Episode Transcript
Speaker 1 00:00:09 Hello everybody. This week I want to talk about the big topic that everybody's talking about, which is ai. And the reason I want to bring this up, we're gonna be introducing a very comprehensive white paper based on dozens and dozens of interviews with vendors on the three generations of AI in the HR tech market. And so I wanna preview that for you now, and you will hear a lot more about this as we get the white paper published. And we introduce a lot of interesting things at our conference in mid to late June. So first of all, let me just sort of do a little grounding here. AI is a very multidisciplinary area of computer science. It's 45 or 50 years old. It started with things like natural language processing, expert systems were big when I was younger, visual image recognition, audio recognition, and then eventually moved to predictive analytics, machine learning, and then large language models and neural networks.
Speaker 1 00:01:11 But the fundamental idea of AI is that is a completely different computing paradigm than all of the computing you've been thinking about. Every computer you've ever used, whether it be a PC or a mainframe or a web-based system, is designed by software engineers to capture information and conduct transactions on transactional data. So you type information into it or it imports information from somewhere else or from a resume or wherever it may be, stores it in a database, a relational database, and you pull that information back up and analyze it as part of your business processing. And so the code that was written by the engineers is designed to give you a user interface and a set of workflows and business rules to capture that information. The data is an outcome of that processing and the data grows over time, but it's usually not that big.
Speaker 1 00:02:04 And as the system gets used more and more and more, the data becomes more and more interesting. AI is the complete opposite of this. The idea of an AI system or platform or application is that it starts with the data and uses predictive analytics, neural networking, and other techniques to classify, understand, predict, and eventually generate information from that data. In other words, it's a data system first, a transaction system. Second, in fact, it may not be a transaction system at all. So if you look at the payroll or learning management systems or applicant tracking systems or HCM systems that you have all over your company today, none of them were designed for ai. They might have machine learning algorithms that try to learn from the data that is collected, and they're typically used to recommend content, recommend courses, and maybe recommend job candidates. But those are AI features or AI algorithms added onto those systems.
Speaker 1 00:03:06 They weren't designed at an AI core. Well, among the many, many fascinating innovations that have happened in ai, and it's really an interesting domain, I've been spending a lot of time learning about it. The area of AI that is now red hot, white hot is neural networks. And a neural network is a mathematical algorithm that takes a whole bunch of data and uses the segmentation of the data into small what are called neurons or pass and stages and tries to use mathematical algorithms to classify and predict what that data might be about. So suppose you have a big database of images and you train the system and you say, of these 500,000 images, here's 500 of 'em that are cats. Can you figure out if there are any other cats in there? And so what the neural network will essentially do is it will look at the bits, literally the bits on those 500,000 images and it will use a neural networking algorithm to try to figure out, while it knows that there are 500 of them that are cats, and it will say, what parameters and weights and balances do we have to apply to this neural network to try to predict that these cats are cats?
Speaker 1 00:04:25 Because if we can figure out the weights and balances in the neural network that correctly predict these 500 cats, then we should be able to find other cats. And based on what I've learned, most people didn't know if this was gonna work. And neural networks weren't super successful until we had very, very large high-powered computers and discovered that the mathematical algorithms that compute these weights and balances actually get better and better and better with scale. And so the neural network was one of the first algorithms that could scale and got smarter and smarter with more and more data and therefore more and more processing. So the way I like to think of a neural network is this essentially like a, you know, spreadsheet in a way that has this model in it that's fairly easy to understand, but very, very large. It has billions of parameters, billions of tuning knobs.
Speaker 1 00:05:16 And what it's doing is it's playing with the tuning knobs, using calculus to try to figure out which weights and balances are going to best characterize and predict or whatever the process is you're trying to accomplish on the data that you have. What the process of training does is it takes a known database of classified information and it trains these weights and balances to figure out, uh, what that pattern should be. So in the case of open AI, G P T three, G P T four, it has been trained on the English language and it has studied what are called tokens, which are words or pieces of fragments of words like i n G might be a token or, and it starts to figure out how these words go together in various statistical frequencies and gets essentially smarter and smarter and smarter by making these weights and balances better and better and better.
Speaker 1 00:06:11 And there are a billion of them or more. So you can't look at the model visually as a human being and understand what these different weights and balances are doing. It's impossible. That's why explainability isn't so easy. You actually need an AI to look at the AI to figure out why it came up with the numbers that it came up with. So anyway, without getting too much more bogged down on that, I'll give you a couple of videos to look, look at if you're interested in it. That's what's going on. So any of these ideas that the AI system is going to take over the world or kill us all or kind of silly, this is a mathematically optimized intelligence system. Yes, and it can do some spectacular things, but it doesn't have any personality per se. It has learned the personality of the data that it's been trained on, which of course leads you to issues that if you train it on biased data, you get biased results.
Speaker 1 00:07:10 So if you were to build a performance model on your workforce and you were an old school northeast Blueblood, Ivy league company, and all of your executives were white males in their fifties and sixties who went to Ivy League schools, the system would believe that that is a criteria for success. You would have to train it deliberately not to include the race, the school, and other things in order to get rid of that bias. And basically that's what these AI companies are doing is they're trying to de-bias these systems by algorithmically eliminating things that should not be used for selection or promotion or other applications of people. Now, as you know, AI does a lot of things. Most of us think about it as a chatbot, but it does a lot more than that. It can seemingly answer questions, but the more interesting thing to me is it can identify many, many things about you, your skills, your potential strengths, the adjacent skills that you have, a potential career, job or opportunity that would be appropriate for you because it knows through all of these data sources in a large database, many, many things about the job and career and opportunities in a history that you have relative to other people.
Speaker 1 00:08:31 So a second generation AI system like Eightfold for example, is trained on billions of employee records and not just employee records. Today, the time series history of these employee records. So if you took a billion people and there's only about two or 3 billion white collar workers, so this is a large number. It doesn't include oftentimes blue collar workers or construction labor, et cetera that are not really easy to find. If you took these billion people and you looked at their careers over 10 years and you had a snapshot of them year by year, you know, it's actually 10 billion employee records if you think about it or more because people change jobs more than once per year. You are being, you're able to look at who the high performers are, who are the people that moved into this role, into that role, into this job family, into that job, family.
Speaker 1 00:09:19 And then if you enrich that data with all sorts of other data like salary data, performance management, data software, data code that they had written, if there were software engineers, nursing certification data, et cetera, et cetera, this thing's getting smarter and smarter and smarter and more and more useful. Now, you know, why do I get into these three generations? What I'm referring to is the neural network. And the neural network, which uses massive amounts of data and indexes and analyzes it very, very quickly, uses vector calculus. If you ever took any math or you didn't need geometry, you know what a vector is? A vector is a string of numbers, which means something. And in the context of ai, this is called an embedding. An embedding is a vector that has meaning in the corpus of data in the neural network. So every individual has embeddings that tell the system things about that individual person or object relative to all the other objects in the system.
Speaker 1 00:10:19 So these large language model neural networking systems use vector databases, vector calculus, they need vector enabled hardware. That's why NVIDIA chips are so expensive and hard to get. They are very specialized systems, which gets me to the three generations of ai. If you are a large payroll or HCM provider and you have a relational or object jointed database storing your data, you really can't do this kind of processing very easily. You can, but it's very expensive and very difficult. And by the way, if your SAP, Oracle or Workday or any of the payroll providers and you claim to have a lot of data, you really don't have very much data at all. In fact, your company, which might be 10, 15, 50,000, a hundred thousand employees is a teeny, teeny tiny database compared to the 2 billion employee records and 10 years of history that might be available in a large language model.
Speaker 1 00:11:16 So these traditional systems, and I don't mean that negatively, I just mean they're transactional systems that we have in our IT infrastructure are not AI systems. You can add AI algorithms to them and you can use machine learning to predict things from the data you have, but they will never be as accurate or interesting or useful as a large language model or a neural network because they just don't have as much insight or as much data to draw upon. Nor can they do the kind of processing that a billion parameter neural network would need. And so there are really three generations of technologies. There are what I call emerging AI vendors where they're bolting AI features onto their systems and they're, you know, bolting on a chat bot or something from open AI or Google, which I wanna talk about in a minute. The first generation are companies like Workday and, and to some degree Oracle and SAP where they have machine learning built in, but they're not neural networks.
Speaker 1 00:12:11 So they can do a reasonably good job of predicting some things, but they're working on very limited amounts of data and very limited number of data elements only that that's available inside of Workday, which is actually very limited compared to the data that's available in the public domain about individuals. And then there's what I call the second generation systems, which are basically AI systems first. And these are vendors like seek out eightfold, gloat, Beamery and they're, you know, tend to focus on a particular niche. The area of AI in HR that is the most interesting is recruiting. And the reason recruiting is the most interesting is not because recruiting's the most important thing, but the number of data sources and the number of data elements in recruiting is massive cuz you have the sourcing data and all of the candidate data for everybody who's ever had a job.
Speaker 1 00:13:01 So they can develop insights and intelligence that a talent marketplace vendor could not develop. For example, a talent marketplace vendor could certainly see careers and paths and roles that are trending upward and that are predictably valuable for somebody with certain skills, but they wouldn't have as much data necessarily as a recruiting vendor. Now that's not to say that all of these vendors aren't gonna move in the same direction they are, they're all gonna try to get more and more data and and build these larger language based systems. But right now we have these three categories of vendors, the emerging ones using AI for small applications on top of the system. They have the traditional first generation AI vendors that have uh, relatively large databases doing machine learning, may be doing some generative AI and the second generation that I consider to be the real AI model vendors.
Speaker 1 00:13:50 And when you read the the white paper, which won't be out for a few more weeks, you can see more about how these three things stack up. Now why should you care about this fairly arcane, complicated, constantly changing topic? Because it's really important to HR every problem or question or answer or or solution that we develop in HR has some need for better data. I hate to tell you this, but there's never a perfect solution about who to hire, who to promote, how much to pay somebody, what performance rating to give them whether they should get a new job or not, how much training they should get, what training they should get, blah blah blah. I mean there's never a perfect solution to any of that. It's all judgment and it's all based on some amount of limited data that we have. If we can get massive amounts of data, we can make much, much better decisions about how to organize teams, who to hire, who to promote, what are good career paths, what skills are trending up, what skills are trending down, et cetera, et cetera, et cetera.
Speaker 1 00:14:48 So you know, my personal belief knowing this stuff fairly well is that the neural network, large language model types of application systems are going to be maybe 10 to a hundred times more valuable over the next five years than the transactional systems. That doesn't mean the transactional systems are going away. I mean we have to have them but they are not going to give you the value of these other systems. You know, an example of this is last week Eightfold launched their workforce planning and workforce intelligence system for looking at the layout of your organization, identifying skills trending up, skills trending down and so forth. We've been doing this with them for three years or more with the global Workforce intelligence project. That is very, very, very valuable information. You're not gonna find that information in Workday. It's not there, it's just not available. Some of it is but not very much.
Speaker 1 00:15:39 And by the way, this is gonna get even more interesting because last week slipped amongst many other things, Google announced their models. So the Google Palm model, which appears to be as sophisticated or more than open AI is being licensed through Google Cloud. Any one of your IT departments or your vendors can go out and get it as they can open AI and Google built multiple versions of the model that can run on different size computers. So there's a mini version that can run on your phone and then larger ones that run on larger networks. So the vendor community suddenly has the ability to re-engineer their systems to move to the second generation solution like never before. It won't be cheap for them because building a large language model neural network based system is very difficult and different. In fact, there's a BCG study that I can point you to that said that 78% of the corporate IT departments have tried to build their own LLMs failed.
Speaker 1 00:16:39 Not just because it's hard, but there's new skills required, there's new technologies required, there's just new software that they've never touched before. But that's gonna happen too. All of you that work in big companies are going to have LLMs in your company analyzing customer service data, customer purchase data, perhaps serving as a chatbot or I mean I just read not only does Wendy's now have a very intelligent L L M to help you order food and I'm sure it will learn things about you and recommend food, that's probably something you might like. I read that one of the airlines is building an L L M to recommend discounts on seats and fairs to get people to buy more tickets. I mean the amount of creativity and innovation in this is going to be spectacular and it obviously is dependent on elimination of bias on data privacy and protection.
Speaker 1 00:17:31 But that's why I think a lot of the most interesting l m implementations are gonna be inside of companies working on data that they own on behalf of their customers or their employees, which, which will have some impact on the vendor market. You know, if you buy a second generation HCM system and you turn it on and you take the data in the system and you enrich it with the data inside your HCM platform, you now have this very, very rich and potentially risky dataset about your employees. Do you trust your vendor to maintain the security of that? I think you will, but I think some companies may decide to do it themselves. To be honest, IT departments are probably gonna come back and start doing more of their own processing because they can license these models over the cloud just like any of the software vendors can.
Speaker 1 00:18:19 So you know, I think this is just a spectacular revolution in our domain. The final thing I will say is about your role in HR and how do you learn more about this. Now most of you know we've spent a massive amount of energy on our academy. We're gonna be doing a lot more education on ai. We have an AI for HR course in there, actually two of them. And we'll be adding more. We will have a curriculum on generative AI and large language models and neural networks. And I think it's important you understand the fundamentals. You don't need to be able to program one obviously, but you kind of need to know what it is. And we will also start to explain our perspectives on the vendors. I spend a lot of time with the vendors and we advise many of them on this and you're going to see them do very interesting things.
Speaker 1 00:19:05 Just today SAP announced a relationship with Microsoft where they're using the Microsoft co-pilot and co-developing tools for recruiters and other HR business professionals based on the generative AI parts of uh, G P T. So you know, there'll be a lot more of that going on. And I think the IP and value add by HR vendors is going to be much more tilted now towards ai, uh, models, content generated from their data sets and their customers and specific IP that they have within their company as a vendor that they can share with you. And by the way, there's all sorts of fascinating tools like Text io, Grammarly Otter that are already available that use AI for recruiting and job descriptions and emails and all sorts of interesting things. So you know, all of that stuff I consider to be very valuable, some of which will be commoditized.
Speaker 1 00:19:58 But there's lots of opportunity there too. So, we'll, we'll do everything we can to educate you on this and later in the year when we launch what we're calling JBA 2.0, we'll explain more about how that's going to come together for your professional development, your career, your certification and your expertise and and basically your growth as an HR person. So let me stop there. We'll be doing a lot more on this final little plug for the irresistible conference. The irresistible conference is almost three quarters full already, so if you want to come, please sign up in the next couple weeks cuz it's gonna fill up. We are going to have a whole bunch of interesting things there. We recently decided that to invite Keith Saunder Lang, the commission of the e ooc is gonna talk about ethics and ai. We have some of the pioneering users of AI and Microsoft HR coming to tell you about how Microsoft is using AI within hr.
Speaker 1 00:20:50 We have some faculty members from USC in the media industry talking about how AI is transforming the media industry. We will be announcing three significant product offerings at the conference. One that has to do with our research that I'm not gonna tell you what it is, one that has to do with the academy and then one has to do with leadership. We're going to introducing our new leadership model and our leadership research there. And frankly, it's just a great experience. This is not really a conference, it is an experience. The USC campus is open to us. We'll go to the Olympic Coliseum, you'll see the space shuttle. We have a couple of other special little trips you're gonna take that are really, really fun. I can't wait to get down there. So June 20th through 22nd. If you're a corporate member, you have an extra day added on for some special events on Thursday and Friday if you'd like to come to that. And I will have a chance to talk to you very specifically about what's going on in AI, the conference. So I hope you found this interesting. As always, I'm always interested in your feedback If you'd like to debate any of these issues with me, I'm a debater at heart. There's nothing I'd rather to do than have a nice heated debate about a controversial topic. Have a great week and I'll talk to you guys again soon.
