Stephen Messer on AI, proprietary data and economic foundation models

If the open source model are so close, what are these US frontier labs doing wrong? I think there's two things. I think there's, one, the, the marketing belief that this is the path to AGI, despite the fact that everybody in the field does not believe that. LLMs or agents, which one is more overhyped right now? LLM. I mean, you can't talk, uh, about agents without LLMs, and the amount of capital going into these providers, that is an easy win, hands down. If a CEO in our audience has one AI budget to deploy this year, where is the highest return dollar?

That is a hard question. If I am a CEO, the only thing I would- Welcome to another episode of 51 Insights, today with Steven Messer. Steven, welcome to the show. Thank you so much for having me. It's exciting to be here. Excited to have you here again, Steven. Last time you sat in that chair, you told us that large language models would end up like web browsers, free everywhere and worthless as a moat. That is exactly what happened. You invented affiliate marketing, you sold LinkShare for four hundred and twenty-five million. You're an internet entrepreneur since over thirty years.

Since two thousand and eight, you have co-founded and led Collectivei, one of the earliest AI companies. Steven, it's awesome to have you back, and we're gonna talk all about AI today. I'm so excited to be here, and I feel like I'm in witness protection because when I said that, a lot of investors who had put money into the frontier models, I had to go into hiding. And then I put an article out on my blog, reloadnyc.com, and I called it Peak Token, and I think now I need to stay away from starting my own car in the morning and, uh, walking out across the street alone.

Let's talk about this first. So you called the commoditization of LLMs eighteen months ago. Can you just unpack that for us? What did you mean by that, and did it actually play out like you expected? Language models, by their nature, are commodities, right? The training data that they use to train the neural net is publicly available. And what you saw was the really, um, compressing of the gap between frontier level skills, the, the latest, hottest models and everyone else, right? Even if you think about it, two weeks ago, three weeks ago, Mythos, oh my God, it's going to take over the world, and it's so advanced.

And, uh, in the span of the last week and a half, we've had open source models that are so close that it's indistinguishable, the gap. Coming out of China, coming out of Japan, coming out of all these other places, because a lot of them are using a technique called distillation, and they don't necessarily have to do the hard work to get to the same outcome. And that's a really hard place. If, if your gap and the premium you're asking people to pay for a language model disappears in two weeks, but it took you a year to make, that is not the sign of a moat that is going to be defensible.

And I think this is where we're seeing all the benchmarks everywhere tell us that these open source models are so close, um, that this is not the area where, you know, real value will end up getting created. If the open source model are so close, what are these US frontier labs doing wrong? Why are they spending billions of dollars of investors' money to build these models if models that can be created for a fraction of that money can compete with them? I think there's two things. I think there's, one, the, the marketing belief that this is the path to AGI, despite the fact that everybody in the field does not believe that.

There's this marketing belief that this is how we get to superintelligence, and whoever gets to superintelligence, it doesn't matter if you have a week lead. If you've got one week, superintelligence would then expand so quickly, it would go so fast that that week would, would turn into years, and there'd be no way to catch up. And this is why you constantly hear the, the frontier model players talking about how much of the code is being written by the AI itself, and this idea of a recursive org that's growing. And there is a piece of that that is somewhat true in that sense, and we haven't reached a limit yet to the scaling laws, and we haven't reached a limit yet to the model experience.

So their argument still holds true. And I think, I think that's where people are, are arguing that this is where the underlying value exists. Then there's other people that are looking at world models. Yann LeCun, Fei-Fei Li are looking at it from a robotics perspective. I think at Collectivei, we're looking at proprietary data around understanding economic transactions. Um, everyone's sort of going down a different road of saying, "Okay, what will get us to that marvelous wonderland?" I think the other thing that, that was unexpected was this, the GPU strength to be able to do a frontier model at scale.

This is, is not inexpensive, right? We, we've watched this crazy build-up of data centers, and the argument was that in many ways, the only way you can get to these scaled models that are so great is you have to have this massive compute, and then you have to have this massive inference. This was Chili Pepper, uh, which was the, um, OpenAI's new chip, and all these other players that are coming out there arguing that. And then you had this announcement that comes out of this small company called Nvidia with another small company called Microsoft about a completely new approach to computers and how they will function and where you can-- what kind of local model you can run.

And all of a sudden now you ask yourself the question again, do we need these massive data centers if I can run a lot of this on new laptops that are coming out that are gonna be much more efficient for GPU-like models? So this is where the balance of battle is taking place, and it, it is a all or nothing game that people are playing. Are you saying that building frontier models is gonna be essentially a, a software problem and not a hardware problem anymore? I think this is the argument that... Well, let's play, let's do the two arguments.

If you're a frontier model company, your argument is that- There, we have not reached an end to scaling laws, so we don't yet know what will happen. For example, we don't know how humans, um, got consciousness and decided to do it on their own. In theory, maybe if we get the models big enough that the machine will have its own consciousness, it will reach the super intelligence, but that's gonna require huge computing power. And if you get to super intelligence first, well, yeah, this thing is gonna be a recursive thing that's gonna keep until you get better and better and better, and that's a game that's worth investing in because if you do that, you will essentially be the, the intelligence for the globe.

A-and that's a really big value. And so that's the argument of massive data centers, massive compute, massive training data, massive, massive, massive, massive, and that's gonna be the barrier. There's a counterargument. The counterargument from the open source players are, one, I've been able to create competing models for a fraction of the cost. This was the DeepSeek moment a year ago, where all of a sudden everyone looked and said, "My God, that model is so good. It was done with a small team and a fraction of the cost." And then you look and you start to see the, the newer hardware that's coming out from people like Microsoft, which will eventually affect Apple and other players, and they're able to run these extremely high-parameter models locally.

And then you have the argument of, okay, are we starting to see a break of the market where most of your jobs will be run locally on your own machines? This is sort of the OpenClaw and Hermes and argument of where I'm gonna run my own harness, and I'm gonna keep my data to myself. But if I have a problem that's so big, okay, I'll pay for tokens for the frontier models for something that my local model can't run, which seems to be getting smaller and smaller, that I'll need that Mythos level or whatever the next model will be, that I'll pay a premium for that.

But ninety percent of what I do I'll do on my local machine. But that blows up the economics of the big players. That's debate that's going on right now is which is this gonna be, is can the local models do enough, and that I'll-- it'll gut the model. This is sort of my idea of peak token, which is how much can you charge for that little incremental improvement. And y-you know, in certain areas, maybe you could charge a lot if it does something where it, it finds all your holes in your security. Okay, that's a premium you're willing to pay for because security is so valuable to you.

But are we getting to a place where that's getting smaller and smaller? This is why you see them focusing on biology, you know, the, the most recent move from DeepMind over to OpenAI, Anthropic. You're seeing these, these fi- battles of trying to find places that are getting more and more narrow to get the value out of that peak token. My question is, can that cover the cost of everything else? Well, speaking about the cost of a token, Palantir's CEO, Alex Karp, was on CNBC yesterday, and he basically spoke about the AI industry's token model. He said it's insane, and he said that enterprise are livid at Frontier Labs over token maxing.

Do you think these big enterprises have realized that they're overspending on tokens, and that the better solution might be to build their own internal AI compute clusters? Look, uh, that is the-- we're at the very beginning of people's realization of how these systems work. And I say that because we had an announcement a month ago of essentially forward deployed engineers and building essentially a replacement to the Accentures and the McKinseys, and a lot of companies were eager to get involved in that from private equity on out. But what that says to you is that they don't understand that what they're doing is basically bringing in a bunch of engineers who are trying to build barriers to exit.

For example, if I build all of my agents using Claude Code, and all of a sudden there's a cheaper alternative, I have to rebuild all those agents. If I used an OpenClaw or some other harness, I could just switch the underlying model and never have to change anything. But a forward deployed engineer coming out of, you know, OpenAI or Anthropic, they're not gonna be incentivized to go build these abstraction layers that anybody else in the world who understood token costs would say would be the core foundation. You would absolutely wanna build these abstractions. Take, for example, uh, Open Router or Neuromorphic, two great companies that are basically saying, "I'm gonna build an abstraction layer away from the underlying models, and we'll optimize your costs based on what model you need." These are all areas where you're starting to think about tokens and how do I get the efficiency out of it.

And I think what Alex really hit on was two things. One is they know this. OpenAI and Claude know this is a risk of commoditization, which is why they created this forward deployed engineering. And he put it in a much cleaner way, which is if they thought they had something so valuable, they would say they want thirty percent of what this thing was valued. The problem is it's so commoditized that if OpenAI says, "I want thirty percent," I'm just gonna go to an alternative model and get it for free, which means they don't really have the pricing power.

And that tells you a lot about where these models are. Take Collectivei, for example. We have proprietary data that trains our model. We're not a language model. We focus on predicting economic outcomes. And in ours, there's no number two. There's no alternative model because it's proprietary data. We had to get it from all of our customers contribute it, and we did in exchange for making them better. But we could go and say, "Okay, if you wanna do anticipatory shipping, here's what this model costs," because there's no number two. Now, I don't bring that up to say, wow, Collectivei is a great company.

We are. But what it shows is the difference between the commodity trained models and proprietary trained models, and the gap that you're starting to see emerge between where people have invested their money wor- versus where value is getting created. And speaking about Collectivei, you mentioned at the very beginning, uh, economic foundation models. You also mentioned world models. What we were s- just speaking about are LLMs. Can we take a step back and unpack that for the audience? What are these different models? How do they compare against each other? For your audience, you've been told by everybody that AGI, this, this, this amazing artificial general intelligence, will do everything.

That is not where we are at today. What we have are what's called narrow intelligence. We have expertise. So each model only becomes good at a narrow field that it's been trained on. Words and images can give you the illusion that these things are better at things. So I'll give an example. A lot of people are saying, for example, that they're using an LLM to d-- predict the stock market. Now, anybody who knows these models is having a heart attack when they hear that, because an LLM has no concept of time. Stock markets are time series problems.

This is crazy to us. When we hear you're thinking, you're using a model that studied Reddit discussion groups, and you're using it to analyze the trading value of a stock, all it's doing is rep-repeating words that sound like it makes sense but have not done any of the analysis. It's crazy to us, but it's being positioned to you as, "This is gonna make you a ton of money." And really what you're getting back is a feedback of what everyone's talking about on Reddit. That's kind of scary if you think it's actually done some fundamental analysis of your business.

That's not what it's doing. So what you'll see is that all these other models are emerging to solve different business problems. There's a biology model, AlphaFold. A language model will talk to you about protein folding. It will hallucinate what the, how, how a protein will fold. You will end up with a drug that will kill you. AlphaFold is a model that's designed to predict how proteins fold that will create drugs. Very different model, trained on very different data with very different kind of skill sets, and it does it really well. But if I ask a AlphaFold model, the thing that won Demis the Nobel Prize, to predict the next best word, it won't do that well.

And that's what you see going on is-- and every one of these is a different model. We, for example, are predicting economic outcomes. What does that mean? Well, I'll give a simple example. If you've ever had salespeople and they're trying to predict which deals are real, that's an economic outcome. If you're a private equity firm and you're looking to buy a company, and you're talking to a whole bunch of them, and you're trying to figure out which one's real, which one's really going to do a deal with you. Because a lot of private equities buy family-run firms, and so maybe they wanna sell the business, but then the family gets into a fight.

These are expensive endeavors. If you're trying to figure out which ones are likely to happen, you need to know information about how they make decisions, not at a general level, but how this group of people makes decisions. That's where you need a model. And you can keep going. Yann LeCun, for example, and Fei-Fei are focused on w-- a world model. This is how do robots learn to navigate the world they're in. And to do that, you need to figure out how-- what ha-- When you and I enter a room, we use our eyes to spot where things are, and we as a human can figure out very quickly how to navigate a room.

Machines can't. And so they're learning how to, how to navigate different situations, whether it be for manufacturing or whether it be to deliver you your, your restaurant order. These are things that they, they're trying to now build models around to make it easier for machines to learn. Steven, you founded Collectivei in two thousand eight. At that time, no one was speaking about AI. How did you build that company, and when did you realize that it's economic foundation models that you need to focus on? From our perspective, we knew neural nets were gonna be a big thing.

Um, we had a lot of friends in the industry. I used to be in an area of AI called expert systems. Nobody needs to know that stuff. It was old. Um, and, and we knew it was coming back. I think we looked and said, "This stuff is pretty amazing. Um, it's still not totally here." And we just started experimenting. For a long time, we kept playing with this technology with a small group of people, and what we realized was that i-if you train it on something that's a commodity, it becomes commodity. And we asked ourselves, "How do we compete against Google, Microsoft, uh, you know, Oracle, some of these huge companies, Meta?" Um, we just didn't think it was, it was the right place.

We asked ourselves, "Where would a model be highly valuable?" And if you think about it, when we looked at it, we said, "Okay, a hundred years ago, Adam Smith talked about the invisible hand. Why, why are we still making decisions based on this invisible hand when everything in the world has got a sensor on it? Can't we make better decisions? Can't we avoid economic catastrophes like, uh, recessions if everybody has a better sense of what's going on? How do we know that we can stay away from making bad decisions economically that affect all the employees of our firms?" And this became the foundation of where we focused.

And it turns out we were right. You can actually do all that stuff. And something, just to come back to what, uh, Alex Karp said yesterday on CNBC. He also said, speaking about proprietary data, it looks to me, listening to you, it really all comes down to having proprietary data. And one of the things these big corporations are also saying is that, "Why should we pay millions of dollars for these models and at the same time also train them with our proprietary data and basically lose our competitive advantage?" He's right. I'll add one thing that Alex isn't saying, which is, you can have proprietary data that's still not valuable, and I think this is what gets lost all the time.

For example, a lot of the, uh, medical firms are like, "We have the on- we're the only ones with this data." Yeah, but it's biology. I can use synthetic data from a small sample set and train a model by doing that. Is it valuable? Not as mu- as valuable as they think. And the challenge with a lot of really large organizations are, they believe because they're large economically, that they have a lot of data. The challenge is most of their data is still dramatically too small. No one company has enough data. I think we chose economics because what we're studying is an externality.

And what I mean by that is, let's say I'm the largest company in the world. I'm a Fortune One. I'm, I'm pr- the biggest player out there. If they're studying their own process, an internal thing, they may say, "I've got the most data about how to manufacture a car." You know, I'm Ford, I'm Toyota. Um, I've got that. Okay, fine. Maybe, maybe you are, and I'm not gonna argue with that for now. But if you're studying an externality, you have nothing, right? If Toyota's trying to figure out how to make a car that more customers will like, they don't have enough data.

They only have the data about their customers. Toyota, as big as they are, is still a small percentage of an overall market, and as you saw with the Chinese, that didn't stop them from, from gaining huge market share. So what ends up happening is if you're JPMorgan Chase and you think, "I have more trading data," well, you're gonna be dead, because everybody else who's out there is gonna pool their data toge-together. They're going to get an insight from a broader base. This is why, for example, marketers use Google. It's not that they don't have a ton of data.

If I'm P&G, I have a lot of data about my buyers buying Ivory soap or whatever P&G sells. I just don't have enough data to make a good decision. And so a lot of people misunderstand having data and having insights, and I'll give a great analogy for your, for your audience. There are a lot of oil wells that are out there. Saudi Arabia has tons of them. The US has tons. Where does all the value tend to accrue? At the people who refine the oil. Your oil is junk, right? When you look at the oil, the oil that came out of Venezuela, it's a heavy oil.

It's got tons of problems. It has to go to one of two refineries in the world that can handle heavy oil and convert it into gasoline or jet fuel or whatever you want. If you can't get into those two refineries, you have nothing. And this is what Standard Oil learned a long time ago. Whoever controlled the railroads and whoever controlled the refineries was where all the value accrued. Companies have raw data. They have oil wells, but they don't have refineries. You need massive refineries to turn this into value. That's, I think, what a lot of them miss.

And so Alex gives a part of the answer, 'cause his goal is to sell tools to you that let you use your data more. He's not an AI company in the s- in the true sense of creating models. He basically takes those models and uses it for your own data. And so Alex, I, I, I love Alex. He's fantastic. He's so smart. But he's definitely selling his own book as well. I hope the analogy I gave gives your audience a way to understand where value actually accrues. You're saying companies need refineries, and that's exactly where Collectivize plays into, right?

As I remember from our conversation last time, you're combining data sources across different companies. Your model watches something like five percent of the global economy and prices in things like rate moves before a company feels them. Can you walk us through an example of what this actually means for a company that uses Collectivize model? I'll give one that I think is more general, 'cause, uh, give me, giving an example of a private equity firm knowing when a potential company can be, uh, is interested in selling or not, that's gonna be for a, a small group of, of listeners.

So let's say you've got a sales team. And this is a great example from one of our customers. They're trying-- As a seller, you're trying to figure out who's real, because customers will waste your time forever. If you are out there, uh, venture capitalists will waste your time. Uh, all-- Think about all the people you're trying to convince to do something, and they're out there, and you've gotta go decide how much time, money, and effort am I gonna invest in these people to try to get the deal I want. Well, that's really hard to do, because my guess is most of the time you're dealing with people for the very first time.

And some people are passive-aggressive, and they'll say, "Oh my God, Mark, that's great. I'd love to invest in you. Come back to me with more information." And they'll drag it out and, and keep you going forever. They'll cancel meetings. They'll do all stuff. And you're so excited because they've told you things that you wanted to hear. "I'm interested. I like what you're doing. I wanna see if this can grow." Nobody knows. But here's the deal. You can know. I'm gonna bet that your loved one, you know their patterns pretty darn well. If they say, "I'm interested in doing something," that could mean they're interested, or if they say, "Yeah, let's do it," it means definitively they're interested.

You start to learn that pattern. Well, if I see the same people across multiple individuals, multiple companies, who are both try-- all trying to sell the same person, guess what I start to learn? I learn their patterns of behavior. I learn what they say when they wanna do something. I learn how fast they move. I learn even the people they bring in. And I'll give you a little cool insight. Did you know that most people, when they wanna make a decision, will bring in one group of people, and if they don't wanna make a decision, they'll bring a completely different group of people into the meeting?

So when you go meet a venture capitalist or you go meet someone or you're trying to get a big deal done, the group around the table that they've selected is their way of making the decision without having to make it directly. So if you're a venture capitalist, you don't wanna be the one who said no to what might be the next Google or OpenAI. You may bring in a group of people so they say no, and you can say, "I really wanted to do that deal." And you would never know. You just think, oh my God, they're, they're now bringing in five people into this meeting.

This is gonna be a good meeting. And those five people, just by having them brought in, means this meeting's over. You just didn't know it. Our ability to see across all the companies we work with lets us know, oh God, you got the bad D- team. And you got that, you find that out before it even happens. This is the stuff that we're helping people understand. But yeah, we do it because we recognize that we need to see this buyer so many times or this decision-maker so many times across so many different people that we just learn How they make decisions, and it makes the world predictable.

Steven, what is your biggest challenge when you explain this to your potential customers? Today, everyone believes that there's language models are AI, and they are one model. And so when I have to explain to them that, look, you're going to have many different models, and you're gonna string them together for their insights in the same way that when you're doing a big deal, you bring in a lawyer, you may bring in a CFO, you might bring-- You're bringing in expertise, and everyone's contributing to a good outcome. This is what's gonna happen. You're gonna have multiple levels of intelligence inside of your org, and then you're gonna string it together to get advantages that allow you to behave either as a faster org, smarter org, or some combination of all of that.

And that's what's happening today. So they'll say things like, "We already have Claude," or, "We have OpenAI. We have AI." And you're saying to yourself, "Okay, you've got something that can write pretty language. Great. What does that have to do with, with me?" Like, if you wanna grow revenue, Claude is the most dangerous thing you can bring in. It doesn't even understand time. And what's the one thing we know about sales? Time kills all deals. And you've chosen an AI that doesn't even understand time exists. Good choice. And I think this is what's happening, is people are learning so much so quickly that our, our biggest thing is we have to help them understand, no, you need a super intelligence that understands your buyer, that understands the port co you're trying to bring on.

It understands all these different economic challenges, um, that are happening in the world, from the Straits of Hormuz down to inflation. You need something that understands this to get an advantage, and they're gonna learn this bit by bit and start to learn how to string these things together. Steven, you've been an entrepreneur for most of your career. Right now, we have an explosion of new AI startups building tools, building SaaS solutions that many say are just wrappers around frontier model. What do you make about this argument, and what would you tell these people what they should focus on to create value?

I think the wrapper phase is rapidly sort of dying. Um, I think even with Cursor getting bought, people were excited. But if you look at Cursor today, they are trying to build their own model now with Groq to have their own coding model. But it-- That, that's a rarity. Uh, most of the players you see out there are, are sort of dying a, a slow, painful death, mostly because they're not getting funded anymore, and secondly because the, the-- all the players that were out there that had their own models are just waiting for you to build it, show the market, and then they're coming in and aggressively building Claude Design or Claude something else.

And I think that investors now have clued into that. I think founders are cluing into it as well. Um, can you have models that, that work? Yeah, you probably could, right? There, there were Instagram is one of the few, WhatsApp is one of the few that survived the app market, where Apple didn't just take the idea and build it themselves. Um, so there will be some that will survive. I don't think those are gonna last more. And the ones that I think are the ones at most at risk, which I've written about, is the SaaS players.

You know, their, the-their idea was, "I'm gonna just build a wrapper of language around whatever my SaaSy thing is, and now I'm an AI company." I think they are, they are the ones that are most at risk because they actually believe they have AI. In speaking about bubbles, AI names drove about eighty percent of last year's market gains. The financing between Nvidia, OpenAI and the clouds is visibly circular, and we just watched a chip sell-off halt trading in Seoul. So to investors, what is your take there? Is this a bubble, or does the answer change whether the technology delivers?

What's going on here? Look, nobody can say this is a bubble based on the revenue that's been produced. This is not-- In the early days of the internet, the bubble was companies were being valued at multi-billion dollars and have zero revenue. That still turned out, for the most part, to not be a bubble if you looked at it. You could actually call it a correction, and then if you held on to Amazon or any of the good names like Netflix, you would have done extremely well. There was a valuation gap, but that, that three years later was pretty darn good.

So if you're a trader, you can argue that maybe there's a valuation mismatch, but I, I think this bubble talk is crazy given the amount of revenue you see going into these markets. Now, the circular nature of the investment to growth. Here, I would argue it's a trade, if you want to say it's overvalued or not. SpaceX was a trade, right? You-- There is no justifiable reason other than Elon seems to keep coming up with more ideas that are gonna be multi-trillion dollar businesses. I'm banking on him. Some people will be right, some people will be wrong.

Jeff Bezos was a trade that people banked on, and it worked out pretty darn well. And Tesla has had its ups and its downs, but it seems to be working. SpaceX seems to be working. So I think people are banking on that. But I will put one thing out there, Mark, that I will only tell your audience. Are you ready for this? Yeah, of course. Nobody really wants to talk about the hedge that all the buildup of data centers is. If I were to ask you, Mark, what would you put the risk at China's encirclement of Taiwan in the next four years?

Very high. Okay, give me the percentage. Fifty percent. Okay, fifty percent. I would argue even at three percent, that risk is so high that the buildup of chips and data centers in the US, even if it's paid for by frontier models for a short period of time, what is the value of those assets to Oracle, Microsoft, Google, Meta, if there's an encirclement? Five times? 10x? Right? Because these are general purpose chips. You could argue during the e-commerce boom of the '99, 2000, that a lot of those warehouses from Webvan and-- couldn't be repurposed. But a GPU is a general purpose chip.

Having a stockpile of those chips might mean the difference between two or three years of continued growth, shifting from one use to another, or having the economy crash. And so maybe an overbuild isn't a bad risk allocation if you're Google, Meta, Amazon, Microsoft with a massive cloud business. And I think there's a lot of ways that a lot of people are thinking that this is a way to ensure where the street will allow you to do CapEx, where we can make money from it today, but it also acts as a, as a little bit of a backlog if something were to happen.

If Terafab in the US and, and TSMC's factory, which is still a three nanometer chip, I think a lot of people look and say, "This could be worth more," and that's a future value that you should take into account. If it's fifty percent, then this is an undervalued stock because those chips will be worth gold. And I think Europe doesn't have that backlog built up. I think parts of Asia haven't built up that backlog. Right now, the US has a, has an asset that can be leveraged for everything from your Gmail need to your hosting needs to, to AI.

And I think that's an undervalued appreciation in the, in the asset ownership. Another thing that's also playing out, Steven, is these assets, whether it's chips or whether it's frontier labs, are also becoming increasingly national security topics because they become so important and so strategically relevant. And now we just saw the latest happening with, uh, Fables, the latest entropic model, how that got stopped by the US government. It is now back online because it was apparently so powerful and so strategic. How do you think this will play out in the next months and years? Because there will be a next Fable coming out, and there will be different frontier labs bringing models with that capability.

I think the idea of locking out a model is insane. One, it drives people to other markets. I think if you look at other countries now, they ask the question of, "Can I rely on the US?" Um, and they will find alternatives. Right now, China isn't really an alternative. If China was not China, a draconian government, you know, that is looking for power, those models would have won by now. They're, they're cheaper, they're better from a price to value, they're more efficient. I think there's more creativity in some of the work that, in particular, DeepSeek has done, but other models too.

They would be the winner right now. The person driving a lot of this is, is actually Dario and Sam. They need to try to create a lockdown on open source models. A lot of what you hear, I think, is more marketing than not. I think, uh, you know, yes, they got shut off. Yes, they, they had a fight with the Department of War. Yes, they did all these things. But the reality is, they're trying to say, "These are weapons of mass destruction," and that you need to lock these down, 'cause if you had the biggest commodity model in the world, and no matter how much money you were spending, you couldn't get a lead that was more than point three percent better, I'd think you'd be nervous.

I don't think the rest of the world will buy that. I think what will end up happening is, if they get it locked down in the US, the rest of the world will start shifting away from them. I think it's a, it is a very dangerous strategy. One that I think is BS. Locking down knowledge is not the way you win. Um, you, you win by building better products and being ahead of everyone. Do you think that there's actually some truth here, that there's a legitimate concern about the security risk of these latest models? I believe, like all-- like any time that there is a zero day discovery, that is a huge risk.

These coding models like Mythos are starting to find long run problems in code. Don't hold back. Like, it means we have to move quickly to build fixes. And I like the approach they took where they said, "Okay, we're gonna give it to some of the bigger software providers, so we can find these problems and release bug fixes before the world knows about it." I think that's responsible. I think that's great. But, but we should want tighter code. We should want security found. There's a reason why countries like y- like Israel started red teaming, um, and that spread then to the software community where you try to find vulnerabilities.

That is not a bad thing. Finding vulnerabilities means you can fix them. This idea that w- it's so dangerous that we can't release it, but yet we're gonna let these flaws sit out there is to me, uh, uh, not a great thing. I think as the more of these tools get out there, the more we'll have perpetually tools looking for security vulnerabilities before a third party does, and I think that's a healthy thing. By the way, it also comes out on the verge of Q day. So are we really worried about, like, finding our flaws when RSA is likely within the next two to three years to be broken, which is the core of all security?

I think our goal should be to get ahead of all this stuff as fast as we can. Yeah, definitely agree. And Steven, we're almost at the end of the show. I have a couple of very juicy, uh, questions before we end this and, and go into a lightning round. Uh, the first one is, if a CEO in our audience has one AI budget to deploy this year, where is the highest return dollar, honestly? That is a hard question. Mark, if I am a CEO, the only thing I would probably ask is, what is the one thing my competitor can't keep up with?

And that could either be speed. I need my company to behave faster. I believe an advantage is speed. Uh, this, this I'll give for everyone who might not be in tech. If you can go into a hotel that allows you to check in and get your key and go right up or sign in with your-- with an app versus go to a front desk, which company do you choose? ... faster one You're going the faster one, right? So for some companies, speed will matter. For others, it might be that you can provide something that's int- more intelligent.

So you have to look and say, "What is the one thing my customers are going to value more than anyone else? And if I can do that first, I will win." Amazon did this with lots of little wins, right? I didn't wanna put my check-in information or checkout information in, one click. That was all it took for me to switch my shopping from other places to them. You've got to find the one thing, and that may be language, it may be an image, it may be an economic model like Collectivei. You're gonna find that thing and then go all in on that thing, 'cause winning the big thing will allow you time to get all the other things.

That would be the way I would think about it. What's the most important thing happening in AI right now that the market has completely mispriced in either direction? I think it's all the other models. I think, um, right now 90% of all venture capital in the world went to two companies, and it's causing a dislocation inside of all the others. The other thing, if you look, is venture capitalists aren't funding these things anymore. It's companies. It was Nvidia. It was, you know, Microsoft. It was other... Who are starting to realize, okay, there's an advantage here.

I think a lot of the big tech companies should be the ones who are investing in the other models because that's what's gonna drive their business success. I think that's being missed right now, and there's a hole in the market, and that's not being filled yet. I think the other models, what-- Yeah, you could talk about Yann LeCun, but there's models in almost every area that are getting built that are gonna own sectors of the economy forever, and I think people are just missing that. Then I have another one, Steven, which is very important because it affects a whole generation of young people coming into the jobs market, trying to figure out what they wanna do, trying all this AI stuff, building with AI, building new companies, seeing tremendous job losses across the board.

If you were 25 again, and y-you were at the start of your career, and you wanted to build companies, what would you do? What would be your advice to a, a young person today? My first piece of advice is don't believe anything that anyone has told you about job loss. That is insane. Yes, you see people switching jobs and changes in the roles that they are in. That's always taken place. There's always... The internet changed roles. The, uh, mobile changed roles. All this stuff is always in change, and that's the nature of evolution. Be glad with it because it opens up opportunities for you.

This is the first time in history that individuals and small groups can build massively. I don't even think people fully appreciate even other revolutions that are taking place. Three-D printing. Imagine today, if you wanna produce something and manufacture it, you've gotta fly out to a place like China and try to find a manufacturer who won't screw you, who won't take your money, won't make something bad. Well, you're getting to a place where you can design something, come up with the idea, do the marketing, and have it manufactured locally and have it delivered automatically, and you don't even have to hire a whole staff to do this stuff.

You are entering a period of time where your ideas are unlimited. You are literally limitless. How could that be bad for someone young with ideas for where the future is going? My goodness, this is the greatest time in all of history. I would kill to be starting my career now because there's nothing you can't do, and there's nothing you need huge resources to do. You can make almost anything happen today on your own without the need of venture capitalists and everyone else. You will dominate the next hundred years. This is the world in your hand.

That's what you should be excited for. I think there's no greater time than today. It is amazing. I get excited by just what could be done. Amazing message, Steven. That's a, a great ending. And let's do a short lightning round as well. Very quick questions, very quick answers. The first one is LLMs or agents, which one is more overhyped right now? LLM. I mean, you can't talk, uh, about agents without LLMs and the amount of capital going into these providers, that is an easy win, hands down. Second one is economic foundation models, a category in five years or still just Collectivei?

I mean, selfishly, I always like to believe that, that it's, uh, it's one model. I think, uh, because scaling laws do matter here, I do think it ends up being predominantly one model, just because of the lead we have. The more customers use us, kee-keeps getting better, kinda like Waze or, uh, or Google Maps. But that doesn't mean an Apple Maps can't come out. So I, I'll still hold it out there. And last one, one book or idea that shaped how you see the world and has nothing to do with tech or sales or AI.

Cat's Cradle by Kurt Vonnegut, an amazing book about religion and, um, and how, you know, people find a way to believe and have that no matter what. Um, a book that, uh, changes your life forever. Amazing. Thank you, Steven, for coming on the show. This was very interesting. Uh, where can people learn more about you, about Collectivei? You can go to collectivei.com. You can sign up for the free version if you want to get access to all your relationships in a way that's accurate and verified, uh, at intelligence.com. And if you wanna hear my crazy ideas, you can go to reloadnyc.com for my blog.

Uh, sign up. I put out two blogs a week. They're, they're, they're full of crazy ideas that some will be right and some will be just crazy. Crazy and interesting. I confirm that I read them every week, so highly recommend it. We'll link them in the show notes. Steven, thanks for coming, and all the best. Mark, thank you for having me. Always great. Um, you're amazing. You obviously like this video enough that you got to the end. Listen, do me a favor, hit that like and subscribe button because I think you'll like it. And if you want even more, with more I mean incredible alpha research and digital asset market updates, subscribe to our newsletter on 51, that's the number five one, insights.xyz and get the most actionable insights on digital assets.

See you next time.