Episode Transcript
[00:00:00] Good morning everybody. Today I want to talk a little bit about a concept that you're not going to probably like. But I think there's an argument for it and that is that the technology that we're all circling around, I.e. aI, is actually starting to become a little bit of a commodity. And the idea that keeps getting promoted by the AI vendors and a lot of the investors, that this is going to be super intelligent and it's going to be smarter than human and it can learn anything and it can embrace recursive self improvement. In other words, it can program itself is kind of baloney in a sense, because it is technology, it's not humans. And the evidence of this is a bunch of things. Let me sort of run through them. And the reason this is important, by the way, is the expectations that you may have for AI as a tool in HR just need to be moderated with a dose of, I wouldn't say sanity, but maybe rationality. So let's talk about software for a minute. When I was in college in the 70s and I learned about software engineering, we used a program called Fortran. And then I went to Cornell and there was, well, there was Basic, the program that was developed by Microsoft, and then there was program at Cornell on The mainframe called PLC, which came from PL1 programming language 1 and PL1, if you read about it, was supposed to be the last programming language ever created. It was the one programming language that everybody would always use. There was Cobol and all sorts of high level constructs and then we had SQL and other programming languages in business and then we had C and then we had Java, and then we had HTML and then we had Python. And what we were basically doing with all of these languages is creating and I mean we being the whole community of engineering people and computer science, is building higher level tools to avoid having to get in and program at the machine level. Because ultimately what happens with all this stuff is it turns into bits that are interpreted by a chip, which nobody wants to do by hand. Very few people, some do, but very rarely. So anyway, it's a big, it's a big exercise over the last 60 years of trying to abstract away the complexity of a computer so that a human can tell it what to do, which is basically what AI is. AI is just one more step toward abstracting away what we want the computer to do versus what we feel comfortable using ourselves. And what you find with the coding tools in AI, much of which by the way, are beginning to get kicked out of companies because they're too expensive is that people have become so sloppy with their problem definitions that they're using the coding tools and building junky stuff. If you don't tell the computer very precisely what you want it to do with the ifs and the ands and where to find this and where to find that and who to ask for permission for this and who has permission for that. And you know, what do you need to know before you do this? What data should you look at for that? It doesn't know. I mean, it could try to find stuff, but it won't know.
[00:03:04] So what software engineering and product management has always been about is specifying in business terms or problem terms what are we trying to do with this system and then pushing a button and trying to get the system to build this thing that we defined. Now in the older days of engineering, before we had AI, you did that in a software engineering team and you would build sprints every day to try to see how well we're doing at building the stuff that we specify. Now the sprints take place in real time with AI and you can actually watch the AI iterate. But we've had some experience with it here where I hate to say it, and I won't mention the sort of project, but you get sloppy in a hurry because you start asking the AI to do things willy nilly because you just can. And it does it, but it doesn't do it quite the way you wanted it. So you then ask it again, then you ask it again. The next thing you know you spent a lot of money on compute cycles and a lot of your own time wasted. So this is no different from sitting down with Cobol or PL1 or Fortran, which we did a long time ago, and just whipping up a program and throwing it into the mainframe and getting an output that says error on line 2 and having to start over. So yeah, it's faster. Yeah, it's more miraculous. But it's not going to feel miraculous in a few weeks or months or years because it's going to be normal. I remember when we first started building websites in the early 2000s, there were companies that were charging millions of dollars to build a website. And then we got HTML Development tools were so easy to use like Dreamweaver and other ones that we didn't need them either. So this is not really anything new now, you know, then the question is, how intelligent are these engines under the covers? Well, you know, even though they all behave differently and they have different strengths and weaknesses and they're, they're spiky in different ways because they've been trained on different data sets and they have different focus markets. They're all very similar. They all do a lot of the same thing. The difference is they're trained with different data. So if you think about the engine as the scalable part and the data as the unscalable part, then what makes a, an AI model or an AI system unique is how it's trained. And you know, I know this because of Galileo. We get compared, Galileo gets compared to ChatGPT and Claude all the time. And it outperforms in a human capital related things because it's really, really well trained in that. And I think that means there's going to AI systems that are well trained in genetic, genetic engineering. There's going to be well trained AI systems in mechanical engineering and in power engineering and you know, lots of things. Like I made the comment the other day about my credit card getting hacked. You would have thought by now that American Express had built a powerful AI for credit card analysis, but they didn't. Or they maybe they haven't tried hard enough. But sure they should because it's probably going to cost them a lot of money when they have to forgive a $2,000 bill that somebody ripped off from me. So my point is that we as users, users of this stuff and developers of new and better HR solutions and human capital solutions have to assume that this is not that different from any technology we've had in the past where we have to apply it, make sure we know what the problem is, clearly articulate the problem, work with the tool vendor or the engineering team or the IT team to specify what we want in as much detail as possible and think logically about what we're trying to do. And to give you an example, I was talking to Rahm about this over the weekend. So big, big retailer, you know who this company is. I won't mention her name. During Christmas they have to hire five times as many people as normal, which happens in certain industries. So they hire 30 or 40,000 recruiters in the fall to start recruiting people for Christmas. And it's a terrible process for them. It's expensive, distracting, unscalable, but they do it because that's what they're used to theoretically with AI. And this is, you know, absolutely doable. There could be an agent, a super agent, that knows the likely hiring per city, per store, that knows the sources from the prior years, that knows the assessment process and selection and screening process, and manages this and could literally tune it up and this is in HR 2030 by the way, as a use case and would save them a fortune in all of this extra hiring training and cost of hiring recruiters. And you know, this could make them a much, much different company because once they got that working, they could use it for other variations in demand for other reasons. Summer, fall, different stores in different cities that have different periods of time where they need more people and so forth. So these are really useful systems once you get them, because you can apply them in many different ways. But you have to define what the problem is.
[00:07:56] And when you think about your job or your role or your skills in AI and what you're trying to accomplish and what tool you want to buy and what vendor you like, and you know, all of those gritty issues that are very kind of difficult, I think you have to take a step back and just go back to the fundamentals of what is the problem we're trying to solve. And just because you can iterate and build 15 different versions of the same thing doesn't mean you're going to end with a, with a good solution. What I've found over my years working in engineering, in some cases software companies and then, you know, in the consulting industry and other places is the most important part of a problem solving exercise is defining the problem.
[00:08:38] I mean it really is what is the problem we're trying to solve. Let's define it very clearly and articulate it. And what you usually find is you don't know exactly, you kind of think you know what it is, but you've defined a sphere of problem, but you haven't defined the specifics of what, what is causing that problem. And, and that's not going to be known by any AI. That's going to be known by you or you're going to have to do more research. In the case of that hiring situation for that company, their high volume hiring is not the same as a restaurant chain or a trucking chain or a bunch of haircut salons. I mean it's just not, it's unique to them. And so they probably somebody in that company knows here's exactly the kind of people we need and here's why we need them and when we need them, and roughly how long we need them, how much we should pay, I mean that kind of information is available. So as you get into your HR tool selections and vendor analysis and you work with us on HR 2030 and roadmaps, we're going to really force you to think about the problem and not get too enamored, the technology now, you know, I've also got a lot of opinions about the human part of this. And the book that's coming out in the fall called Superhuman, which we have a preview of now, is about this and what we're going to probably find. And I can't predict this 100%, but certainly this is where I see the world going, is that no matter how super intelligent, quote unquote, these tools get the human tuning, direction, training, alignment is always going to be a part of the solution. The self driving car is a good example. It's not perfect and there are people behind it and there will probably be people behind it for some time to come. We still have airline pilots in the plane. Even though the plane can land itself anywhere in the world where something is completely automated, there is likely a person involved in the process, not necessarily in running the process or monitoring it, but making sure that it's getting the right answers, making sure that it's getting improved. These are tools, they're not humans. So I think Satya Nadella is starting to talk this way. I'm kind of excited to see him sort of picking up this topic that the organizational issues, job issues, the cultural issues, the training issues are still way more important. I mean, if you think about any AI type solution you buy or build and go back to the retail one for a minute, as that thing starts to work and it starts to do what it's supposed to do, maybe 80% of it is working and 20% is a little bit off. Someone's going to have to look at it and say, you know, how come in this city we didn't get enough people? Or how come in this city we got a bunch of the wrong types of people? Or how come we have high turnover over here? Or how come we have to pay people more over here versus over here? I mean, the AI may not have been programmed or specified to think about that use case. You will see it and you as a human who sees the bigger picture and understands how your company operates, will also see it. So the human element or the human ad or the human factor in AI is not going away. And it will never go away. That's my opinion. I mean, even the space shuttle and all these rockets and things, there's always humans involved in these things. And I think the vendors are doing a great job of convincing us that these tools are so miraculous that we're willing to spend a fortune on them because we only need people. But I don't see that playing out Anywhere. I don't see that in any of the companies we talk to ever. I mean, I haven't seen any AI solutions except for maybe some of the high volume hiring ones that absolutely eliminate hundreds and hundreds of people. Sure, we won't need so many call centers answering silly questions. We won't need a bunch of schedulers scheduling interviews. Yes, there's all sorts of tactical work that's going to go way. But the ultimate effect of a lot of these tools, including Paradox and call center automation systems, is what they allow the people in the company to do is to think about bigger problems and to move into higher level problem definition. It's all about problem definition. So, you know, the reason that I think it's good to sometimes just take a step back and, you know, consider AI as a little bit of a commodity, even though it's, you know, not exactly a commodity, but you get the idea is it gets you out of the weeds. There's a, there's a problem. I mean, I think this is true in every domain where if you listen to a bunch of AI researchers talking about their work, they're really wrapped up in the intricate details of the features and technologies and futures of the thing they're building. And it's very hard for them to take a step back and look at the bigger picture. And I know this because I've done it myself as I've gotten deeply involved in assessment or training or tech stuff or data stuff, I can spend days and days and days on something and then all of a sudden I need to take a step back and say, wait a minute, what was the problem I was trying to solve here? I got a little bit lost in the weeds, getting too excited about what I was discovering, like a scientist. And I need to go back and remember the problem we were going after in the beginning and maybe we already identified the solution to that problem and we're going off a little bit of a dead end. And I think the AI industry is so hyped and so much money and so much investment and such high salaries that there's a lot of endless researching going on. I wish there was this much research in every domain of the world. I would imagine that at this moment there's more money going into AI research and chip research than everything else. Maybe not the genome, maybe not the medical world, but certainly many, many areas. And that's an artifact of this bubble like experience that we're going through economically around AI. The bottom line of this particular podcast is I just want to remind you to be pragmatic and don't sight of the fact that you know what the problem is and the AI doesn't. And so the tool or the vendor or the IT person who's working with you who says, oh, we've got this perfect thing and it does this and it does that and it's an automatic coach and it'll make everybody smarter or whatever. Yeah, under the right conditions. But if you don't know what your situation really is, it may not do that at all. Or you may end up spending an awful lot of time trying to get it to round the edges to do exactly what you want. And the role we're going to try to play here, and we're trying to do this, we've done it for years, is to just keep you focused on the best possible track to do this. And we do this through examples and case studies and lots and lots of tools and research and stuff. And so if you've been doing any AI stuff that's a real breakthrough, just let us know. We want to write about it. I want to interview you on the podcast. We're going to have an in a really great podcast with Lockheed Martin. Pretty soon you're going to hear how they're doing it. We learned a lot about this from a lot of companies at the conference a couple weeks ago. And now that we're talking about recursive self improvement and other scary ideas, I just want to kind of try to calm everybody down and say, don't get too hyperbolic about this. That's my recommendation. And think about it as a technology, like every other technology we've ever had in the past. Thanks a lot. Bye.
