Regeneron ISEF 2026 - The Future of AI: A Conversation with David Holz and Maya Ajmera

Um, I think all as just like a random kid in Florida, there weren't a lot of scientists around me, but every adult that cared and that listened and encouraged, it has a bigger effect than you think. Um, obviously outside of that, I would try to look to different historical figures. Um, Richard Feynman did something weird for me where he had this ability to not just think deeply about topics, but to feel deeply too and to kind of do so both at the same time in a way where I didn't as a kid feel like I had to either be sort of purely logical or purely emotional. They could kind of come together as one thing.

Um, I think the biggest thing was just coming out here and meeting all the other kids. Um, definitely take as much time as you can to meet each other. It is an unusual group. Um, more than you would think. Sometimes you come here and you might think that, oh, it's like this everywhere. It's not. Uh, and so I think, yeah, that was really it. And then obviously doing science for over many years and finally getting to the ISEF was by itself like a just a really cool thing.

My project was um it was about uh locating and tracking sounds in 3D with microphones. So my board had a bunch of microphones built into it and they wires going into a computer and we would and as we would walk around the board and talk about the project it would show in 3D where we were based off our voices. So it was pretty cool. Uh yeah I mean I was really into sensors and like tracking and understanding things and that sort of ended up going into my second my first company which was Leap Motion which did hand tracking and so me being kind of into sensors uh is definitely related. Uh Midjourney is not as much into sensors. It's sort of the opposite of sensors where we're not like listening or making things, but it's sort of related. It's just the backwards. Um, and uh, Midjourney is working on a lot of hardware now. That's more sensor stuff again, too. So, it's coming around.

Yeah, so I never really heard the word entrepreneur as a kid. I was just like I wanted to invent things and I wanted to make things and I had this idea that um I would go to college, I would get a PhD, I would go work at NASA, I would eventually have my own lab at NASA and then one day I would have my own lab like outside of NASA and I had this sort of like long story that that would go through and then at one point for a variety of reasons I was like what if like I just sort of skip a lot of those steps and go straight to trying to make my own thing and Um, you know, it might feel risky at the time, but there's a lot of risks in your life that are sort of two-way doors, as they say. Meaning like, um, my school liked me enough. My other like my school was like, 'Oh, you can take a leave. We'd love for you to come back.' NASA was like, 'Oh, you could, you know, you can take a leave. We'd love you to come back.' Max Planck was the same. And so, everyone sort of was like, they didn't really want me to go, but that meant that I could come back. So, I was like, 'Okay, I'm going to take six months, 12 months, see what I can do.' and I uh I guess never came back.

I think I remember in 2012 or probably 2013 um was the first time I heard of like neural networks and there was somebody who was saying hey why don't we use neural networks for that 2013 is really you know before a lot of that became popular and I remember sort of saying well why would we use those I asked a lot of questions they didn't really have good answers for me and eventually it was it was kind of bad it was kind of a fault of myself at the time I kind of judged them for not having good reasons and I was like well if you don't have good reasons it's probably not a very good idea but man if we had used neural networks in 2013 it would have been really cool and I think over the years like looking back at many things like that um I would sort of find that in many situations you might like overanalyze somebody's arguments or even overanalyze the person themselves when what you really should be doing is trying to just seek inspiration from everything around you. So in this case, if instead of saying I think your arguments are wrong, I was like, hm, why are you like you're saying this thing that's an unusual new thing like is there something there, you know? Uh and then I could look into those things and I could be like, oh, there's something that works here. There's something that I could be inspired by. I could use this tool uh even before anybody else is using it. There have been many cases where I would say um almost every opportunity that I've missed in some ways was sort of this um was sort of like um I would say enacting judgment instead of seeking inspiration out of a given situation. And in some ways like you know life isn't necessarily about giving you answers as much as it is about giving you inspiration. So the more that you can kind of take anything that comes at you, whether it's good or bad or right or wrong, and say, 'What kind of inspiration for new ideas can I draw out of this,' uh, everything just works better.

Um, I'd been working on hand tracking for maybe 10 or 12 years. Um, and I had this idea that it doesn't matter like what computers do or what we can do, but how we work together is really what kind of changes the world. And I wanted to work on things specifically where you think about bringing those two things together. But over time, you know, working on that for so long, I realized that even if I gave people the best hand tracking and the best user interfaces, they still wouldn't necessarily think or make things like I do. And so I wanted to kind of step back from the interface between people and technology and just think about the process of creation itself. Like what does it mean to come up with ideas? Like what does it mean to reflect on ourselves about what we want? How do we work together? and try to just sort of I wanted to build a new company really focused on that on the sort of creative act and I wanted to spend time thinking about that and exploring that and so I yeah decided to go in a new direction.

So, I really wanted to work on this sort of this creative process like where do ideas come from and how do we think of them? And so, there's these different processes. There's like you reflect on what you want, you imagine what could be, you work with others to get there. And I had a bunch of projects and ideas for each of those different areas. And so, at first it was very unpinned. It was in just a purely explorative uh state. Um, and somewhere along the way, um, I was reading all the papers and there was a paper on diffusion models and it was a really new thing and I was like, hm, these seem different. There's like something here that's it's not the same as GANs. It's not the same as the old neural networks. Um, there felt like something significant. And so I talked to a lot of my other friends and I was very privileged, you know, be out in San Francisco. I knew all the other major researchers at the time and we all said together like, yeah, like these seem like a big deal. Someone should work on this and there's a lot of things to work on in AI and you know I think in this case I was like well there's something fuzzy here if we we're trying to help people make stuff with it like what's what is a good like what's a good picture what's a bad picture it's like not well defined what is the interface that allows you to make those things um these are like they're not like purely quantitative tasks they're very fuzzy problems and as somebody who worked in human interfaces for a long time I kind of like fuzzy vague not well-defined problems um I'm sure many of you work on the same sorts of things. But I would emphasize that in the world, most people are not good at that. That's a rare skill to be able to work on something that's not particularly well-defined, that maybe doesn't have a good number for it yet. People are very good at optimizing numbers. The second the numbers are gone, most people feel lost. Uh, and so I was like, well, there's something very David shaped here. And so I decided I would uh try to just do it like explore diffusion models and the creative process. And Midjourney um is big in many ways because it was the first company to actually uh ship diffusion models at scale. Um like there was nobody there was no diffusion model that more than like 100 people could use before Midjourney. Um and it was a tricky thing to do. We had to like I had to go to Google and I was like Google I need 10,000 GPUs to do the you know week one launch of my product. Uh, and I mean normally, you know, you wouldn't really just ask somebody for that in like 2021, but um, you know, one nice thing about being out in an industry for a long time is they kind of knew me already. They're like, 'Oh, David, I know that guy like and he wants GPUs.' Like, 'Okay, cool.' And, uh, it took a while to kind of like, you know, get from my friends down to somebody who could actually flip the GPUs on, but when they did, you know, we had the compute we needed and we we launched the first diffusion models.

Um, it's really I think in AI a lot of people use the word like intelligence and I think we don't actually have enough words for that concept. Um and and and and as you as you play with AI models, you realize that there's some more nuance there. Um like what does it mean to like there you know there's some some some things I'm sure you've heard of like people talk about the reasoning models and so there's some models that just sort of like babble and then there's some that try to go through a reasoning chain and so there's like the reasoning is something that's discreet from just intuiting. You could say that the no models they intuit it and then these sort of newer models they reason. Um but then there's also sort of other things where um uh you'll find that sometimes the models are like really good at something and then they're really bad at something else and the thing that they're good at seems so smart and the thing that they're the thing that they're good at seems so smart and the thing they're bad at seems so dumb and it's not really clear what's happening. Um if you kind of go under the hood a bit, um a lot of this comes from sort of like other concepts like are you um does the knowledge required to do a task uh is it sort of in between two things you already know or is it outside of those things? For example, you know, if I've seen uh if I've what's a good example? Um, you know, I've seen a lot of I've seen a lot of if it's images, I've seen a lot of cats and I've seen some blue dogs, which is weird, but then I can probably make a blue cat. Like some new things, even if they didn't exist, are sort of between other concepts. And when you do whenever your situation is like that, the AI will just solve the problem. It's sort of the sort of interpolating. It's sort of blending between things that already exist. Um but then sometimes something is extrapolating like it's sort of outside of the set of things which has been seen and the cognition required to do a task like that is different and we don't actually fully understand yet what it is and how we do it. But if you can sort of understand sort of what your situation is is this like a situation where I'm blending things that are known or am I really going outside of the set of like stuff that people have done before? Any situation where that is the task you're doing, the computers just can't do it. Um, and they might not be able to do it for a while. We really, the AI field, we don't know how to solve that problem. The best theory we have is like maybe you can teach the computers certain algorithms and certain tools and maybe those algorithms work outside of the set of situations that you built them for. For example, you could hand somebody a hammer and they could hammer nails or they could hammer other things and like, you know, you can use a tool outside of its situation it was built for and that's like something different. But I do think that actually when you look at the sort of total set of like where like new ideas and where do they come from? Um, there are other things happening too and we don't fully understand it but like sometimes you have to rethink the whole world from scratch. Sometimes you kind of just have a moment of of randomness that just happens to be right. There's like a lot of ways that new ideas can form and that's really where all the a lot of the opportunity is. If you can sort of take your talent at those things and then leverage the tools that are just good at everything else, you can do like crazy things now.

Um, you know, I I think just I mean I think there's there's there's two categories. There's if there's something you can automate and go superhumanly fast at like uh maybe there's some new laboratory things. Um there's like there's a lot of situations where you could get uh some kind of experimental loop that happens really quickly and in parallel in a way that humans can't do. Another one that people don't talk about enough is anytime where you can simulate something um even if it's slow or it's unrelated like there's a lot of ways to then leverage AI to then sort of um to do real world tasks. Sometimes they say well you have to simulate exactly the right thing but you don't. You can simulate a wider set than what reality is and then the AI will just learn everything including reality. So there's a lot of really cool types of things like that too.

Um I mean in some ways you could if you think of AI as like a friend that thinks they know everything but they often don't. Uh they probably have a productive relationship with your AI. Um there. So, it's hard to know an easy way to explain this from a technical standpoint, but neural networks are truly they can neural networks can kind of uh understand if they have seen a lot of data that disagrees with each other. They can un they can estimate their variance, but they can't estimate their bias. And so bias is really it's very hard to estimate and they they often can't estimate if they've never seen something before. So there there's a lot of sources of just complete, you know, uh all sorts of weirdness. So you kind of have to treat it like Yeah. Like one of those.

Um, so I think there there's a lot of angles to come at this, but to me I think we know we're always going to build bigger machine bigger and better machines. And I think the to and and like that that's an interesting process, but to me personally the most interesting question is not so much like what do we want to build, but what but the big the more interesting question is what do we want to become? And so there's this idea that everything is going to change over time, but you know what? But but but that implies we're going to change too. And so trying to understand like what could we become, what do we want to become? Um yeah, it's just a thing I wish more people were working on. And so I think Midjourney is a little bit at trying to to poke around at that. Um, yeah. And, uh, you know, sometimes people think that they should be the same thing, but if you think about it, like if an AI on average is going to be sort of like an economic agent, like it's there to kind of do economic tasks. And you can optimize it to be the perfect economic agent. Uh, okay, let's do that. Let's pretend it happened. But then, okay, what do we want to be? Do we want to be the perfect economic agent? We might want to be something else. In the limit of things. I think where humans want to be in a million years is different than where a perfect economic agent would want to be in a million years. And so I think it's very normal to believe I think it's reasonable to believe that like there's sort of two destinations that are really important and I just uh we should be thinking about both of them equally at I think at least I'm a little biased towards the humans because I guess I am one.

Um I think a lot of um interacting with a big model is a weird thing where it's some combination of your ideas and the ideas of like everyone who's ever lived and then some other everyone who's ever interacted with the machine. Um, it's a weird new thing, but sometimes I like this metaphor of like it's like a garden sometimes, you know, it's almost like it's a you don't train in some ways you don't build an AI, you grow an AI. And in some ways when you make things with AI, you're also sort of growing things. So, it's kind of like a garden and you're like, you know, did you make that tomato plant? I don't know. I guess you you grew that tomato plant. There was some part of your work there. And that tomato plant is a reflection of yourself in some ways, but it's also a reflection of every tomato plant that's ever existed. And that's kind of cool, too. Like there's something weird there, something bigger than us, but still a part of you.

Probably the future. I'm very I like the past, too. This I guess I'm just like I feel like I have more of a connection to the past. Like I I read a lot of ancient literature. I think about it a lot, but I really don't know. Like at some point you would rather travel to the future.

H that's interesting. Um well actually some of the people on the original Midjourney team they have this thing called aphantasia. So I think like 5% of people or uh can't actually visualize things. I'm sure somebody in the audience is wondering this now. They've never heard this concept but uh most pe Yeah. And then some people I have hyperphantasia like I wildly hallucinate things. Uh and I think there's actually like we can see how people look different but we often can't see how they think different or how they experience the world differently. And I think there's a lot more diversity there than we actually uh think about. Um and I think a lot of Yeah, I think there's definitely an aspect of these tools that I think I mean they interact with us in really different ways. So for my friend on the team, he would he would never been able to visualize anything in his life before and he was now visualizing with uh this diffusion model and for me I'm like hallucinating all the time and I'm like hey man this will hallucinate with me and it feels really different. Um I don't know it's really interesting. I think we're going to see all forms of that. I mean, in some ways, um, even the language models probably do that a bit already. Like some people they hear voices in their heads and some people they just can't and so that so like you're having this conversation with outside of your mind. Um, and so there's something there about like um yeah, sort of building I think this is where that you build these things sort of become tools that extend you. Um, it's interesting like when you hold a stick you often think the stick feels like it's a part of you when you poke at it. So I think ideally we could build the tools in a way where they feel like these extensions of ourselves and they aug and they they they augment our they they they amplify ourselves. Um and I sometimes uh I think that's a really important direction to go that's not just like um building like other creatures that are not you. Um so like this yeah that's it's yeah there's a whole important set of questions there.

Um, so we were the first ones to come out with an image model. Um, and so it's definitely been interesting seeing like everyone else doing it more. Um there's definitely um I think the biggest area that I'm excited about is this concept of not necessarily making images that are more accurate but sort of trying to imagine the explorative process happening through these images and asking how can we amplify that process. So, a thing that I talk about internally at the company is I say, I would like us to try to reach something I call the aesthetic singularity. Like, what would it mean to have more aesthetic visual exploration in one year than all prior human history? And everyone's like, 'Yeah, you could do that.' But it's funny because like if you don't put the question out there, everyone's like they don't even think about it. Uh and so for us we're thinking about how do you um put these feedback loops up where you sort of watch everyone exploring and then you try to kind of amplify that force. So instead of sort of converging it you sort of diverge it but then you still it's hard to say exact you know I don't we haven't figured it out yet exactly. I'll just say that but uh it's a problem you can actually sort of start to pick at um how do you amplify the explorative process of millions of people in a way where collectively it's all going through one system and so you can say I want collectively us to explore more in one year than all prior human history. So that's like pretty cool.

There was this Pope picture where the Pope had a big puffer jacket and like a big metal cross and stuff that went really popular. Um, I was delighted to hear that uh uh yeah, one of my friends, Dario at Anthropic, he he actually showed the pope this picture on his phone. He was briefing the pope on AI and he goes, 'I want you to show you this.' And the pope goes, 'I don't remember having that jacket.' And then one of the bishops leaned in. they go, 'Oh, we know all about that.' Uh, so that was that was great. Um, yeah, that's probably my favorite one. There's a lot of Balenciaga stuff, but you know. Yeah.

So, Midjourney, I mean, it is one of the larger revenue AI companies, but it's also we have no investors, which is really weird. I mean, it only happens every 10 or 20 years. There's hap something like this happens. Like Craigslist was this way and Valve like they make Steam and, you know, stuff is this way. Um, you know, you really you never have to do things a particular way. Um, for me, I was like, it wasn't that I was against investors. I was just like um uh I guess I was like I'm gonna use my own money first and um you can do funny things like you can loan your own money to your company and then the company could pay you back in the future if you raise money like you can do all sorts of tricks but like you know I think I just um I spent all my money until I was like negative $1,000 and then it's very motivating to make money when you're when you have no money left. uh and uh but you have to understand when you raise money sometimes the motivation to make money goes away and then instead you get the motivation to just raise more money and raising money is very different than having a product that people are willing to pay for which I think is kind of a beautiful thing like in some ways if people aren't willing to pay for it is it really adding value you know sometimes but it's a nice way to kind of it's a proxy for knowing like am I adding to people's lives and so yeah