AI Models Will Tap the Whole Economy in 2 Years | Aidan Smith

It's so sad. I miss it. No, they're not. They turned into an AI lab. Exactly, exactly. Their AI went very well, I hear. Okay. Wow. So why did you choose the route of data efficiency over any other strategy that a new lab could have taken? Yeah. I think for us, the biggest reason to go off data efficiency is that we look at the broader economy, and we look at the parts of the economy that AI can touch today, and it's a tiny fraction of the overall value. And there's this weird paradox where the models are so much smarter than you and I, and also completely inept. They really can't do anything actually useful in the real world.

And most of the value is, frankly, stuff in the real world. It's these places where you actually have to be very, very data efficient to learn a new task or to adapt or to even kind of pick up this knowledge that's germane to a given domain. And so we thought that by investigating data efficient methods, we could go and tackle the other $30 trillion. As other people focus on code and what have you, we can go and focus on the long haul. And we really think that there's actually an advantage here for a Neolab because most Neolabs are trying to compete on these action practices with their folks. This is great, and I really respect our peers and all this, but I don't think that you're as likely to be successful there versus when you put yourself in a place where you have to innovate and you can't be on the nice stage like everyone else doing the same thing with the same methods with the same people, you can actually start from scratch.

Yeah. I mean, two things I'm particularly curious about. I think that what you guys came out with is the primitives for your GPUs. You guys can actually have different types of algorithms be really efficient there. And specifically with your background at Neuralink, I mean, it kind of hints to what's coming, and I know you guys didn't particularly release a model quite yet, but when you do, does anything from your background at Neuralink kind of shape what this is going towards? So it was hard to bring my dead horse to the studio to beat. It's a little bit too heavy, but the Neural stuff is a good inspiration.

Folks have been inspired by the brain forever in ML, and I think that justifiably, it's the only system we know that actually attains this track record and this performance in the real world. i think that we are much more inspired than than stuck to the brain and that like you look at it you say okay here are these properties it has can we replicate those properties versus like i am now looking at the hodgkin huxley I'm going to be trying to port that over to my GPU because, frankly, you have different substrate. The best algorithm will change based on substrate. That gets back to our thesis really about how new GPU primitives are really unlocking algorithms. The brain is actually this motivating thing where it's like, nature had these tools. These tools, which I might add, are actually really bad. If you go and you look at how slow neurons are, how much energy it takes to any communication at all, you're like, how did we end up becoming reasonably smart even a little bit? This is remarkable.

The brain is so low wattage and it's so small. It's like 20 watts, right? Same as a light bulb or something? Yeah. I mean, even more than that, though. Like, one thing I like to see is how slow do you think it is to go from one side of the brain to the other side if you were just to send a message? Probably much slower than the speed of light. Well, that's a very good guess as far as nothing is going to be faster than the speed of light. But, like, how – can you guess in terms of, like, orders of magnitude? Or give me a number in milliseconds. In milliseconds? So, like, in terms of, like, speed of light, it would be, like – what is that?

299 million meters per second or something. So that would be, like, what, nine orders of magnitude, less than a second? Okay, I'm going to guess it takes, like, I'm going to guess it takes eight milliseconds to get from one side of the brain to the other. It's not a bad guess, but it's actually wrong by, like, two orders of magnitude. It takes 150 milliseconds. And when you think about that compared to a GPU, in a GPU you're talking about microseconds. It's not even close. And when you even, like, look under the hood and you scrape back the white matter, and you look at it and you're like in order to even get that latency your brain has to pull some unbelievable bullshit your brain has to do this stuff where it's like stacking up all these fibers and spending you know 20x more basically energy on these areas to make these information highways and it's get tapping out at like 150 milliseconds that's so slow that like you would not be able to watch this video if it was streamed at that frequency right it's like yeah and you really

Aidan Smithyou look at this and you're like holy crap everything fits to the hardware that it's given everything is downstream of the primitives that you give it and so by unlocking gpu primitives

we can make models that are much much smarter than today's and also operate in different ways like have different latency characteristics and it can incorporate knowledge in more interesting matters why not try to do literal whole brain emulation like we have we have proof of concept for this with much smaller animals we had the sea elegance worm and now the fruit fly i have so many robin hansen has written about this you know age of m one of our favorite books of all time so like why not just do this so i love this stuff i think it's really cool i have many friends are working on this you should get my friend love chish off on the show if you want to see some really really crazy sci-fi as for whole brain emulation stuff i i think this is a really noble path but maybe it's slightly misguided insofar as there's two hard parts with whole brain emulation the first is actually much much harder than the second but the second is also really hard

Aidan Smiththe first is how do you get the data of what the heck's going on in the brain

our best scanners are woefully inadequate here whenever you get mri or pt or whatever this is really quite high resolution it's great and yet temporally everything's smeared to death because it's really the blood that you're measuring right and if you look at a device like a neural link a device that has immense potential to you know raise the floor for the human experience and create the transhumanist future like okay you can get high resolution neuron activations but localized to a given area and maybe that company will move towards getting more coverage and this of thing in the limit i don't know we'll see where they go but even then you're not going to cover every single neuron you're not going to get the full dynamics of the brain like that and so you have a really hard time even just in a really naive way measuring what's going on and then secondly if you were to actually have the measurements how do you run the thing if you think about number of parameters the brain has it's a lot and they're not necessarily parameters the same way that you might get a neural network they're kind of constrained some of their connectivity and things it's there's you cannot just analogy like 100 billion yeah so the question

you're talking about like neurons or synapses or what have you um but if you were to try to model them in the completely exact same way like with like a hodgepain hodgepain type thing it might take like 200x more compute i think something rough like that you should actually be on this this is something i'm not totally sure off the top of my head uh it's not actually a thing that you can map one-to-one in this clean way because your brain is taking advantage of physics it's taking advantage of these really really weird dynamics that make a lot of sense When you have the substrate that's kind of evolving forward and it's really grainy and actually doesn't really have a good analog when it comes to our computational substrates.

And so I think we should come with algorithms that are not like literally exactly what the brain is doing, but algorithms which are, well, for one, probably much better because we have a way more bandwidth, right? We can send a lot more information around. We can actually be a lot greedier. The amount of information exchange in your model is so much more than your brain really can do, at least, you know, per millisecond. And then these new algorithms... What was the other thing? The second thing is actually running the thing, like I said. So there's like getting the data about what is going on, and then there's running the thing faithfully, both of which are really, really hard, nearly intractable problems.

And I think I have some very smart neuroscientist friends who are working hard on these things, and I wish them the very best of luck. Fair enough. I want to shout out our sponsor, Lovable. You know the app you've been meaning to build? The internal tool, side project, or product you'd otherwise lose a weekend to? With Lovable, you get software you can ship today. That includes an editable code base you can inspect and change, plus two-way GitHub sync. Lovable also handles the backend and infrastructure, including managed Postgres, auth, storage, hosting, payments, and more, so you don't have to wire it all up yourself.

And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable. Lovable.dev. Now, on to the episode. These new algorithms you mentioned, what are you guys working on? Well, I can't speak too, too much of this because that would leak a bit of alpha, But you can imagine that we really are a vertically integrated shop. We look at GPUs and we're saying, okay, how do you go under the floorboards? How do you go rebuild everything? We don't use Torch. We don't use NVIDIA's normal CUDA runtime stuff. We basically have built everything from scratch.

We don't use Nickel. We don't use any of this. And as a result, this gives us a really, really unprecedented level of vertical integration. This means you can do some really crazy stuff like go and have your experimental training models run at the highest level of production inference stack at a big lab. And because we have this deep co-integration, the systems guys sit next to the post-training guys, next to the pre-training guys, it means that we really think a lot more like this brain analogy, where how do you fit the substrate you're given? We only got one shot. We only got one substrate. And frankly, the GPUs are blessedly a really, really flexible substrate, right?

There's so much that can be done with them. And yet everyone's kind of like, I don't know, it's just a mammal machine. What a waste, right? We really see a lot more potential in the GPUs than other people. I mean, given today's announcement, $1 billion raise, I mean, given what you guys actually announced, I mean, new primitives, GPUs, the KVM with exactly a whole vertical integrated shop where you can get the most out of new models with different types of algorithms, it seems like potentially there's some sign that there's some algorithmic breakthrough that you guys will probably be announcing coming up eventually.

I know we can't talk about it live, but this is where we're going. However, I am very, very proud of our pre-training team. I think they've done really remarkable work. I think all of them are biting the nails on their computers, waiting for the chance to go and hit the post button. When will we hear? Do we have a tiny deadline that you can give us? I won't tie ourselves down to a hard deadline because I think that I want to make sure they release stuff when they're ready to, and I want to make sure that we have a chance to scale and really set ourselves up in a spot to capitalize on our advantages and go and build a sustainable business.

But within the year, are we thinking? We'll see. I don't know. I'm just trying to push as much as I can. So how worried are you about your work leading to a major compute overhang? If you do figure out how to make current models 100,000 times or a million times more data efficient, couldn't that lead to a sort of very rapid discontinuous increase in capabilities? I know a lot of people are worried about the need to pace the frontier and allow alignment to catch up with capabilities and such. Yeah, so there's two questions there. The first is the compute overhang. The second is the alignment thing.

I'll adjust both.

Aidan SmithAs for the compute operating, like I said, we want to unlock this 90% of the economy, which is untouched right now.

That's so much more demand for compute just waiting to be tapped. I mean, if you think that we're seeing a lot of demand for compute now, imagine what it's going to look like in two years. Even if we're able to run 1,000x more tokens, imagine the demand for compute when you're actually able to tap the whole economy. That's huge. As far as the safety case goes, we are really thinking about it. I think it's worth taking seriously. It's worth setting up careful systems for observability and making sure that you're doing careful scaling science. I think that I have the privilege of being a lot of really, really smart folks who are thinking about this a great deal and who I get to talk to nearly every day about the state of safety and really make sure that we scale responsibly. But I really think that our goal is to make models that help people and that actually go touch that 90% of the economy that are actively useful. And I think in order to be useful, there's two desert data, there's actually go out in the world and do things and keep their world to do things in.

So for now, we're much more in the former case and we're taking lots of precautions on the ladder. I mean, I guess when it comes to the former, in theory, this now opens up a bunch of new possibilities to add AI into areas that like right now the capabilities aren't totally there or it's just not getting diffused into society in different ways. And so like what more do you see happening? Like the longevity of this? Well, the world's going to be a crazy, crazy place. I mean, I'm very EGI-filled, so to speak, in that I believe that capabilities are going to continue to improve and improve discontinuously.

Maybe our data efficiency will contribute to that and that models that can learn from more limited samples can update themselves better and help make the next run going better. I think we're well on track for this, even outside of Flappy. But, yeah, I think that it's hard to predict what the Jevons effects will look like, what the places where demand for tokens will leak into, and frankly, what the regulatory environment will adapt around that to make it tractable. It's not at all obvious what the future will hold here, but I'm really excited that Flappy gets to be a part of it, and I think that we'll really work hard to make sure it's a positive vision for that future.

Yeah. What kind of timelines are you working under the assumption of? Like, it seems like if some people are right, if the people who are most bullish on timelines are right, then Anthropic or Opening Eye will just, like, do RSI and, like, rapidly improve to superintelligence within, like, a year or two years. Do you think in that case Flappy airplanes would have a chance to succeed i mean tons for what's worth super respect our our colleagues and friends at openanthropic they're shipping great models we use our models every day i i i think their success is broadly great for the success of the industry um i think there's absolutely a case for coexistence and i see ourselves as being a very complimentary asset because at the end of the day what we're bet on is actually that at the age of research is not dead that there is a lot of stuff left to be discovered that is off the critical path for the capability stuff of these labs for example with coding it's actually this very narrow domain where you're it's easy to verify what's going on you can kind of just scale the thing and we've seen people scale a thing basically no one i know programs anymore myself included i love programming and i still don't program anymore

Aidan Smithand so what we see our alpha as being is is we do not have to focus on the same areas where everyone else is really having this dogged competition and driving cost down to near zero

and really frankly doing the super intelligence in code or otherwise we see that oh there's all these places where actually no one's really making any progress the models are still not that great at creating these new research ideas they're still not great at interfacing in these places where there's limited data the fact that data labellers like a job and a growing job even indicates that there's something something wrong right and it suggests to me that we will have a lot of room to operate in a separate space with a very different kind of tan yeah so what does ml research actually look like right now its models are so good at coding that you aren't writing any code anymore but also like you've described them throughout this whole talk as like being useless at real world tasks so i think again it's very spiky right they're very good at some real-world tasks and completely useless at others.

If I wanted to bring in that dead horse, can't do it. Impossible. You know, they can do backflips on stage with your unit tree or whatever because you have this program routine in that has a nice little PID or whatever. That's fine. You want to bring the dead horse they left outside? Nonsense. I can't even beat it. Like, I think that what it looks like using these models is actually taking advantage of places where they're superhuman and then really this almost like classical economic story of where's your comparative advantage? My comparative advantage is an idea generation. It's in actually deploying these models out in the world it's an it's about actually going and if scaling the things that need to scale and a lot of human things that frankly are quite hard to do at least we found in systems these models are still nowhere near As one, Benjamin Spector, and two, the rest of our inference team, I think that we've been both surprised and disappointed by the models.

There's some places where they're so spectacularly good for things in distribution, and then when you take them to slightly more novel things, like some of the things we're working on, which are really quite weird, it's a bit hard for the models to adapt as fast, and they might be more likely to make rookie mistakes and things of the sort. We'll continue monitoring right after this message from our sponsors. 11labs, AI that communicates at human level across every channel and modality. 11labs.io slash mts. scale your startup on neon millions of developers and startups have already chosen neon for their back end start on the free plan or get up to a hundred thousand dollars in credits for your startup at neon.com slash mts special thanks to our sponsor kong the ai connectivity platform connect apis llms agents and systems with serious security and governance konghq.com finally prove what your ai investment is worth with lairdan head to lairdan.com today to start measuring your ai's bottom line mongo db trusted to run critical applications by 75 of the fortune 100 the intelligent data platform for the era apply today for access to up to 15 000 in atlas credits dedicated technical support and exclusive events and go-to-market opportunities mongodb.com

so when you guys launch models are you doing general purpose models are you building them for a specific use case well so the most obvious sort of data efficiency is in data limited regimes like if you were to just go and ask me like i would say there's these places where you just have almost no data and you try and train a conventional model on them and like what it barely works uh for example i'll go off the one from yesterday i think it's kind of cute like DNA. There's actually not that much DNA data that's collected out there. Like robotics, actually. Very, very limited amounts of hours of robotics data in the world.

Like trading, actually. If you think about all the trades that have been made. Tiny, tiny drop in the ocean of data that is all LMs. And so, this being just a tiny, tiny piece of the broader pie of all the data in the world makes you realize, wow, we really are not actually covering the broader data regime. My co-founder, Ben, would like to say that.

Aidan SmithWe think that we have 99% of data in the world in these models, but we really have under 1%, because there's so much individual knowledge, like how you run your show.

Think about all the bespoke knowledge on the data to actually make these things happen or like like one thing that models are really bad at actually is writing interview questions for guests and i tried to try so hard yeah you know i try what is most load-bearing about your company yeah it's like it's a bunch of this shit it's like yeah like if if i asked it like write questions for like aiden smith of flapping airplanes it would come up with questions along the lines of like oh you know so like how do you you know plan to market your product like it's it's very basic it's very rudimentary um and the models have not improved at this at all since, you know, the first time I did a podcast interview probably three years ago.

Yeah, I mean, one natural story you can see is that part of the spikiness means that there actually are some regressions in some places. Yeah. Where, like, the model's entropy has collapsed. RLVR collapses entropy at a large scale. You take these incredibly smart models and you pass a K at them. It's basically the same thing as pass a K. Maybe that's because we're optimized for pass a K and you just had some teleological argument you could make here. But, like, realistically, the thing that the model's rewarded for is somewhat of a conversion objective. And there's lots of ways you can train models to get around that.

And many, many smart people have published lots of research trying to do that. But the current thing that works at scale and that people are really scaling up doesn't really follow that recipe. And I think that it's a big part of the story for the next few years of ML of like how do you maintain the diversity in these models while scaling their intelligence? To what extent does learning the data distribution of like web text actually teach you to learn the underlying distribution of like reality? You know, this is what like Ilya Sutskever talked about three or four years ago. He said like basically all you need is language.

And once you've learned enough language, then, you know, understanding language is understanding the world I think there is a lot of cool evidence that kind of supports this for example if you pre-train just on text and then do a few video samples or image samples you see that you're able to learn a lot faster and you can get zero shot transfer where you kind of learn that okay this ball is red I know ladybugs are red That thing is probably a ladybug. Like, that kind of analogical reasoning actually shows up a lot, and there's some evidence that the representations are kind of conserved in many, many spaces. Actually, people do a lot of work to make this true. This is a really, really hard craft. I don't mean to say, oh, it's simple. It works really easily. But I think there is a lot of truth to that. I think that the more data you're able to get to the model, the better it can be. And I think that as models have gotten better and the algorithms that optimize them, you have seen, actually, this become more true over time. So I'm pretty bullish on it.

Yeah. One thing I'm curious about, especially when it comes to new neolabs, is how do you actually get talent? You're now part of the AI talent wars. And it seems like from what you shared, you guys have a lot of breakthroughs when it comes to new types of algorithms, new type of technological breakthroughs. Where are you finding talent from? A couple places. I think the first thing to know about flapping airplanes is that flapping airplanes is a place for people who believe in themselves. This sounds kind of corny, but really, when people interview with us, we counter sell them a ton. Because, like, frankly, if they want to go work at OpenAire Anthropic, they should go do that.

Like, you should go to Flapping Airplanes if you want to be part of the last age in human history where you can do research. And it's actually really easy to go into a big org and just go hide under the floorboards. It's really easy to go into the big org and be like, okay, I contributed to that. And realistically, you had an incredibly small business impact.

Aidan SmithAt Flapping Airplanes, oh, you own this thing that's really, really important.

You know, this project will live or die on your success and a lot is demanded of you. And so you kind of have to have this entrepreneurial spirit. This kind of gets into the process of maybe with Ben where he's incubated so many great companies like, you know, Mercore, Etch, Cursor, et cetera, that it attracts these people who look out in the world with these hungry eyes like, man, I want to go and change things. I believe in myself. I want my work to be highly, highly visible. They believe in themselves. If they think that they're differentially important, then it shouldn't be part of this big monolith necessarily where their work matters less.

And of course, I have lots of friends who actually do really important work at these places and drive the frontier and all that. But it's a harder story for a lot of people. I also think that we have the privilege of having many, many very smart friends. And our friends have very many smart friends. And their friends have very many smart friends. And we also are willing to bet on different kinds of people. Like, we've had some people inbound us who are very radically different. Like, a Clash of Clans world champion. If you interview, you're like, holy crap, this guy is really, really smart. Yeah, I mean, like, I think if you're a world champion at, like, a highly competitive thing, like, you have to be, like, really extremely good.

Especially if it's a strategy game. I mean, like, this is how... The story of how we became really good is incredible. And I wish I could tell it to you, but it might cause him to be investigated by several authorities. I mean, it's like one of the reasons why a lot of people hire from like the MIT battle code winners, right? Like if you're insanely good and competitive there, you're probably good in a bunch of other areas too. Yeah. It's just an unfakeable signal of aptitude. Absolutely. Yeah. But even lots of spiritual signals are not necessarily good. For example, we have lots of IMO medalists, but also lots of IMO medalists are maybe not the best machine learning people.

We really want, like, diversion thinkers, people who are willing to go and work incredibly hard to make a differential change in the world and believe that they can actually do that. And there's a small group of people, really, who has that set of criteria. Yeah, yeah. I know we're on time. I have a last question, but I'll let you ask if there's anything you really want to share because this is an exciting announcement. I want to make sure that we get asked. I do have one last question. Like, what product might you eventually ship? I'm sure you get asked this a lot, but. it so i think there's interesting questions about on the timeline for this i think in the long long run like i said we look at the world 90 percent of the tam not going anywhere many ways to address that i think i detailed a few of those earlier in the short term though there's lots and lots of things where i think frankly it's on us to decide how soon we want to commercialize i guess to risk leaking a little bit of alpha uh you can imagine that we have models which are paired with performance of many frontier models or many many open source frontier models that exists at different points in the throughput interactivity curve.

And so there's an interesting question of like, should you go and go really, really upmarket and sell models that run at unbelievable speeds that literally no one else can offer? People will go and they'll be like, hey, can you do this? And ask many inference people and they'll say no. And they'll be like, yeah, it's already working. I think that part of our ability to do this really deep co-design means we can ship artifacts like this. But at the end of the day, we really We really want to go and conquer the world, and so we want to make sure we don't get mired down in these smaller markets without an obvious path of how we tackle the long-term one.

My last question for you, what can people expect over the next few months? Well, for those folks in the company, you know what to expect. It's going to be a very, very exciting few months. For those outside, we'll see. I wish I could promise you something concrete, but I want to make sure that we announce our things on the right timeline. I don't want to force artifacts we're not quite proud of as an org yet. And I'm unfortunately going to have to keep you waiting because I want to let the cake bake all the way and not ship you guys some E. coli. Fair enough, fair enough. Well, very exciting. I can't wait for what Flapping Airplanes comes out with.

Yeah, I've been following you guys since the start, so it's great to have you in the studio. Aiden, thanks so much for joining us. Thanks for having me. It's wonderful. Thank you. Big congratulations. One billion dollar raise. We don't hear about that every day here on MTS. So thank you so much for joining us. Thank you so much. And we'll be back with MTS. A huge thanks to MTS sponsors. Blitzy, autonomous software development for enterprise codebases. Ship 5x faster. Blitzy.com. AdQuick, make your brand a billboard. Out-of-home advertising as easy to scale as digital. AdQuick.com. Arena, measuring AI performance in the real world.

Arena.ai. Support for the show comes from VCX, the public ticker for private tech. The U.S. stock market started history's greatest wave of wealth creation. From factory workers in Detroit to farmers in Omaha, anyone could own a piece of the great American companies. But today, our most innovative companies are staying private longer, which means everyday Americans are missing out. Until now. Introducing VCX, a public ticker for private tech. Visit GetVCX.com for more info. That's GetVCX.com. Carefully consider the investment material before investing, including objectives, risk, charges, and expenses.

This and other information can be found in the Innovations Fund prospectus at getvcx.com. This is a paid sponsorship.