Episode Transcript
[00:00:00] For the last couple of weeks, there's been a lot of questions about where the business models of the frontier vendors are going to go and how they prevent their models from doing bad behavior and whether they're really intelligent or not. And while that's been going on, we've been doing a massive amount of optimization of Galileo and learning something that I think is very profound, and that is as follows. In our particular case, we have a massive library of very narrow, highly interconnected information about management, leadership, recruiting, HR and related technologies. And it's not just science like numbers, and it's not just opinions like blogs or New York Times articles. It's fact from real stories of real companies and conclusions that we draw from thousands of observations and meetings and surveys. So it's a form of applied knowledge or applied science. I always think of it as social science based on tens of thousands, thousands of interviews and conversations and surveys with real companies correlated back to a given industry, company size or role in a company. So it's a very powerful corpus of knowledge. You know, you may think it's just a bunch of reports, but the reports are very interleaved and connected. And because we've had 25 years or more of common taxonomy in our library, it's very easy to draw conclusions and, and see how some of these trends change. So how to recruit people and what technology to use and how to decide how to pay people and how to be a good manager and how to organize a company and how to drive change and how to be more innovative and how to save money. I mean, those are all things that people think they know, but under different conditions, the correct or best solutions are different. Just like for a doctor who has to diagnose a patient, two patients with the same symptoms might need very, very different diagnoses and very different solutions because of their age or their body or the true underlying problem. Because when you say any of those business issues like who should we hire or how should we save money or how should we grow, it's unique to that company. There's no general solution for that, even though you could read a book on it. So that's what we've done. And as we've built a new structure for the content, which we're going to talk about in the next release, we found that the level of intelligence is going up by orders of magnitude because of the purity of the content and the way we've arranged it. And this gets to this issue of how AI works. If you point your large language model at the whole Internet and It tries to index the whole thing. It's very blurry. It sees fuzzy things. It sees lots of duplicative information that contradicts with each other. And it has no way of knowing which one is correct or inflammatory or perhaps just an advertisement or an opinion or perhaps deliberately incorrect or perhaps written in a form that sounds correct, but the writer didn't know how to write it clearly. So it interpret. The AI would interpret it differently. So these embeddings become very, very large and very imprecise. And that's why all of the models seem to feel and sound and talk the same. They're all homogenizing, very unique points of view and averaging a world of experts. It's really kind of silly in many ways. We honor experts and scientists and gurus and senior leaders and, you know, people that know a lot about different things. And then we throw it into AI and it all becomes homogenized into this big mush. And it. The AI is. The reason the AI seems so obsequious or kind of bland is that it doesn't know. First of all, it doesn't know anything. All it is is statistical model. It doesn't really have a brain, but it doesn't have any way of differentiating a unique, highly powerful insight from something else. Now, the way Google did this in their search is they just looked at how many people were pointing to it and therefore assumed that if a whole bunch of people are pointing to this, it must be really, really good. But that gave way to SEO and people gamed the system and stupid stuff became very, very popular because it was either entertaining or highly optimized for Google. And so the same thing's gonna happen here where if these guys go public with this stuff and they come up with a monetization strategy, people who don't mind sharing their content are going to optimize their content so that the Frontier Labs find their stuff first. Now, I know the New York Times is suing OpenAI.
[00:04:50] I know that News Corp has licensed their content to OpenAI. And I think what they're thinking to themselves, I think News Corp. This is basically, this is money. We're the source and they're the distribution. So they're basically saying, we need channels of distribution, and if we can get paid for all the hard work we do generating content, great. And they're big enough to get away with that. Although the revenue they're getting from the Frontier Labs is not that high compared to the subscription revenue they would get by selling their own magazines. So, you know, I think The News Corp people are very smart and they know what they're doing. But there's a fairly big risk, in my opinion, that it just doesn't work out that well for them in the long run, because all OpenAI has to do is attach some other content in and give it a slightly higher priority because somebody else is paying them more money and News Corp gets demoted. You don't want to lose the control of your relationship with the customer. I learned this many, many years ago. That's the reason we don't put a lot of our content out into these systems. We, we sell it uniquely to companies with our own services around it, because we can frankly make more money doing that and provide more value than just giving you a whole bunch of stuff to find on the Internet that gets homogenized in everything else. Now, if you sort of understand this and think about it, you're going to see a lot of this come out when we launch next, the next version of Galileo. Then you think to yourself, we got two companies that are worth a trillion dollars, maybe three, maybe four. Google, Amazon, OpenAI, Anthropic. And then Microsoft's playing a different game. Microsoft's playing more the game we are. They're selling tools, and they want you to build your own AI with the tools. They're not trying to get into the content business per se. And so where would they go? Well, if the Internet of AI goes the same way as the Internet of non AI, you would anticipate that one of these companies would become so good at finding stuff that they would be the expert, but that's not actually how it worked. What happened in the Internet was we had Kayak and then Expedia and then Travelocity and then Airbnb, each of which became very, very financially successful companies simply focused on travel. And they all jockeyed with each other and bought each other and merged with each other. And eventually that became a massive market. Now we have booking.com and others, but they moved down the value curve with their content. They didn't just produce the content to make it easy to find. They put products and services around it so you could book a flight, not just find a flight. And you notice that Google didn't do that. So Google makes a small piece when you click on a link, but they don't want to get into the booking business themselves, just like they didn't want to get into the recruiting business. They tried that too, and stopped because it was too much work, because they can make money just by selling links. So that's Number one. So that would imply that where this is going to go and this is the way I see the world, is there's going to be AIs that are very specialized. Galileo will be an expert at the labor market and management and leadership in HR. There'll be Harvey or somebody that's really good at US labor law or US law, and there'll be other AIs that are really, really, really good at certain things. And those companies will have to spend a lot of their time making sure that their content is really well optimized and really accurate and up to date. And they don't care. They won't care too much about what LLM they use. They'll use whatever one they can find that's reasonably cost effective. And then it will mimic the real world where we will go to experts and those experts will be using somewhat off the shelf tools and those experts will get better and better and better. And there's no way in my mind that OpenAI or Anthropic or even even Google could do that. They just don't have the business model or the interest in getting into all these unique domains. So they become software companies, search engines, advertising companies, and very, very good tools that we use for a lot of things in the consumer life. But they don't become experts. And then these vertical expert companies could, when they get really good, license their expertise to the bigger frontier vendors at a very high price, not a low price, because they'll be uniquely powerful. And, and I'm not trying to just push the agenda that we have. I'm really just reflecting on what we've learned. We have a pretty good sized business now and everything that we do in HR. You could go to ChatGPT and just look it up and just do whatever ChatGPT says. But what you'd find is that it's pretty hard to make it into a really actionable solution because it's not clear if it's correct or relevant to your situation.
[00:09:33] So the value of getting a general solution is okay, but not that great. Just like going to a doctor who's not a specialist. When you have a unique situation or a unique problem, it actually could hurt you because that doctor could easily give you the wrong advice. And your unique need could be not only unfulfilled, but perhaps harmed by you believing you got good advice when you didn't. And these doggone things are so nice to you that they very happily tell you things that are not really correct or accurate and they don't really apologize or learn when you tell them something's wrong and there's no way that they can because they don't know who you are. So there's no way the LLM could know that you're an expert and what you say is correct because they don't know who you are.
[00:10:21] And for every one of us that are expert in certain things, there's probably 10 people that just like gaming these things just to kind of cause havoc or make money through hacking. So I see no way that this doesn't go into a more vertical market. Now that opens up a lot of doors. First of all, it's a little bit threatening to the frontier guys because they put a lot of money into these software tools and they want to monetize them in a big way and they'll be fine because there's such a big market for the sort of general search stuff that they do. But I think there's going to be much more money spent and much more value created by on the domain specific AIs that are built by people like us or companies. I was actually talking to somebody about this the other day who works for a data labeling company. And I was telling this person that I think the data labeling industry is a little bit of a dead end. And you know, the person said, well, why? I said, well, let me give you an example. Let's suppose you're Ford Motor Company. I just made this up. I wasn't even thinking about it. You're Ford Motor Company and you want to build an AI that allows all of your dealers and repair people to take a video of any one of your cars and diagnose the parts that need to be replaced, the procedures that need to be fixed, and the other perhaps tuning things or maintenance things that need to be done from the images. If you're Ford and you know all your cars and you have parts inventories and photos and diagnostic information and engineering information on every part in every one of your cars, you might decide to build something like that. It might take you a few years, but once you built it, it would be spectacularly useful. You could sell it to mechanics, you could sell it to customers, you could use it in your engineering organization to decide how to design things better. And as your AI gets smarter about physics and heat transfer and power, your AI would become better and better and better at helping you engineer your cars. By the way, that's what we're doing with Galileo. We're now using Galileo to decide what we need to write. We don't just throw stuff in there and cross our fingers anymore. Galileo is teaching us or telling us what it's missing. So it's suddenly becoming a much, much more strategic tool. Now, if that Ford Motor thing happens, which it will, I mean, absolutely will, I'm sure the pharmaceutical companies are doing this. And you know, customer service, by the way, this is why I think these generic customer service tools, like the stuff that Salesforce sells, are really not that useful. I don't want a generalized customer service bot. I want a customer service bot that knows everything about company and what we do and every product we sell that can give the customer a very pinpointed answer to a question and help them buy something else, not just say, hey, how's it going today? How do you feel about your life? And did you like my service? It's not a generic problem. I mean, if you're not using the AI to go vertical, you're probably not really getting that much value out of it. And I think we all got very enamored with the, you know, initial human, like, experience of the AI. But very quickly, after a few weeks of playing with one of these things and you realize it's not that useful. You just stop using it. Because it's nice to talk to somebody who actually is listening and can give you some value. But somebody who just mimics back answers isn't as useful.
[00:13:41] The other thing, of course that happens is these vertical AIs inside of companies become a huge market for the vendors. So this is where Microsoft's going. They call it. They don't call it Frontier. They have another name for it, but they are really seeing the market the way I am, which is that I want to help each one of our clients implement the human capital HR system of the future, which we call HR 2030, that knows so much about their company and so much about their people that it's telling the CEO and the heads of HR and the heads of different business units what to do, who to hire, how to train people, how to optimize sales, how to create better services, how to improve retention, how to create better frontline services, et cetera. We, it can do that. I mean, we know it can do that. We're, we're seeing it do that. And I think you apply that to supply chain finance, operations, sales, marketing. And we have this new market of AI consulting tools in a way that are used by companies to build groundbreaking individual solutions. Where do the Frontier vendors go in this? They don't really play. They're more like the Google search engine. And I think from the years I'VE been doing enterprise stuff. Very, very, very few companies succeed at consumer businesses and enterprise businesses because they're so different. In an enterprise business, you have to get to know your clients and you have to build products and services to help them create great things and teach them how to use your stuff so that they can build great things of their own. The consumer market is completely different. You need to answer questions with no support because there's no way you can support millions of consumers. And you have to build things that are instantly valuable so that so the impatient consumer doesn't just wander off and disappear. Very different kind of business problem, different business model. So this sort of academic, very scientific debate about reinforcement learning and self learning and how fast these things learned, I don't think that's really the issue. The idea of general superintelligence is that a superintelligent AI is so smart that it could learn anything. Well, from what, what is it going to learn? From what content is it going to use? Playing chess a billion times to learn how to play chess is a very tiny little thing. If Elon Musk really did capture the accident history and driving history and videos of every Tesla in ever built and put that into an AI, he would have some sort of a database that would understand the patterns and needs of cars, not, not drivers of cars, because this is about transportation, not the driver, and what they should and shouldn't do under different conditions. And so, you know, that is the value of the data in say, a car business. But you could do the repair. There's many, many examples. And I think the ability for the AI to learn from its own history is kind of a commodity. I'm not saying it's easy. It's obviously extremely complicated and there are going to be different implementations of that. But if it's not getting good data, the fact that it knows how to learn, isn't really that the value? The value is the data. And so, you know, in the case of the Elon example, how many Teslas are driven in the snow? Not very many, because the battery doesn't work well in the cold. So my guess is that the Tesla AI doesn't know a lot about snow, and it may never know a lot about snow until they have a lot of cars driving through snow. You could talk about rain, you could talk about narrow streets in San Francisco. Those are different kinds of problems for cars. They're similar, but they're different. So where this is really going to go, and this is why I'm so excited about it, is you're going to be able to specialize with your AI and focus it on a problem that's very unique to your company. And in the case of HR, the 2030 Reference Blueprint, which is pretty much done now, is a spectacular tool that'll show you what's possible.
[00:17:47] And over the next four or five years, and I think it'll take that long, we're going to have dynamic enablement systems that really, really help our people and our companies grow with strong human capital, expertise and intelligence behind them, just like we will in supply chain and we will in finance and we will in other forms of business operations. I know this was a little bit of a pondering podcast, but I really wanted to write this down while I was thinking about it over the weekend. And I hope you guys find this interesting and I'm more than happy to debate this with anybody who's an AI engineer and talk through and compare notes on what we've been doing versus what they're doing in their engineering projects too. That's it for now.
