Episode Transcript
[00:00:00] Today I want to talk about all the hype about AI and whether maybe we should be a little bit bearish on these big frontier model vendors. And particularly I'm talking about OpenAI and anthropic. Most of you know that Anthropic got significantly hammered by the federal government last week because their model was jailbreaking systems. And Amazon tattletaled on them to the feds and told them to shut it down, which they did. And Anthropic then sheepishly took it down. There was also a whole series of lawsuits by the state of Florida and other states against OpenAI for disclosing information and other business practices. You know, this is fairly typical when a company goes public. They have a lot of money, so there's a lot of people going after them. But I want to sort of give you something to think about in the enterprise space that occurs to me while we're hanging around this weekend. In the early days of the Internet, the predominant players were a company called Excite, AOL and Yahoo. And those companies pretty much thrived for almost four years before Google was even founded. And I just looked it up. They were each 250 to $300 million companies by the time Google started, which is probably the equivalent of a billion today. So, so they were pretty good sized companies and they were getting massive amounts of traffic to these web portals for search, but they couldn't personalize ads, so the ads were banners and they couldn't monetize the ads very well. And what Google did through the PageRank algorithm and deeper indexing technology and a whole bunch of other things was change the paradigm to be a highly personalized experience which then allowed Google to sell personalized ads, which then allowed Google to become a, a huge company for many, many reasons. Better experience for users because you could have a very simple experience that was personalized to you. Better experience for advertisers and companies because they could pinpoint an audience and spend money to go after them and sell them things, which led to many, many other businesses, of course, that Google's created. Excite is gone, AOL is gone, Yahoo's around, but they're not doing much. They're part of Verizon, I believe. So here we are three and a half years into open AI and Anthropic, and we might be in a similar situation. Here's why I even bring this up. The technology that we're using for large language models for all of these AI applications is probabilistic technology around language. It's called a large language model because it's treating bits of code like language. And so even the photos that it creates are probabilistically created based on the probability of this bit being next to that bit, or this word being next to that word, or this letter being next to that letter. And for simple applications, this is just spectacular. You get this human, like experience where the AI can respond and generate things and you can tell it what to do, and it can look for probabilistic, likely replicas of the thing you're trying to do. And it's extremely fascinating and useful. It's very useful for looking things up, for education, for training, for problem solving, for writing, for editing, for summarizing information, for creating podcasts from text, for creating videos, for creating images. And all of those are real problems that we have in companies.
[00:03:43] And we need real solutions, just like the type ahead and automatic spell checkers we have on our phones and on our computers. Things, those are very useful tools. We don't consider them to be AI, but they are. I had an interesting thing that happened this weekend. Someone hacked into our American Express account in the company and purchased a pump and a massive tool for tearing down houses. And that particular stuff that they bought was shipped to a house in Salinas. And all this happened online. The American Express people canceled the card and we reported them. And I was sort of laughing to myself. If you looked at all the things that I've bought on American Express for the last 25 years, I've never bought anything that looks like a pump or a hacking tool to not to knock down a house. Yet the AI in American Express didn't notice that this was a very, very strange purchase. It should have, I think, but, you know, I don't know how it works. And it's been around a long time. So I assume American Express has a lot of security around this. So what I'm getting at is that AI is way far from perfect. This super intelligent stuff, I don't get it. I'm not sure why people keep talking about it, because there are many, many things that these systems are not optimized to do. I mean, what they do do, they're good at, but they're not good at everything. And we tend to have this generalization belief that if it's good at this, it's going to be good at that. But, you know, that's not actually the way humans are either. Mathematicians aren't necessarily good at writing poetry. Some of them are, but some of them are not. People who are good at singing may be good at Dancing, but they might not. So anyway, so we have these $2 trillion market cap companies about to go public and then we have Microsoft and Google. And I would put Microsoft and Google into another category because they're very large sustainable companies where AI can complement their business. By the way, this is true for Workday, Oracle, SAP, UKG too. They can generate revenue from AI by adding value in different ways than selling a frontier model. But the two big frontier model vendors are OpenAI and Anthropic. Thinking about the enterprise market, I have a thesis which I will write up that maybe they don't turn out to be as big as we think. And the reason I say that is as follows, going in the very similar vein to the Google displacement of Excite and Yahoo. When you think about what UA want AI to do in a company, of course we want it to look things up and serve as a service center chatbot and provide training and provide knowledge and show me compliance rules and handle customer inquiries and look things up in the customer support database. And you know, maybe it can have a conversation with people to help them understand a job for recruiting. I mean there's lots and lots of good things it can do. But over time those features, in fact very soon those features in the next year or two will be as commoditized as the spell checker.
[00:06:48] They'll be common, they'll be ubiquitous and we will assume they're there. And I'm not saying they won't be good. They will be good, but they won't differentiate you from anybody else. Just like the spell checker doesn't really do that either. What's really going to differentiate your company is the vertical specific solutions to your company's business. And I noticed that Satya is starting to talk about this at Microsoft a bit. But I firmly am in this camp that the real value of AI, I mean business value, not cost reduction value, is going to be different. It's going to be a system that can model highly complex numeric data or alphanumeric, but probably numeric, and, and show you what you can do to run your company better. Let me give you an example. I talked about this at the conference and this is an example that everybody could relate to. I've got a hundred people in the salesforce spread out around the world calling on different clients. 20 of them are killing it, doing really well. 50 of them are doing okay. They're sort of fair to maybe mid level performing compared to last year. And 30 of them are in bad shape. They're behind, they're selling to accounts that are not profitable, they're just not hitting their numbers. And this is probably a typical distribution in any company. Now, the most natural questions you would ask, you could ask a lot of questions. One question you could ask is what can I do to take the top 20% and get them to sell twice as much? Which is probably a good question to ask because they're already good and, you know, you could probably just support them and ignore everybody else, or you could try to take the people in the middle and turn them into high performers or, or you can figure out why those people at the bottom are having problems because you might learn something that helps everybody. So let's assume you take the third approach and you ask the AI explain to me why you think these 30 people are underperforming. Now, it doesn't know anything about your company or your business or nothing. So what it's going to do, and I know this, it's going to give you a very generic solution. It's going to say onboarding, training, leadership, pay, employee experience, productivity, you know, enablement, et cetera. When in reality in your company, it might be something completely different. It might be that they're in a part of the world where there's a really fierce competitor that no one else has. Or it might be that the market they're going after, the industries or the types of companies are going after are just not suitable for your product. Or it might be that the product that you're selling, that they're selling in their spheres of the world is overpriced because they're selling in a geography or geographies that are more price sensitive than they were last year. Or, you know, I mean, there could be a hundred reasons for these 30 people to be underperforming. Of course, maybe they're just the wrong hire, poorly trained, their managers aren't paying attention to them, they're having family problems at home. I mean, there's just, you know, many, many things. But I want to know what they are because that's what it takes to run a company well as we. That's what operational excellence is all about. So in the theme of dynamic enablement for growth, that's what we want our agents to do. So what we really want our HR 2030 agents to do is identify this anomaly, first of all. So know that it's happening, because sometimes we don't know and isolate the 30 people that are in this group and start doing analysis based on historic data on what we know about these 30 people. Whatever it may be. And what in the past would those characteristics or attributes possibly infer would lead to lower performance? Perhaps we hired them with the wrong background. Perhaps they're in the wrong educational background. Perhaps their pay is not keeping up with the pay and other parts of the company. Perhaps their managers are all young. Perhaps their managers are all old. Perhaps their managers have different characteristics and backgrounds than the other people's managers. Look at the skills and capabilities of their managers or their individuals or their senior managers. Sometimes it's senior leadership. I mean, one of the studies of sales productivity that I found interesting was actually a pretty important one that in one of the companies that did a big analysis like this, they found that the salespeople that were selling the most and their particular company had more internal relationships inside the company. They knew more people to help them. They were really good at internal collaboration, which is something that is important in sales, getting other people in the company to help you work with customers. So anyway, you would want the agents to ripple through different data sets and give you suggestions and recommendations on what you could do. Now, there's two problems with this today. Number one, we don't have any agents that do this. So it's got to get done by hand. And you know, the people, analytics people in a small group, and they can't do analysis of everything all the time. And they can't do it in real time either. So there's some of that going on, of course, but we don't have time to do it a lot. But the agents could, and that's certainly part of HR 2030. But then there's another problem that's actually a lot bigger and that is that the horizontal LLMs were never designed to do this. They don't work this way. If you look at what they do, they don't have Windows data Windows that are big enough to analyze this much data, especially if you're going to do historic analysis. Most companies have lots of history of sales and customers and products and marketing programs and so forth. By the way, the other thing that could happen is maybe the marketing department in that in this group where the 30 people happen to be just sucks and they're not doing any marketing, so they're not getting any leads. So there's a lot of things involved here. So what we want the HR 2030 world of agents to do is to look at this systemically and say for this anomaly that we don't want to deal with and we want to fix, what are the statistically correlated or caused factors that we believe would improve this situation. And once we know what those are statistically, we can literally tell managers, suggest to managers to change the characteristics or behaviors or attributes that are causing those problems. Because once we know what causes the problem solving, it's usually not that hard. But there's a lot of spray and pray, let's try this, let's try that, let's have more meetings, let's have more QBRs, whatever. That seem to be the normal behavior. Now, it turns out when you dig into this, and I've been digging into this over the weekend, the LLMs that we have today are not designed for this. They just aren't designed for this kind of numeric analysis. And also they're not vertical. So we know, for example in Galileo, Galileo very, very vertical, it's very smart about a narrow range of data skills, hiring, recruiting, pay, training, leadership, that kind of stuff. It's not an expert on supply chain management, it's not an expert on telecommunications network analysis or whatever. It's just. But it is really smart about human capital stuff. So you take a vertically intelligent system like ours and you give it a bunch of data and it actually can make very smart conclusions. By the way, I've tried all this prompting in Claude and OpenAI to do these kinds of things. It's not very good at it. So we are now in the middle of showing off and we did this in the conference last week, a demo of Galileo on a real company. And what we've been doing, and we've been demonstrating this now to more and more of you, he's showing you what you can do with Galileo with these kinds of problems. And we're going to keep pushing on this and we're going to basically but financial data and human capital data into Galileo and show you that in a real world situation where Galileo is connected to your real data, it can give you spectacularly intelligent recommendations, solutions and advice on how to improve the performance of your company. And that gets back to my theme of what are we trying to accomplish in hr? If you think about your job in HR as dynamic enablement for growth, then all of these other things we do, training, onboarding, compliance, pay equity analysis, dei, et cetera, are really means to that end. And as we design and build agents, we should be thinking along the lines of what will enable the company and the individuals to grow. Now, going back to the very beginning of this podcast, what does this say for OpenAI and Anthropic? Well, there's no question that those two companies are going to go public. There's going to be a spectacular valuation. We're going to see how profitable they are, or maybe are not, and there'll be a lot of analysis and there'll be a lot of people with a lot of opinions. But I do have this funny, nagging feeling that we're in a little bit of the same situation we were in 2000 when we had Yahoo and Excite. Maybe, just maybe, the real enterprise value of AI is going to come from somewhere else. And I'm not going to tell you where I think that is, because I do have some pretty informed ideas. But it could be that the companies that are in the market today are doing a spectacular job of the first generation here, which we need. But there's another generation to come that's going to move way deeper, in a more vertical way than we've ever seen before. And that's where I think the future value is going to be. Even in the consumer world, if you're out there trying to book a hotel or a resort or a flight or a cruise, and you get a generic answer with a whole list of hotels or cruises, that's not really what you want. What you really want is the answer highly attuned to your exact personal situation. And that is exactly the same thing with a company. No two companies are the same. So we want these models, whatever technology they are, to understand what we do and how we operate and what makes us successful and what our glitches are in our company, and then teach us through statistics, through data, or through external benchmarking, what we can do to be better. And that's going to be the type of solution where the cost of the tokens is going to be tiny. We're not going to care because the benefit is so high. Okay, that's a little bit of food for thought for today. I hope you guys are having a nice weekend, and we'll talk to you next week.
