Beyond LLMs: How AI is Set to Reshape Global Business with Stephen Messer

[on hold music] Welcome to another show of 51 Insights, today with Steven Messer.

Steven Messer, a pioneer in AI, uh, an American entrepreneur. Um, Steven, welcome to the show. Mark, thank you for having me here. I'm so excited to be on 51 Insights.

It's one of my favorite, uh, podcasts to listen to, and I'm just excited to be a part of it. Yeah, likewise. Uh, today we're, we're gonna cover a wide range of topics focused on AI.

You've been an AI pioneer, and I'm super excited to dive in. I'm honored to have you on the show. Steven, can you give us a little bit of background? You started Collectivei, your current company, in 2008.

You were very early in AI, very early in big data. What's your story? Yeah, it actually goes back even further.

So all my life I've been involved in technology, but early in the, on, uh, the days of AI, I started in a field called expert systems, which today is something that's been forgotten.

But it was the idea of studying someone who's an expert in a field, learning to build heuristic algorithms around their decisions, and then applying it to bigger categories.

And we watched how that world evolved, and so after that sort of

grew and then collapsed, one of those nuclear winters of AI, I left and started a company called LinkShare, uh, which is the world's largest affiliate marketing company, sold that to Rakuten.

At the end of selling it to Rakuten, a lot of my friends who had been in expert systems ended up at other places like Google Brain and brought me back into looking at what neural nets were doing.

That got us heavily involved in, uh, different kinds of neural nets, which is what led to Collectivei. That's a very impressive journey, and, uh, as I said, you've been early in data.

And 20 years ago, I remember- Yeah... uh, I think it was a headline story by The Economist, "Data is the new oil." Is that still an accurate statement today? Which is funny.

You don't even have to ask me, just ask all the Middle Eastern countries that have made investing in AI the core of their platform.

These are firms that understand where, what having commodities means and how to maximize the most out of it, and these are the players that you see today investing in everything from Anthropic to OpenAI to, you know, Midjourney, to all the major players.

The money is not coming from venture capitalists these days. It's really coming from Middle East countries. In a talk in 2012, uh, Imagine, you mentioned that data is really meaningless without the right context.

What did you mean by that? Well, that's a good question. So I think what ends up happening-- In fact, in that presentation, if you remember, we ga- we asked people the question, "Who looks more like a terrorist?"

And of course, the terrorist was not what people expected. And what it meant to show is that oftentimes data with a, with a lack of context can lead to wrong in, um, insights.

And in fact, this was the world of machine learning for the last, you know, 30 some odd years. A human being would go in and do something called feature selection.

They would pick the data elements that matter to them, and then they'd pick a model, and then they would try to figure out correlations, almost always overfitting a model to get to the answer they wanted, not the actual true answer.

Deep learning is a n- totally different approach.

It says, "Give me as many data elements as you can give me, and I'll use this thing called stochastic gradient descent to figure out what the most accurate way of looking at the world is."

That's why they call it learning algorithms.

So in that example I gave, which is everyone has a belief of how things should operate, including how AI should operate, and oftentimes that belief is really what causes some of the underlying problems.

If you get good data, and you give it enough depth, you can really get to new kinds of insights that can change the world. How, how has this changed now over the last 10 years?

If you look at big data, if you look how AI came into that, what is different than 10 years ago? The world is completely different.

There-- You know, when we started this, I think, look, we've had great success in our career, and even great friends in the technology fields would turn to my co-founders and I and say, "What are you guys doing?"

Like, "Why? Why are you here?" And I think the, the second thing that changed was OpenAI released ChatGPT, and we went from being what they thought was the dumbest people in the world to the smartest people in the world.

And then we discovered that they all thought the world ran on just LLMs, and that was gonna be everything.

And in fact, I still see today that most people talk about LLMs as if it's going to be the basis of artificial general intelligence. And it may be, but it's unlikely to be just the only piece of that that gets us there.

Even though everyone now loves AI, they love one particular kind of AI without the realization of all the other things AI is currently doing that may even be way bigger than language.

And I think that's what's exciting is you're still in this world where people's eyes are open, but the, the magic is still yet to come. Yeah, you mentioned LLMs.

LLMs had experienced a huge boom since ChatGPT got released two years ago.

If we take a step back and look at the last, let's say, 15 years of AI, from your perspective, what have been kind of the cornerstones and breakthroughs that you experienced with Collectivei and as an expert in that field?

Well, I think you ca- you can ne- you can't stop-- Even though I, I said LLMs are not the be-all, end-all, the transformer model that came out of Google Brain was clearly one of the big turning points.

It opened up the proof points that people had been waiting for around neural nets.

I think the diffusion models were the s- were a quick follow on, and I think what you're now seeing is people realizing where and how far can we push these models.

But maybe for your audience, can I take a step back and lay out- Sort of the land of where AI is and sort of ways to think about it that may make it a little easier. Yes.

Because I think people just struggle with this concept. And by the way, the place I hear about the struggle more than anywhere else is Silicon Valley. Silicon Valley missed the whole AI thing.

I know people will say it came out of open-- it, it came out of, you know, California with OpenAI, but, you know, in reality, the-- some of the biggest breakthroughs have come from DeepMind, um, that's based in London.

Um, if you look at, at the Facebook Meta team, they're all based out of New York. This is a broad group, and so people wanna believe there's one thing, one place to do it.

Let me give your audience a rare look into a framework that will help them understand how to think about AI. Is that a good idea? Definitely. Let's do it. Okay. Let's go back to first principles and start off.

The best way to think about the AI landscape is to really think about models being trained in three different ways. The first group of models I would call commodity models. Now, you may say, "Steve, whoa, whoa, whoa.

What do you mean by commodity models?" What I mean is the data that trains the model. So a foundation model is nothing more than a trained neural net, and a neural net trained on one particular insight.

A commodity model is something where the training data is freely available. Language, images, audio, video, and you could argue math. They're freely available. Where does the data come from?

It comes from scraping the web. There are complexities around stra-scra-scraping the web, there's no doubt about it. You have to label and annotate the data.

You have to clean out the dark net stuff that you don't really want the AI to be trained for, and, and probably now you're gonna have to pay for high quality, excuse me, content that you'd have to get to train your model.

But the reason why I call them commodity models are they're easy to train. If you've got enough money, and that price is coming down pretty steeply, anyone can train a model.

This is the shock that everyone's watched with the launch of ChatGPT, and then how fast everyone around the world has caught up to near parity. That's the first model.

So people will say, "No, no, no, those are gonna be the models that change the world." I might make an argument that I think OpenAI and Anthropic, this is the new browser war. I don't know if they're gonna be the Google.

I would probably put it more like Netscape. Hugely important, everyone thought Netscape was gonna be the control of the world. In the end of the day, they were roughly parity to all the other browsers.

We all settled on whoever gave it to us for free, who gave us good enough. Great, fine. That's my, my belief on commodity models. The second kind of model to think about is what I would call commodity plus complexity.

This is more where DeepMind went. DeepMind is probably, if you, if your audience is not familiar, is one of the most important companies in the world. But these are things like protein folding and material science.

The periodic table is-- does not require an NDA to sign, but it does require a lot of chemists to get together to figure out what are the properties of materials as they're, as they're mixed together.

And this is gonna be important for a model. If you're Samsung and you want to create a material that has a certain conductivity at a certain temperature,

well, an AGI model that's designed around material science requires great data scientists and great chemists to put that together. In biology, it's protein folding.

You have to-- These are, these are complex systems where, yes, the commod-- the data's commoditized, but to turn that into something requires great skill. So there are a lot of models that are gonna be like that.

Much more complex, fewer of those models, harder to replicate, but highly valuable. And then the third kind of models that'll get created will be around proprietary data, and that's a whole different group.

So I lay that out because if you're trying to understand AI and you think one company is already gonna win it all, we are far from that.

And in fact, I think the one player that everyone thinks is the winner might end up being nothing more than a Netscape, but we'll see.

What do you say to people who, um, argue that the race of AI is basically already decided?

You have the big tech companies building massive chip farms, um, working with that data, and they're basically impossible to catch for smaller companies.

They will provide the f-foundation models, and then startups can start customizing those models to their specific needs. Is that the scenario we're going towards?

You know, look, I, I have learned long ago, as Yogi Berra said, that predicting the future is hard, um, you know, including the prediction of being hard. So I don't wanna make that call.

I think what I would say is this: In every revolution I've been in, people want to believe that one player is going to dominate. When Netscape started with the browser, everybody said, "It's game over."

The browser wars happened, nobody remembers Netscape anymore. They ended up donating it, became the beginning of F-Firefox Foundation. When Yahoo! was the leader in owning the web, everyone said it was over.

Google emerged and won that one. When Friendster launched and said, "Oh my God, this is the, this is the first social network, it's the only one that matters," Facebook was the fiftieth. They won.

I would say if you are looking at the world based on who owns it, the first thing is remember there are going to be many foundation models.

In those three categories alone, you are going to see the rise of different models with different value.

I would not-- If I were banking on is the foundation model area finished and one player won, I think you've missed the entire way this AI world works. So you, you mentioned those three categories.

A lot of people talk about LLMs. Where do LLMs fit in in all this, and why are LLMs important, or why are LL-LLMs overrated at the moment? Look, I think LLMs really were opened up the category, right?

It's, it's amazing to be able to dialogue with something that mimics humanity's intelligence, and use it as an interface to doing whatever it is you wanna do.

But here's the challenge with an AI that's purely based on an LLM language model. If you ask any question, let's put it into practice. In my industry, we, we study-- We have an economic foundation model.

We study how the world does business. So an LLM, if I were to ask, is kind of like a professor in a gilded tower.

If I say, let's say I'm the CEO of a company, and I turn to that LLM and say, "How can I improve the success of my business?" The answers that will come back will be like, "Hire well.

Structure your business so it's clean and smooth. Try to make sure there's good education for your team. Come out with good products." If you ask it what products are good, it'll recycle what it tells everybody else.

So yeah, it-it's pretty cool to be able to do that. But here we are two years after, as we just celebrated the second-year anniversary of ChatGPT, and people are angry. They want it to do more. They're kinda frustrated.

It-- Sort of the illusion has worn out. Now, if you take a language model and combine it with, say, an economic foundation model, when I ask that same question, "How can I improve my, my business?" It doesn't tell you

information that makes you feel good but it's unactionable. Now it starts saying, "These four people have a problem. These 10 people could use support. These deals are at risk.

This business could be better if you had less of this inventory, more of this inventory." It ties it in with another form of li- uh, uh, of, um, uh, of intelligence. So as a interface, it's helpful.

As a provider of insights, it's not so helpful. So this is why I worry a little bit about the overhyped nature of the LLMs, which is getting a language model today, they're a dime a dozen.

The language models like Llama that are, are very good, the incremental value you're getting is, is tapping out because we're running out of words to train these models with.

And so even the things that they're trying to layer on top are getting more challenging 'cause they don't add so much more unique value.

Combining them with other forms of intelligence, if I were a doctor, a, a, a model that studies scans of the body, having a language model that can interpret or work with the, the other model can now start producing intelligence.

But I worry that the language models are a little overhyped and will probably end up being more interfaces, like browsers that are commoditized to zero dollars.

You, you mentioned economic foundation model, and I think here we really get into the meat of what you're doing with Collectivei.

Can you explain what you mean by that, and can you explain how this works together with other models or an LLM, for example? Yeah.

So, so as a, a language model is trained on language, and an economic foundation model studies how the world does business.

So in the land of proprietary data, at Collectivei, which is just short for Collective Intelligence, we use a grouped or pooled data set to start studying how the world does business.

And the way we do that is we connect to proprietary data inside of companies about how they grow their revenue.

The model combines it all into a universal data model, and then studies how the same buyers are buying across different customers.

That means that we're basically taking Adam Smith's invisible hand and making it visible for the very first time.

Imagine an AI that not only studies how the world does business, but can then actually start guiding how the world is operating.

It may sound fantastical, but imagine if you could actually run an economy by having an AI optimize every single day of how business is being done. That's what we're talking about.

We do use language models as a way for our users to ask questions. So if you wanna understand something as complex as, How is the war in Ukraine affecting my business?

It's easier to ask that question and then have that question translated to an economic foundation model that can give the real answer. Mm-hmm.

It makes it easier for a user to have that dialogue, and in fact, we allow the user to talk to our economic foundation model.

You can literally have a discussion as if you're talking to Nouriel Roubini, one of the greatest economists that are out there.

You know, you can ask questions that might seem overly complex, and it will do the work for you, but it'll do it by learning across the globe, by actually observing how commerce is taking place.

What's the difference between training an LLM with proprietary data, let's say you, you, you feed him all the data that you got on your business continuously, compared to using an economic foundation model? Yeah.

There's a lot of differences. So if you think about it, transformers were designed really to optimize for language. Diffusion models were designed to optimize for images.

An economic foundation model has to look at time series problems. So a way a company behaves during high interest rates versus low interest rates can be dramatically different.

How it, how, how it behaves after war breaks out can be vastly different. How it behaves when a new CEO steps in can be vastly different. So understanding how time is affecting outcomes is really critical.

Transformer models don't care about time, nor do diffusion models.

If I say, "Write this article as if you were Shakespeare," it sees Shakespeare as a style, not the fact that, you know, in the, the 14th century, there was a writer who wrote plays, et cetera. No, it doesn't.

It can repeat that back, but it doesn't understand the concept of time at all. It's just style. And these kind of models require different kinds of problem-solving capabilities.

So each one of the models is gonna be set up differently with data that works differently- To form an outcome that can make predictions more effectively.

When you founded Collectivei in 2008, you obviously thought about data, you also thought about AI, but how important was AI at that time? It's the only way to solve this problem. Mm.

It's funny, when you, Mark, when you think about it, companies don't want to share data. Deep learning enabled us to go to companies and say, "You can share your data, but in a way where the AI will never share it back."

It's a black box by its nature. So in the beginning, the thing that people were most concerned about, about deep learning, which is it can't explain what it knows, is also a benefit when you're giving proprietary data.

It means the AI can learn from everyone, but without ever disclosing anyone's information. It becomes a trusted third party so that everyone can get smarter. Now, you've, by the way, experienced something similar, right?

When you use Google Maps or Waze, it's learning from every driver on the road, but never disclose it who's on the road with you.

But you get the benefit of finding the shortest route, or avoiding the accident ahead by learning through this collectiveness.

And this is really what deep learning enabled for B2B companies, which is nobody wants to give up data if they can help it, but if they can all benefit from each other as a collective without disclosing the raw data and knowing that it will remain confidential, they're in a situation where everyone can get smarter and improve.

So data is obviously the oil that we need to train those models, and it also became clear that data authenticity and data traceability is becoming more important than ever. How do you solve that at Collective AI?

So in our world, we're able to solve it because we're going to companies that already-- It's their data, they own it, and we're able to go and talk to them about the rights, make sure that their consumers and their customers understand the rights, and we go through that process when we're onboarding a partner.

I think it, it's a lot harder and where things like blockchain, blockchain and other places can help is understanding who owns what data, the authenticity and the quality.

Remember, an AI can be trained on both good and bad data. And we remember this, people forget there used to be this Microsoft chatbot that came out called Tay.

Tay was chained, trained on, unfortunately, a lot of bad data, and it didn't take long before that chatbot turned into a neo-Nazi. It was pretty ugly.

And this is the stuff where we wanna make sure, just like you're raising a child, you're raising it with high-quality information, because the better the information is that trains the model, the better the model behaves.

And I think authenticity for most organizations is a challenge. We make sure we do that through our partnerships and direct connections.

Going back, um, to economic foundation models, something y-you mentioned is how they are connected to AI agents, another topic that has been very hot in recent months. What do you think about AI agents?

Where do you see potential, and how is it connected to economic foundation models? So agents are really where everyone's fixated on.

Agents are the idea that I can optimize almost every aspect of my life and my business by having agent that's seeking to optimize for whatever I value.

And today, agents start off as something as simple as doing reasoning, like tree of thought in, in your products so that the neural nets can actually start making decisions in the way you might want them to on your behalf, all the time, updating themselves.

All of these things are great.

I think when we're looking at things like payments or looking at things like buying products on the fly, on decisioning, even reviewing legal contracts, these are areas where agents will start to take over processes because they can both accelerate the speed of what people are doing, and they can remove transaction costs and do that well.

But you're right, there is a lot of fear that some of these agents will go off the deep end, and so this is a bit of the challenge.

Economic foundation models are important because the agent is learning to make decisions based on predictions that are being made for it.

Now, LLMs and diffusion models are predicting whether it'd be the next word or the next pixel, but they don't really understand what they're talking about.

So economic foundation models, on the other hand, have to understand the products you sell, the people you sell to, the business you're running, the cost structure that goes into it.

It really has to understand the world of business. And so there, it's able to make business decisions well ahead of where people would've. Can I give a simple example? Yes. All right. Sure.

Today, if you are a company that manufactures a product, you typically wait until the order comes in, someone signs an agreement and says, "Okay, now I'm gonna, I'm gonna get paid, and when I get paid, I'm gonna build this generator," let's say.

That is a extremely expensive thing because as soon as that order comes in, you have to rush to build it. And that, if it's a long cycle to build it, you, means you have to warehouse it and wait for it.

That can mean companies sit with a lot of inventory and things like that.

A neural net, on the other hand, using an economic foundation model like Collectivei can come to a conclusion that within the next six weeks, seven weeks, whatever it is,

this customer is so likely to buy that we should start manufacturing it now and ship it to them before the com-- before the product c- uh, gets purchased.

An agent's more likely to do that because it can make decisions really quickly based on what's happening in the world. Mm-hmm. And it can put itself in a position where it can say, "This is the most cost-effective way."

Now, let's imagine along the way that while it's shipping it out, someone cancels that order. It can immediately reroute it to the next best place.

That's something where you could have an agent do it automatically for you in a way that humans just probably- Wouldn't be able to do at the speed or with the certainty that an agent can do.

That's just one of many examples, but this is an example where studying how the world's doing business means you can react to wars, inflation changes, uh, CEOs getting fired or hired or whatever it is.

Y- it means the business can react dynamically in a way that it just can't today. How far are we away from a future like that? In many ways, it's here already, right?

We've been doing this now for years, and the fact that our model has, you know, over a decade's worth of information and training, its predictions are already so good.

Where a lot of companies will have to catch up is a lot of their infrastructure is still based on these old architectures. I've listened to you speak with other guests before.

You know that their architecture is-- I mean, think about the banking architectures already. Half these things are still on COBOL. There's a whole layer that's getting built

around Web3 that is basically just meant to say, "We're just gonna move on from that old world."

And I think a lot of that architecture needs to be there to make these interconnections work faster, more seamlessly with an AI.

Smart contracts are a great example of a place where, you know, if you-- if, if an agent is capable of doing something but the organization is still analog, it doesn't matter if you have great agents who can add a lot of value.

You need that end-to-end where the agent can actually do the work from beginning to end.

In a recent interview, Eric Schmidt mentioned that those AI agents will start talking to each other, will start speaking their own language, and at that point, we might have to pull the plug.

Do you see that similarly, or what's your perspective on that? I love Eric. I've known Eric forever. He is a smart man. I would never bet against Eric.

That said, I believe that when humans talk about the risks, we protect against them. Where we usually run into problem are unseen errors, unseen problems.

I think humanity is smart enough to build in the circuit breakers and the protections to make sure these things don't go out of control. Now, I've heard the arguments.

The arguments are, look, if it's an AI that's smart enough to know we're gonna do that, it can fool us for a long time or it can create a scenario where competitive natures drive you to look past it.

This is the idea that if we start using robots in wars, that, you know, to win the war, we'll start giving it more and more freedom until one day we wake up, it's out of control.

And I can't stop and say that can't happen. I've read too many dystopian novels in my life to believe that the chance is non-zero. Do I believe that we're in that world today? We are in an exponential curve.

It could happen. I don't want people thinking that that's a reason to slow down. Um, I think they're, they're more likely we'll find good ways to protect ourselves and live symbiotically with these, uh, great machines.

If anything, maybe it drives us to im- to use CRISPR and other technologies to make humans even smarter, so we keep beating the AI. We do have a better processor. All right.

So t-talking about AI agents, I also wanna know from you, how does blockchain play into all of that? Because at least in, in our crypto space, blockchain and AI are also hyped at the moment.

There are a lot of intersections that people see coming, you know, whether it's data authenticity or AI agents using structured data on-chain. How do you see that?

Look, I, I think what you're seeing, if anything, I think AI will accelerate the adoption of, of Web3 generally.

Mostly because the architectures that companies have run on for the last forty years have reached their max, and a lot of companies have really tried to figure out how to stretch it.

Take, for example, a old architecture like CRM. Right? Salesforce or Dynamics or any of these companies have been around forever. These are not architectures that are designed for AI.

They are old, they are difficult, they are customized, they are, are-- they struggle generally.

There is this AI movement will drive the adoption of chain, uh, of blockchain and other technologies because to get the full benefits of AI, those older architectures have to be upgraded.

I think you talked about some of the big ones, right? Uh, you definitely have to think about, uh, authenticity of data. The last thing you wanna do is get sued because it's unclear where the data came from.

I think you're talking about making sure people have ownership for shared value, right?

Even, uh, think about in our world, transactions have many partners involved, everyone who wants to make sure that they get a piece of the value in accordance with whatever they've agreed to.

I think you can talk about the economics of actually transferring value amongst many different players. All these areas are key places.

I don't see Web3 generally, uh, and when I say Web3, I'm talking about blockchain in general, is I hear a lot of people talking about training models using blockchain. Today, I don't think that's sufficient enough.

As a guy who spends his day thinking about pipelines of data and how to eke out incrementally more value for training your models, any inefficiency has to get cut out because the cost of training is still so high.

That could change, but right now I see the two pairing well and essentially upgrading an arc- a legacy architecture to really enable full autonomy of a business' operations.

I, I don't think it'll survive with the older infrastructure that exists. So still, uh, speaking about AI agents, a year ago, you mentioned in Forbes that all the traditional SaaS providers are terrified of GenAI.

What do you mean by that? [sighs] I think when you look at what GenAI has done It has completely transformed our view of what language can be capable of,

what images can be capable of, and now it's also driving what other types of information can be capable of.

I think when you look at the medical field, when you look at the finance field, when you look at all these gen AI models, and I think economic foundation models are a form of gen AI, what all these models are doing is changing the concept of how we think about information, insights, predictions, and work.

Everything we know is gone. It's just everything, uh, if you think about it, I mean, look, I would imagine, Mark, when you've had guests on in the past, and people are like, "Why is blockchain a big deal?"

They're not fully comprehending what you see every day, right?

You see that every foundational tenet that businesses have been built on can be changed, how the community can be in-involved, as opposed to one central control.

When you get into AI, what you start to realize is every way we think about how the world functions, underlying it changes because it's not about just c-collecting data anymore and us processing it.

It means that we can start bringing communities of people and really start operating in completely new ways.

I don't know if that helps give a good overview, but I think what I'm trying to say to your audience is people form opinions about AI without really getting deep and their hands into it.

I would argue that everybody today should be trying as many different AI tools and getting and playing with them in so many different ways to get their arms around the new logic that's emerging.

And I hope that gives a good sense of imagine having a common logic across your whole org and a speed to get to the, the best outcomes that no one's ever seen before.

That does change the way you think about operating every part of your business, every part of your life.

It is a-- We're entering a paradigm, I think, that really will be meaningfully change in our lives, that we'll look back and think we can never live again without it.

I think the iPhone was probably the last paradigm that was as big as this.

If you look at other big tech companies or companies building sales tools like Salesforce, how are they comparing to what you're doing at Collective AI? Did they realize that opportunity?

Are they working on economic foundation models? Will they be offering that to their business customers as well? I think this is, um,

I think when I look back at Salesforce in particular, and Mark Benioff is an amazing leader, um, but he's in a paradigm sh-shift where he's struggling to understand the paradigm shift. For a guy who destroyed Siebel

by moving to the cloud and making himself cloud native only, he's behaving the way Tom Siebel did with his business, by really trying to keep the older AI paradigm.

And the older AI paradigm is more this simple machine learning. So in the Salesforce model, every customer's data is used to only analyze themselves.

Think of, imagine if knowledge was always kept to yourself, never shared. Imagine if you couldn't learn across the world, you could only learn from yourself. How limiting that is.

And the reason that paradigm exists is because Salesforce was successful by moving people in the cloud from on-prem by telling them that their data would never be usable by anyone else.

Yes, it would be in the cloud, but it will remain siloed for yourself. So their entire paradigm is to take your data and try to understand your data to make better predictions for you.

Those aren't great ways of getting answers. I'm sorry. You end up with really low-quality data, and you end up with low-quality answers, and that's why Einstein's really struggled so hard.

I think when you look at an economic foundation model, what you realize is we're moving away from simple machine learning, regression models, et cetera, to deep learning, and we're learning across the entire network of our clients, so we're able to see the same buyers across multiple sellers and get much better at predicting what they're doing.

AI is about this deep learning paradigm. That's really when people are talking about AI today. They're not talking about regression. They're really talking about these models that are getting super intelligent,

and I think that's what the main difference between us and everyone else.

So you're, you're talking about economic foundation model, which means you are looking at much, much, much more data i-outside of the data that exists within an organization.

When you do that, how do you make sure that the data you look at is the right data? So we have a patent called the Universal Data Model.

So when a customer joins us, we actually denormalize their data to a way that the AI can learn across all the different industries, across all the different regions across the world, and that's that data setup that we're doing.

So in the same way that language models had to go in and find good data, label it, annotate it, structure it in a way for training, we do the same thing, but around financial data.

And so as customers feed data into us, and it's continuous feed, their, our training data that we get from our, from our customers feeds this overall model, making it smarter.

It literally is studying how the globe is doing business every single day by literally observing every transaction across our entire, the entire network. It's not siloed by customer.

We literally study across the whole network.

And so it's as if you have a God mode watching the way the world is operating, and then that enables it to learn more about what's happening with the economy, down to individual customers, um, down to individual deals, and that really is what makes the whole model so predictable.

It's amazing. Steven, um, before we close this session, let's take a step back again, and let's look a couple of years into the future.

If you look ahead Into the next, let's say, five to 10 years, what do you think will be the most significant changes that we'll see in AI? If I were to make a prediction, I think five years from now, our closest friend

will be our AI friend who's helping us with our day-to-day. I think it'll become a part of our family.

Our AI will help us become successful in our careers, our AI will help us manage our families more effectively, our AI will be helping us make sure our health is getting better.

It'll communicate with our doctors where our problems are and, and help me prevent problems from happening. I believe within five years it will be a member of your family.

That is, is something that will be life-changing for all of us, and I hope for the better, if we do our job well. Yeah. That's, uh, actually another question I wanted to ask you. You said, "I hope for the better."

Are you a techno-optimist? Having been through more AI tech revolutions that have, have started off with the joy and the dream of doing great things, only to find they've been misused.

I'm an optimist that the people who are building these things are aiming for a better world. I worry that the world is so busy creating havoc for itself that it outweighs us, and I worry about that all the time.

I hope that I make a world dramatically better, and I hope my peers do the same. It's gonna definitely require a lot for us to make sure we don't screw it up. So yes, I hope I'm an optimist. [laughs] Great.

With a line of fear underneath. Yeah. All right. Another question, Steven, I wanted to ask you, and I know it's something that's always coming up when talking about AI, is job replacements. Yeah.

Do we need to fear job replacements? What's gonna be the future of work when we have all those AI agents doing stuff for us? Uh, how do you see that unfolding?

Look, it-- what worries me is the lack of thinking about upskilling.

Um, you know, when we went through the Industrial Revolution, there was a huge number of people who moved from farming to manufacturing without any training, without any skills, without any jump, and that led to two world wars, and that, that is something we need to keep our, our eye on.

We at Collectivei have an entire group focused on responsible AI, where we focus on upskilling people for the next century. Um, and I believe that that is a responsibility that every AI company has.

Whether jobs are lost or gained, history would say that with each new revolution, there are more new jobs that we never dreamed existed before, but we have to help people achieve that goal.

It's amazing to see how fast AI has been adopted, but I also, when I go into every company, the first thing I hear is fear. And we're not gonna stop this revolution from moving forward.

It continues regardless of fear or not. But the way to destroy fear is to train people and give them a path forward, so they can see where their future lies. I don't think

when social media sprang up, that people thought influencer would've been a job, yet alone one that can create billionaires, and yet it did.

And our job is to help create those billionaires who are building these AI companies.

If you were in an elevator right now with a CEO of a large company, and you had the chance to them to talk about Collectivei, what would you tell them in a short elevator pitch?

I-- You know, in that short elevator pitch, I would probably tell them that the fact that, that we're not already working with them is probably something that they should be afraid of more than anything else.

Um, the speed of adoption is so fast, and in our model, competitive-- world, worlds are so competitive that if their co- co- uh, competitor is using me already, I probably will either have their business because my competitor will take it, or, or they're gonna figure out that they probably better need me.

Imagine FedEx working without Waze. Uh, you know, if UPS had Waze, UPS is gonna deliver packages faster and on time and avoid more accidents, and it's only so long before FedEx would disappear.

So it's a game now of getting to optimize your business as fast as possible, whether-- If you do it, great. If not, your competitor will. All right, Steven, we're almost at the end.

Before we end this, a very quick lightning round. I do this with all the guests. Uh, short questions, very quick answers. First one, your favorite LLM for everyday use. Anthropic. I like Claude.

One encounter with AI that blew your mind. [sighs] I was with a client last week, and they asked-- Our product is called Intelligence.

So they asked Intelligence how all the macro events have affected their business, and within 30 seconds, it had spotted events from around the globe that I didn't even know happened.

And the person looking at it said, "Oh my God, I remember that. That did have a huge impact on our business."

It was seconds before the AI found all the factors that were affecting their business and then predicted what would happen based on the election for their business. I thought that was crazy. I had never seen that before.

Wow. That's super cool. Um, then next one, AGI, Artificial General Intelligence, sooner than we think or pipe dream for now? S- Depends on how you think. I, I think within the next five, 10 years we'll have it.

So if you were thinking 20, 25 years, I think it's probably sooner than that. I don't think it's a pipe dream. I think we'll probably have to keep recalculating what we think AGI is to push it out further, 'cause

I think we can still do so many more things than this AI can do, and that the barrier to looking like AGI is still the Turing test, which is too low. Amazing. Steven, thanks for coming on the show.

It was a pleasure to host you. Uh, if you wanna find out more about Collectivei, go to collectivei.com. Any other places people can find you, Steven? Yeah, we do an event every week called Forecast at ciforecast.com.

I highly recommend joining. It's a 90-minute ask, uh, people anything. It's funny, Eric was on there.

Um, we have, we have some of the best people, including Yann LeCun, who've spoken, um, and Yoshua Bengio, and some of the other greats are there. Highly recommend coming. It's free. It costs nothing. It is, um, uh...

For your audience, and I-- can I just fanboy for a second? 51 Insights is a, is my m- must listen. And, um, and I know you have some of the best people in your audience.

I'm sure they would love CI Forecast 'cause you just keep people on the cutting edge.

And, um, I wanna thank you for having me here today 'cause, uh, as a fanboy, uh, this is, uh, this was important for, for us, and I'm excited to be here. Thank, thank you, Steven. The pleasure is mine.

A pleasure to have you on the show, and all the best. My pleasure. Thank you. [upbeat music]