited about. I think there's a lot to be done there.
Oliver Chen:
Well, for those who are less familiar, what do you mean about a frontier model versus an application model?
Francois Chaubard:
Yeah, so the frontier models are going to be your OpenAI, ChatGPT-5, your Qwen 3, your Kimi 2, Llama 3s, all those are frontier models. They're great models. Some of those are open source, some of their closed source. Some of those are APIs actually, they're not actually a model. It's maybe a whole bank of models. And then you can have access to them via an API.
Now, go build something cool about that, like an automatic way to fill out my form at the doctor or give me recommendations on what to cook for tonight. And like, yeah, I can go to OpenAI directly for this, but there's some things that make sense to have pulled out into an application that is automatically more customized. And Cursor is probably the best example of the mind of this where it made so much sense for Cursor to exist inside the IDE and not just be some chat app for code. And they're at [inaudible 00:02:59]. Yeah.
Oliver Chen:
Should we be most interested in frontier or application models? And what's going to happen with the evolution of frontier models?
Francois Chaubard:
Well, I think it's a symbiotic thing, right? Where you have the frontier models that are sitting, you have the applications that are sitting on top of the frontier. So the frontiers are a key component, that's the database. So being excited about the database is a great idea, but you still need the application layer there to do that last mile piece.
Oliver Chen:
Yeah. Okay, and diving into application layers, Francois, what applications are most exciting to you? And/or on the frontier side, what would you prioritize as key themes you're seeing?
Francois Chaubard:
I would say that the one that's probably going to impact the world the most that no one is really recognizing that everyone is ... It's been so wrong for so long that people don't really see that it's going to actually happen this time is robotics. I think that having worked on robotics, people and they rest on these frontier models, that everyone's using something called a VLA, vision language action models. So the VLA stuff is like really important and we're going to be able to buy Rosie the Robot and have it in our house and do our laundry and wash our dishes and do those things. And that's going to be massively transformative for the economy, for the world economy.
Oliver Chen:
Why now? I think MIT and Stanford have been trying for Rosie the Robot for like decades. And also, does this play into Diffusion Policy and the latest research?
Francois Chaubard:
Exactly right. So yeah, so definitely read the Diffusion Policy paper. It's shocking to me how many people outside of robotics don't know about it, even in AI. AI experts don't know about Diffusion Policy. And it's like, dude, this is almost bigger than AlexNet.
Oliver Chen:
Okay. Tell us Diffusion Policy 101, and I know this could take an hour, but you're really [inaudible 00:04:50].
Francois Chaubard:
We all know what these image generation models, like Midjourney type stuff, where I [inaudible 00:04:58] prompt and makes an image for me. The way that works is this thing called diffusion, and that was largely founded by this Jascha Sohl-Dickstein, Yang Song, Stefano Ermon. And those are the authors of some of those most important work.
Turns out, last year, they figured out you don't need to do this just on images. If you did this on a trajectory, which is like my arm is doing a trajectory right now, it's just X, Y, Zs over time. And if you did this on a human trajectory, human labeled that I can collect via teleops, I have some robot that I'm controlling via a mouse or hands or something like that, then I can record this information. And I need like 80 for it to figure out how to do a new task. And then I can take that data and I can diffuse it, noise it up, and then train my model to reverse out the noise and get back to that trajectory and it works insanely well.
So state of the art was like 22%, 25% before Diffusion Policy, and now we're at like 85%. So that's a huge jump and that's the core technology that they're using for most of these humanoid robotics applications [inaudible 00:06:14].
Oliver Chen:
Why was this so hard or why was this not able to be done before? What was the turning point-
Francois Chaubard:
Yeah, so this is the-
Oliver Chen:
[inaudible 00:06:20] compute or more formulas? It's not necessarily compute, is it?
Francois Chaubard:
Well, Yang Song published score-based matching in 2020. So it could have been done, all this could have been done 100 years ago, right? We just needed to find the right inventions and really this technology, the diffusion score matching algorithms that Yang Song and Jascha invented, and Stefano started in 2020. And then it took us three or four years to realize that we can do this for any format. It doesn't need to be images, we can do this with human trajectories as well.
Oliver Chen:
Okay. And then as you think about this profound change with Diffusion Policy, what does it mean practically or what are the examples you see evolving? Perhaps as you put your investing lens in too.
Francois Chaubard:
Yeah, so basically if you can write down 100 tasks that you want a robot to do in your home, we can go collect 80 examples or maybe 100 examples, maybe 1,000 examples of all of them, train this robot and then you have an LLM that's invoking, doing tool calling on that task and saying, "Hey, Rosie, can you please go and make me a salad?" And then it will invoke, okay, first thing, I have to go get a bowl, then it has to go get the salad, and it's invoking these micro tasks. And those are all little pre-programmed diffusion policies that I've already trained up. So it's going to be able to do these things that we could never do before and even dream about having robots actually do well.
Oliver Chen:
Okay. I guess that does lead us into AGI. Why do people love to bring that up and/or are you horrified or excited or how does it matter or not? And what is AGI?
Francois Chaubard:
There's like a really cool set of blog posts that talk about like what's the right way to think about AI? Is it a tool? Is it a species? Is it a emulation of us, et cetera? Is it a collection of people? What's the right framework to have?
And I think that it's very clearly a tool. And people have been so amazed by these things and they can't think about it well and it sounds like me, and so does it have feelings? Does it have motivation? And the answer is no, it does not. And that's not even close to what this thing is doing.
So the AGI, EA community, effective altruism community, I think has really done a fear-mongering tactic and scared a lot of people on like that, because we've all seen Terminator, et cetera. That's like the opposite of what's going to happen. It's like so far from what is, if you actually understand these things really, really well, and the people that are working on this stuff are like super not afraid of this. But the people that are farther away from it and that haven't really trained these models themselves, don't really understand what's happening under the hood.
Oliver Chen:
Well, but Francois, what about safety and also ethics and privacy because when you think about-
Francois Chaubard:
Oh, those are important things for sure.
Oliver Chen:
... frameworks like the ontological frameworks, utilitarian, and a lot of the reinforcement learning has to be programmed with values in the algorithm.
Francois Chaubard:
Yeah.
Oliver Chen:
So how do you compare and contrast that risk factor?
Francois Chaubard:
Yeah, and those have always been there. Even in computer vision, if I'm doing face verification and things like that, we've been working on that.
So AI safety to me doesn't mean we need to make sure that we don't make these things so smart that they kills us and turns us into paperclips. That's not a valid argument and we don't need to worry about that. We do need to worry that the self-driving car sees a baby on the left and a plastic bag on the right and does the quick math and say, hitting a plastic bag is okay, hitting the baby is not, and does the right thing. Now, that's just a training thing and we know how to do that. And that's why Waymo has had zero major incidents in like a decade or more. So it's possible to be done.
Some people won't do it and that's good. They're going to run out of business like Cruise, honestly. And there's not going to be taking safety serious enough. That's definitely going to be the case, but I wouldn't worry that, oh, well, someone is going to have this magical model, loss is going to hit zero, they connect it to the internet and literally all of our power goes out and the world is over. That argument is just completely wrong and not even close to that going to happen.
Oliver Chen:
Well, I think though globally, thinking about ethics and also the regulation relative to ethics relative to innovation, there does have to be a decision.
Francois Chaubard:
I would say that the right framework to apply is what happens today? What is the human deciding? How do we get the AI to align with the human? And that's actually a very well studied thing and we know how to do that quite well. The problem is that most humans don't agree on what the right solution is.
Oliver Chen:
That's called society.
Francois Chaubard:
Yeah. Yeah. So like what do we want it to do?
Oliver Chen:
Another great podcast because I did the ethics class at MIT and it's very troubling and forever interesting. What about the Y Combinator? It's a premier place. What are you doing there? What's your day to day like? And what are your philosophies there? And you have so much prestigious impact on AI.
Francois Chaubard:
Yeah. YC is just absolutely a phenomenal, phenomenal organization. It's the only ... I had no ambition to be a venture capitalist. I'm a founder through and through, but YC has the first core value is do the right thing by the founder, even at the detriment of YC. And I know no other venture capitalist firm that has that core value and actually lives up to it.
So that's when Gary asked me to come over, I was very happy to do it. My day to day is mostly meeting with great founders, trying to help them, stop making the mistakes that I've already made. I've made so many mistakes running Focal for 10 years. And if I can help them at all, it's like I feel good about myself at the end of the day.
Oliver Chen:
And what do you think the philosophies are in terms of Y Combinator and what's happening with AI? And the experience the firm has, what have been some key learnings and your perspective on what's a good strategy for the firm to follow?
Francois Chaubard:
Yeah, we've been successful. YC is one of the most successful venture capital firms in the world. Maybe us and Sequoia are the most successful. And our ethos is, if we're doing the right thing about a founder and we're really helping these ... Finding these people that maybe are first time founders and they're 19 or 20 years old, like Alex Wang from Scale, for example. And his initial idea, I think was like Zocdoc or something like that, and guiding him. And my partner that I work with closely, Jared Friedman, was his partner and guided him to labeling data for AI. And then now that was a $13 billion acquisition by Meta.
So like that, without YC, did that happen? I don't know. Honestly, I talked to Jared and Alex about that, but probably not. So we're making these things happen when then the counterfactual is that they don't happen.
Oliver Chen:
Amazing stuff. What were some of your key learnings that you're sharing?
Francois Chaubard:
Yeah. I'll wait on that because I'm doing a series called How to Start an AI Startup where I'll have a lot of those things. But yeah, I'll table that for now.
Oliver Chen:
And then back to Diffusion Policy, how might you compare and contrast that to the reinforcement learning models? I'm sure there also work in conjunction.
Francois Chaubard:
The RL avenue is probably the most exciting part. This is it didn't work, now it works. So the why now in RL is hugely, hugely important. And that's where I'm spending a lot of my time and thinking about what is the implication of this thing that now didn't work and now works on the economy?
Oliver Chen:
Well, it still didn't work though, because reinforcement-
Francois Chaubard:
Didn't work. Yeah, 20%-
Oliver Chen:
[inaudible 00:14:26] driverless cars, you mean which parts of it didn't work?
Francois Chaubard:
Yeah, so the driverless cars were extremely ... That was 5,000 engineers for 10 years of very handcrafted features. And it's a very different stack because it has to be massively real time. So it's a much different problem than Rosie the Robot. But self-driving car largely didn't work up until three years ago. We had some Waymos, but not very many, and now we have a lot of Waymos and now the things are really starting to work. So I actually don't know what's under the hood there, but I would bet that the innovations that happened with transformers and diffusion were certainly important for that.
Oliver Chen:
Well, Francois, it was really amazing to catch up. Thanks for helping us understand the future of investing and what you're doing at Y Combinator.
Francois Chaubard:
Yeah. Thank you so much for having me.
Speaker 1:
Thanks for joining us. Stay tuned for the next episode of TD Cowen Insights.