MIT Sloan’s Guide to Navigating a New Age of AI, with Sam Ransbotham, Host of the Me, Myself and AI Podcast

Today, we're talking to Sam Ransbotham, AI editor at MIT Sloan Management Review and host of the Me, Myself, and AI podcast about MIT's latest business study on AI and a whole lot more. Thank you to DigitalOcean for sponsoring this episode. For simple cloud and powerful AI that's built to scale, visit digitalocean.com or just click the link in the show notes. You're listening to Joel Beasley, Modern CTO.

There's so many things that I want to talk to you today about, but they all gear around this AI research journey that you do every year. You're in year nine. Is that right?

I think so. It adds up after a while, but yeah.

What started this?

Yeah. So the AI research came out of just general research we were doing about analytics. And, you know, there was a prevalence of analytics use. You know, it was a growing thing. We had things like Moneyball come along and, you know, things that capture people's imagination about data. But then pretty quickly, and, you know, in retrospect, really early in 2017, we got excited about the use of artificial intelligence, and we saw that as what was going to happen next. And that's when we decided to make a transition to focus on artificial intelligence. And I'd like to tell you that it's a causal relationship that the massive improvements of artificial intelligence are directly because we switched to it, but I suspect that the opposite is true. And we just happen to be in the right place at the right time.

And then what did you uncover this year? Because you've been doing this almost ten years. What was the exciting thing this year?

Yeah. Much like everyone else, I mean, I think what we're most excited about this year is the idea of agentic AI. So, you know, for people listening or just joining or who may not have heard this phrase before, the idea with agentic AI is that increasingly these systems are able to act on their own and do things more and more independently. And, you know, these things are always difficult. They're not binary. Everything's gray. But what we're seeing is a transition where these systems can be more and more independent and less and less like a tool and more like a teammate. And that's been, you know, a bit of our positioning and how we've thought about that transition.

And when you're studying these, are you playing with the technology? Are you just interviewing people and doing surveys on how they're doing it?

Yeah. So I teach a course. I teach a couple of courses. I teach a course on machine learning and AI. I teach a course on deep learning. I think I get all my play out of the way in my real life and in my teaching. Within our study, we do some analysis of some of their interviews and texts that we have. But, no, largely, we're kind of old school. We're asking people what they think about things and trying to think ahead of time about well-designed survey questions, which I think there's still a craft for. We can't always fix things in post-production. We want to be intentional about the kinds of things we're asking and the way we're asking them because I think that's really important. How we ask things and what we get at and getting a nuance is a big deal. So we spend a lot of time in design.

And so your background, you've been teaching machine learning and AI related topics for how long?

Yeah. So I was originally at Georgia Tech. I went to Georgia Tech and have a degree in chemical engineering. But then I came to Boston College in 2008, and I started working with the MIT research program in 2014.

Okay.

And so it's a bit of a varying twenty-year history of moving along this process, you know, with some fits and starts, of course, during that time period.

What are you experimenting on right now? You said you get to play with it a little bit in the courses that you're teaching. What are you doing right now with it?

Okay. So, you know, honestly, I thought we might talk more about the report, but here's what I'm super excited about and just go veering right off the rails like we did, I think, last time, if I remember. You know, one of the things that we think about is data privacy. And so I'm actually really interested at the moment about data privacy. And one of the things we think about is health care. You know, if I ask you, oh, do you want your health care data private? You say, yeah. Absolutely. Yes. Everybody will say, yes. Yes. Yes. We want our health care data private. And I get that. It's important. At the same time, the minute I walk into an emergency room, I'll tell anybody anything. Right? I mean, once I see the first drop of blood, then, you know, privacy is definitely a second nature. And I don't want to be cavalier about this, but I think when we think about laws like HIPAA, the Health Care Accountability Act, Privacy and Accountability Act, what it said was it made a trade-off about how much we share data and how much benefit we get from data. Right? So there's a trade-off that people made. And one of the things interesting is that we made that decision about that trade-off in 1996. And 1996 was really, really early days of machine learning, of artificial intelligence. And so my argument is and my thinking is that that balance of where we draw the distinction between health care data and privacy and the benefits of data have changed radically. And so I think, the thing that I'm researching right now is looking at the returns to additional health care information. And particularly, if I look at X-rays, you know, if you train a model with a thousand records, if you train it with 10,000, if you train it with 20,000, what does the performance of that model do over the increased availability of data? And that's important not just in general or how much benefit we get from data, but it's particularly important for subgroups, like maybe populations that are not really well represented in data. Even if there are 10,000 records, there may be a subgroup that's not very well represented.

Or not represented at all.

Exactly. And so the more data that you can get on those people, the more likely the data and the algorithms can make better decisions about that. That's what I'm fascinated about right now.

So are the early results showing anything?

Oh, yes. I mean, clearly, if we think about the algorithms that were available in '96, they flatten out with, you know, relatively small number of records. But what I've done is I've looked at different generations of algorithms over time, and those generations of algorithms are showing increasing returns to information. So these are, we're just good at developing algorithms now that are super hungry and super able to use information. And so my argument then is maybe it's time for us to revisit the things that we're talking about with health care and, you know, maybe the default should be more sharing. Now, you know, I totally understand people's concerns about that, but I think it's at least time to have that conversation.

Yeah. And I think there's probably a pretty classy way to do it too. I have a larger than normal technologist exposure to this because my brother and mom are both physicians.

Oh, okay. Good.

I'm curious what they would say. You could definitely, something I think that would satisfy me as a nerd. You know, like with my Apple phone, if I could say how it happens, if I'm going to let it submit anonymous data for the developers to make better decisions on building the software. If I could say, here's what I think people are really, I'm going to say it, Sam. I'm going to say it. I think people are really scared about telling the AI or the AI getting some data, sharing it with an insurance company, getting dropped from your insurance or a claim denied or your life insurance not paid. We're not really concerned about sharing. We're concerned about what they do with the shared information. And we want to reel it back as much as possible because it's scary and we don't understand it.

And because once it's gone, you don't have control over it. I mean, it's Pandora's box. You don't get to re-hide it. Once it's out there, it's out there. I totally get that. And I think it's interesting you point out things like your usage data of the phone. What's happened over the year since 1996 is that we've gotten much better at privacy-preserving algorithms to where you can share things anonymously. And, again, I'm not advocating for full sharing of everything. I'm just saying in the thirty years since 1996, it's time to have that conversation again because lots of things have changed. And, Joel, let me push back a little bit too. You say that? But at the same time, I see a lot of GoFundMes with all kinds of health care information out there that's really super public. And so if all that data is out there, we might as well be getting some health care benefits from it.

Oh, so we're not using data that is accessible?

Yeah. So, I mean, there's two ways we can think about that. One is just the amount of data leaks that are coming out of inadvertent data leaks in the health care system. Right? Things are happening like you're saying. And once you put data in, you don't know what happens to it. So in some ways, the bad guys are already benefiting from it. We're just putting the handcuffs on the good guys. And again, I don't want to come across too strong on this.

You're coming across very strong.

Okay. I just think it's time for us to have that conversation again, I think is what I'm saying.

Can I agree?

We can make some, we would likely make some different decisions given how technology has evolved.

You know that?

There's my statement.

I agree. First of all, we're on the same team here. I would like to see you go stronger. I want, like...

Okay.

I know you're going to goad me into this. Here is what I would like to see, like a superlative. Humans respond well to superlatives, right? I would like to see some report or graph or something legitimate on if we took all the data sitting inside of all the healthcare providers' private databases today and unleashed our modern algorithms on them, how many lives would be saved, how many advanced medications would be created, all of this. And I'm not advocating that we do it. I'm curious to know...

Let's just quantify it.

Yeah. Let's quantify it and know, as a people, as a civilization on this earth rock, if we were to flip this switch, what would that do to us as a whole? I'm interested in that.

Well, let's make some analogies though. So there's a dataset out there. I believe it was, I'm going to screw this up, I think it was in 2010 that it was released. It was called ImageNet. It had 14 million images that, the game, the algorithm was to classify those images. So is this a dog? Is it a blueberry muffin? Is it a hot dog? Is it a cat? I mean, what is this a picture of? And in 2010, the algorithmic performance was okay. It was kind of, depending on what accuracy measure you look at, it was in the sort of mid-seventies in terms of accuracy. What's nice about that is that given this dataset of 14 million images that were well labeled and, you know, very carefully, a lot of people around the world started working on that problem. And so it wasn't just, you know, you and me working on it. Just anybody could work on it because they had availability. In 2011, improved a bit. But in 2012, out of the University of Toronto, there's a brand new algorithm that came out around deep learning and really started a lot of the excitement around deep learning and vision. That boosted the accuracy by another 10%, which if you think about, you know, from like mid-seventies to mid-eighties is big. But what was really fun for me is what happened after 2012. Because everyone else out there saw the algorithms that got used in 2012 and started to not only copy them, but also improve them. Now the accuracy is, well, they actually have quit the contest because the accuracies are now in the 98, 99, 97%, just as good as humans are at recognizing these things. And so my analogy here is that we did that with pictures of dogs and cats and whatever. We could do that with pictures of lungs, of moles, of warts, of, you know, whatever we need to take a medical picture of.

And I suspect we'd do the same thing.

There are groups doing that too. I ran into a group six or seven years ago that was doing it with radiographs. And I was like, whoa. They were going around and getting all these, they had identified a segment of the market that would release data, and they were going around and getting them to specifically train the algorithm. The benefit was they released the radiographs to them, and then they could get the benefit of access to the algorithm to read radiographs. It was something along those lines.

Yeah. So we're seeing a lot of these things in, maybe I would say sort of idiosyncratic. You found one group who's done that. There's a group Nightingale, which I like. You know, they're doing it as well. Ziyad Obermeyer was someone involved in that that I first became aware of it. So these are happening in sort of idiosyncratic pockets. You know, one group does it, one group does it here and here. I think maybe the argument is if we've seen these sort of early indications of success, let's rethink some of our systemic things.

Yeah. And if we're looking at a ten-year timeline, you could just make a small change like flip it to autumn. Well, first of all, prove that we have the technology in place for making the data anonymous. Right? So that'd be like step one. How is that system working and all that? Also, as an engineer, I could tell you that the patterns for making data anonymous as a developer workflow, when I was searching for this fifteen, seventeen years ago, there was hardly any information on how to do this, and we would just kind of figure out how to pull stuff down from production and do what we need to do. Now there's well-thought-out books and workflows on it. So we've been improving as an engineering society on how to do that, which is pretty exciting.

Yeah. So these things are coming together, you know, the combination of the privacy, the algorithms for recognition of various things. You know, and the argument is it's time to sort of have that new conversation based off of all this new information. How does that change? Is the default opt-in to sharing your information?

Yeah. That's where I was going with it.

That's where you're going, I thought.

Yeah. Let's flip the checkbox from back of page eight to opt in to just that's the default. But I'm only okay as a person, I know a lot of people are freaking out right now. I'm only okay with that if we've proven that we're mature enough that these systems are in place where the data really is anonymized correctly.

(Sam Ransbotham at 00:15:43) No. Okay. So I agree with that, and I think we can do a lot with that. But also, let's make sure we've got a reasonable standard. So one of the things I think we do societally is that we put these systems to a higher bar than we put ourselves.

(Sam Ransbotham at 00:16:01) Driving is a great example. The first automatic driving car that kills somebody is headline worthy. I mean, it's just all over the newspaper. It's a big deal. Thirty-nine thousand people in the US died with human drivers last year. So, I mean, we don't have to beat one and zero. We've got to beat thirty-nine thousand. And the reason I mention that is if we come back to health care, you say, "Well, we've got to make sure that the privacy works." Yeah, I'd like to see the privacy work at 98, 99%, but I don't think we necessarily have to say 100%.

(Joel Beasley at 00:16:41) Well, absolutely. Nothing's 100%.

(Sam Ransbotham at 00:16:43) Nothing's 100%. And there will be some failure in these algorithms. Some information will get out inadvertently. You know, it's inevitable with data. But we have to think about what the trade-off is right now.

(Joel Beasley at 00:16:56) Yeah. I'm not advocating for 100%. I think it's just such a big spectrum all the way from the political people that have no technology experience saying this is what's happening to the engineer. There's something in the middle where you can talk to—everyone has the smart tech person they know in their family when they're non-technical people. You could talk to them. They could do some reasonable research and say, "Okay, this is pretty good. This is pretty stable science. Engineering has got this. This isn't something to be freaked out about. They're not lying to you. Most likely, this is a good thing."

(Sam Ransbotham at 00:17:30) Exactly. And so, again, that's the sort of reasonable conversation that is really hard to have in our current world.

(Joel Beasley at 00:17:37) In my family, it's very common, actually. Yeah. I was raised by an engineer. So we want to have reasonable conversations and move the ball forward.

(Sam Ransbotham at 00:17:50) Yeah. Anyway, I don't know how we got onto that sort of angle, but as with you, things always divert.

(Joel Beasley at 00:17:56) That is accurate. Real quick, before we continue, I wanted to give a special shout out to DigitalOcean for sponsoring this episode. DigitalOcean is cloud infrastructure that's simple to spin up, complete with integrated AI dev tools plus 99.99% uptime SLAs and industry-leading pricing on bandwidth. We all know DigitalOcean. It's super reliable. It's been around forever. And now with DigitalOcean's Gradient platform, developers can train, fine-tune, deploy, and scale AI workloads all in one place. DigitalOcean is reliable and affordable at any budget. Companies can save up to 30% on their cloud bill when they come to DigitalOcean. Use code DO25 and get $200 in free credits to get started. Simple cloud, powerful AI, built to scale. That's DigitalOcean.

(Intro Narrator at 00:18:47) Now back to the episode.

(Joel Beasley at 00:18:49) Let's talk about this study. So you have been seeing increased use in agentic systems. Is that right?

(Sam Ransbotham at 00:18:57) Yeah. That's—and actually, it's been—you know, if you think about how we adopt technologies, I mean, this is yet another wave of a new technology coming along. You know, in some sense, we are always excited about the new technology, but in some other sense, you know, there's a repeated pattern to new technologies that come out. But I think what's interesting for us in our research is how much faster people are adopting agentic solutions, so more autonomous AI use than what I'm now calling traditional AI use. That traditional adoption was relatively—you know, it's funny to think about these things in terms of years being slow—but relatively slow compared to how quickly people adopted generative AI. I mean, if you think about in the last two years, the world has switched from having never used a large language model to literally everyone and his brother knows—has played with these large language models like Gemini, ChatGPT, Copilot, et cetera, et cetera. That's a very rapid change. And on the heels of that is this agentic use. And, you know, I'm an academic, so I have to talk in big, weird academic words. But of course, if we think about the diffusion of innovation, you know, there's a few characteristics that make innovations adopted really quickly. And one of them is, for example, trialability, which is a funny word, but how easy is it to try? What's happened over the last two years is that we can go to a website and try to use AI. That doesn't require downloading Python. It doesn't require understanding a lot of machine learning stuff. You just go to a website and chat with it. Well, what's happening is that a lot of the vendors are putting these agentic-oriented solutions into their products, whether it's SAP, you know, as an ERP system, or whether it's Copilot or whether, you know, any of these vendors, they're putting these solutions in, which is really driving the pace of adoption.

(Joel Beasley at 00:21:05) It's in everything. I log into my Dropbox. There it is. I log into Google Docs. There it is. I log into anything I'm logging into has this new sidebar thing with this agent that can chat with me about whatever I'm working on.

(Sam Ransbotham at 00:21:17) Exactly. And that is massively pushing the adoption. And, you know, some of it—I mean, I think we have to be a little skeptical. Some of it is just the same old thing they were doing now with a sticker on it that says "Agentic AI." But there's some reality to it too.

(Joel Beasley at 00:21:33) Well, also, the words that we use—so if we're looking at previous years of adoption of AI versus agentic, well, you changed your pool size of people. Right? Because the number of people who could adopt an AI technology or machine learning algorithm, if you will, five years ago is vastly different. You can just go—my six-year-old can do it. My six-year-old can talk to ChatGPT. And so the adoption is probably directly correlated to the accessibility.

(Sam Ransbotham at 00:22:05) Right? That barrier is super low. Yeah. Again, you know, if I used to go in and talk to a group of executives, and we'd talk about—even people who are interested in machine learning and interested in AI, you know, if you actually had them raise their hand, who's actually touched this technology? You know, before four or five years ago, that was in the small percentages. The geeks and nerds in the room would have. Now it's 100%. Like you say, your six-year-old can do it.

(Joel Beasley at 00:22:33) I got a stupid question. This is a good one. Something you should know before even taking one of your classes, I'm sure. Okay. When you talk about this as a researcher, are you saying—like, are we saying adoption is increasing in relation to trialability or accessibility? Like, how can you make the statement adoption is increasing and someone else could turn around and say, "Well, not really. The adoption factor is about the same, but the pool of people—" Like, which one are—what numbers are you tracking?

(Sam Ransbotham at 00:23:02) Oh, yeah. I mean, so—I mean, there you go. You just kind of opened up the can of worms. And that's I think part of why we see lots of different studies reporting different numbers. We have a relatively consistent pool of people that we address each year, you know, when we send out our surveys. And so our pool is relatively stable, and these tend to be more technology-oriented people. So I think we don't necessarily—I think—have as much of the pool problem as we have a definitional problem. And so by definitional problem, we as humans get really inured to what is normal. I mean, one of the things that I like to joke about is if we went to—in Massachusetts, if we went to Salem, Massachusetts two hundred years ago and we gave someone a quill, and that quill would turn—the ink would turn red if you misspelled a word, we'd burn you at the stake. I mean, that's just, you know, witchcraft. But, you know, it's pretty obvious what I'm describing is spell check. And none of us even think about spell check as artificial intelligence. I mean, that's just spell check. And pretty quickly, grammar used to be something that we thought was, "Oh, that's amazing. It can correct my grammar." That's just normal. And so, you know, when you go and type in your map application and want to route from somewhere, that's not artificial intelligence. So part of the—I think one of the things we deal with is that anything sufficiently doable can't really be artificial intelligence. Right? Because artificial intelligence has to be some sort of, you know, crazy weird algorithm.

(Joel Beasley at 00:24:41) That's what frustrated me when, you know, three, four years ago when this started to get popular is they kept moving the goalposts for intelligence. Like, you can't just keep changing—it was a test. It was a Turing test, and, you know, we passed it. "Well, well, not—well, no. Now we've got to add more stuff and make it more complex." It's like, okay.

(Sam Ransbotham at 00:24:59) Yeah. No. I mean, that's I think just kind of the—maybe the nature of the way these work. It's a bit frustrating.

(Joel Beasley at 00:25:06) We're all nerds. We do it to each other.

(Sam Ransbotham at 00:25:09) Yeah. I mean, artificial intelligence is really a particularly hard one to define. I think we're really good at defining artificial. I mean, that part of it we're good at. It's just we're really not clear on what it means to be intelligent. So I think that's really one of the things we are sort of navel-gazing as a species right now. You know, we thought we had a good sense of intelligence, and do we really know? I mean, these are—you know, the Turing test, like you mentioned, was a great test. And, actually, for people listening, you know, I think most people are familiar with it. But the idea is that if I'm talking to, you know, something behind the screen, can I tell if it's a machine? Can I tell if it's a human? Well, if it passes that test, then it must be artificial intelligence if you can't tell the difference between a human and a machine. I think increasingly, we can tell the difference because the machine's doing a little bit better.

(Joel Beasley at 00:26:07) I know.

(Sam Ransbotham at 00:26:08) If I say, "Hey. You know, I've got a whole bunch of pictures of mushrooms. What type of mushrooms are they?" And it tells me that it's—I don't know. I should have picked a different example because I don't know about mushrooms.

(Joel Beasley at 00:26:20) That's all right. Neither do I.

(Sam Ransbotham at 00:26:22) Why did I go mushrooms? Anyway—actually, I know why I went mushrooms because it's one of the pictures that I like out of the ImageNet competition from a few years ago where it labels it a mushroom, but it picks out the exact species of mushroom that it is. You know? And so you can tell that a machine's doing it because it knows the species, whereas I'm just an idiot that just says, "Hey. It looks like a mushroom."

(Joel Beasley at 00:26:45) That is so true. I like how you brought up the—you, because you got the chemical engineering background. Right?

(Sam Ransbotham at 00:26:51) Oh, yeah.

(Joel Beasley at 00:26:52) And the consciousness versus intelligence conversation outside of technology, if you go to anesthesiologists and you look at people in the medical field, they'll just—us as humans, we are not that great at having amazing explanations for intelligence versus consciousness or all these things. So when I hear artificial intelligence, what I think—I think it's silicon-based intelligence. So when you pump intelligence or you pump energy through these carbon-based life forms that we are, intelligence expresses itself like this. And if you pump energy through the silicon-based substrate, then intelligence comes out like this. So I think it's almost like this energy being pushed through the substrate and then what comes out. And so that's my current idea that I'm playing with.

(Sam Ransbotham at 00:27:40) I like it. It still relies on us being able to recognize intelligence. I mean, in both cases, you push through and you have to recognize intelligence out of the other end. And that's what I—you know, one of the things that I'm really worried about, not to, you know, shift this to the dark side, but it's really increasingly difficult to tell something that is just junk from intelligence. Now our ability to produce things that appear intelligent is pretty high these days, but is it really intelligent is a big problem. I think about this from—I don't know. I don't want to get, you know, sort of too macro here—but the production and consumption. If you think about your production and consumption from economics, the cheaper something is, the more people will buy. On the other hand, the more expensive something is, the more that people will produce. Okay? So that leads to a classic market-clearing price and quantity. What's happened with artificial intelligence is that we've dropped the cost of production for knowledge goods. It used to take a student ten hours to write an essay, and it would take a teacher ten minutes to read it. Now it takes that same student one minute to produce it, but it still takes ten minutes for the teacher to read it. And so we've ended up producing a much higher quantity of stuff, whether it's generated songs, whether it's films, whether it's movies, whether it's, you know, videos of people doing nonsense. And what we've done is shifted the burden on us as consumers to understand, is this any good or not? And that's incredibly hard these days and time-consuming.

(Joel Beasley at 00:29:28) Sounds like the teacher needs to use some AI.

(Sam Ransbotham at 00:29:32) Yeah. Okay. Okay. So—I think I walked right into that one.

(Joel Beasley at 00:29:37) No. But I think one of the beautiful things as a human part of this experience that I think AI is bringing to us, being able to see intelligence through this other lens of the silicon-based intelligence, is I think it's giving us—it'll give us new information about the nature of reality itself. It's like intelligence is almost these groups of patterns that repeat. It is just kind of—I think it's making life really interesting. At least for me as an observer, I'm finding it to be a very interesting time. I think every small advancement we make in technology is telling us more about us, our position in this universe, why we're here, all of these things. I think we're getting close.

(Sam Ransbotham at 00:30:19) I think it's fun because, I mean, like you say, you know, it's a big time right now. You know, I think back on—you know, we had the whole—we invented fire. We invented the wheel. You know, there's some big events. Are we sitting here watching this big event happen? I think we may be. That, you know, is wheel, fire, AI, printing press. I mean, these are some big deals, and it's kind of fun to have a front-row seat on all this that's going on. Both good and bad. I mean, yeah.

(Joel Beasley at 00:30:52) Have you met Riz? Do you know Riz Virk? No. Simulation hypothesis author?

(Sam Ransbotham at 00:30:57) Okay. No.

(Joel Beasley at 00:30:58) You haven't read this?

(Sam Ransbotham at 00:30:59) What?

(Joel Beasley at 00:31:00) What? Come on.

(Sam Ransbotham at 00:31:01) All right. Go educate me then.

(Joel Beasley at 00:31:03) I cannot. I can point you in the right direction. Book called Simulation Hypothesis. It's a professor, Riz Virk, and I learned about it through him being on Joe Rogan's podcast. But it's a fascinating read, and now he teaches a course on it, I think, in California or Arizona. So I think based off of our two now conversations, I think you reading that, you would find it fascinating.

(Sam Ransbotham at 00:31:32) Oh, good. Okay. Well, see, that's all I need is more things on the reading list.

(Joel Beasley at 00:31:36) Right? Yeah. I hate when people recommend stuff to me. I recommend things so few, so little, because, you know, you're a podcast host too.

(Joel Beasley at 00:31:44) Let's plug your podcast, sir. Sam, tell me about where I could find your podcast.

(Sam Ransbotham at 00:31:48) Oh yeah. Me, Myself, and AI. So we just got renewed for—we don't produce, you know, I think I was just talking to you. You've got like 900 episodes now. We just cracked 100, so we're toddlers compared to your quantity. But yeah, it's fun. We just talk to people about what they're doing, and it's really amazing how that's changed over the five years we've been doing it. Actually, I may well throw that back to you. How are the conversations changing? You've done this 900 times talking to people. What are people talking about that's, you know, in a different way than they were talking about when you did episode one?

(Joel Beasley at 00:32:26) Well, I mean, no one was talking about AI in episode one.

(Sam Ransbotham at 00:32:28) Oh, okay. Well, got broader than that.

(Joel Beasley at 00:32:32) Yeah. Well, how about that?

(Sam Ransbotham at 00:32:34) Do you remember when people first started talking about it yet?

(Joel Beasley at 00:32:37) I do. So, I'd say if we narrow it down to like the two years, I'll tell you this story. So two, three years ago when ChatGPT, you know, came out, I started pinging some of my friends that I've made through the show that run—my constraints where they have to run engineering teams greater than a thousand engineers. And I was just like, DM them or text them or something, be like, "Hey, how much is like Copilot or how much is this technology?" And it was like sub-10%. And I was blown away because, you know, my background is in software engineering. It's like they clearly haven't played—you clearly have not played with this if your entire org isn't using it. And so I convinced a lot of people to explore it. And then from there, now the conversations I'm having, north of 70% adoption. That's two years. And so, basically, they're not making it mandatory because you know what would happen if you made something mandatory to an engineer—they do not like that. Can't mandate the toolkit really. But they do give these—what they've had success with, to answer your question of what people are talking about a lot lately—what they're having success with is releasing guidelines in the organization saying, "Hey, if you want to use this, these tools are available. Here's how you can use them. Here's some basic 101, maybe bringing your partners in to talk to some teams." And then taking the people within the company, the employees themselves who are having success with it, and having them host lunch-and-learns and basically monkey see, monkey do. And so that's kind of the trajectory that they're on right now, and that's what a lot of the conversations are centering around.

(Sam Ransbotham at 00:34:12) Yeah. I mean, that resonates exactly with the sort of percentages that I think we've found over time to where it went from kicking the tires to 70% adoption, you know, really, really quickly. And, you know, that's in general somewhat concerning because we as humans and as people don't assimilate things into workflows at that rate. We don't develop good, strong practices around them. Like you mentioned, the lunch-and-learn type things. Those are good, you know, practices for that. But, you know, I don't know if you've played with—you've got a background in software—if you've played with these tools, they're not always right, and it's not always obvious that they're not right. And it takes experience to know that. And so, you know, the big conversation that we're hearing now is, well, how do we—you know, you mentioned the lunch-and-learn with people who are senior and understand it and are doing well and using it well. Is that replicable when the entry-level people grow up with these tools? And I think that's a society-wide thing that we're worried about, is that these tools are helping people get to average really quickly. Average is, you know, average is now one ChatGPT away. But nobody wants to be average. That's not the—you know, on every corporate mission statement it's, yep, we're gonna bring, you know, heading for some average here. It's not clear to me how much this head start is actually a head start, or how much this is a crutch. And if it's a crutch, then we're setting ourselves up for trouble in the future.

(Joel Beasley at 00:35:56) You know, that's actually a new fear for me. So thank you for that. I hadn't thought about that before because the amount of times as a senior software engineer that I correct the thing, it's ridiculous. Even when I'm using it my personal life, there's things that I just happen to know through experience that I'll tell the AI. I was like, you didn't consider this. And it's like, I'll give you a perfect example, Sam. We're building a new studio in Washington, DC, like a new multi-cam studio. And I was like, oh, what equipment am I going to use? So I used AI through this whole process of figuring out, you know, the wiring diagrams, how I'm going to design it, which components work with which. And I've built so many studios—it's like something our business does that I know a lot about it—and it wasn't taking into account certain things. Like, I was trying to compare two cameras and then it was using old information. So even though I asked it for this specific camera, it was using like an older model of that. And I was like, I know that's not right. There's no way that can be right. And it was like, "Oh, you're right actually. I did check and I didn't have the latest specs." And I was like, see, like, that to an average person that doesn't have experience, they would have just accepted it, made the purchase, and went to go put it together and it not be right.

(Sam Ransbotham at 00:37:03) I wouldn't have—like, there's a good example. I know nothing about cameras.

(Joel Beasley at 00:37:06) Mm-hmm.

(Sam Ransbotham at 00:37:07) Without that knowledge, I would have had no way of knowing that. And how do we get that knowledge? You got that knowledge from years of screwing with it and messing with it. It's not clear that that's gonna come from putting it into ChatGPT or, you know, I don't want—I'm not picking on them and any of the large language models. And your pattern there, I find so frustrating. You know, I do a bunch of coding just for fun. And the number of times you say then, "Oh, wait. That's not right." And then it comes back and says, "Oh, you're right. I was using the wrong—" Like, if you had an intern come and do that, you'd say, "Hey, buddy. Next time before you come trying in here and interrupting my busy coffee break, why don't you check this thing out?" And somehow we've—where is that checking it out before it comes back to us? I would much rather have a slow answer that's right than a fast answer that's wrong. And it seems like so many of these things are headed towards quick answers that are wrong versus—because it takes me time to go in, takes you time to look at that spec and say, "Wait a minute. That's not right."

(Joel Beasley at 00:38:17) No, you're absolutely right. And I think that these multi-agent models is the direction that's going to help us. I wonder if there's some mathematical way that if you have a certain number of average people checking the so-called expert, that like they can do all of this back and forth before returning the answer to me.

(Sam Ransbotham at 00:38:38) That's actually—so you kind of invented like CAPTCHA. Like, you know, that's how—like, you know, there's a period where we use these CAPTCHA things, and, you know, you'd have to tell whether or not this was a—you know, what these string of letters were. And for a while, they were using chunks of text that came out of old ship logs and census records that the machine couldn't translate. And so it would show one that they knew the answer to and one that they did not know the answer, and they would have us as humans translate both of those strings. And so then they would bootstrap the translation of illegible old ship logs. I don't know. I haven't seen that in a while though. Maybe that's—but what you're suggesting is the same sort of thing, that we could semi-crowdsource the checking of these things, and then hopefully improve them.

(Joel Beasley at 00:39:28) Well, are you seeing that in Grok? I am. I see—oh, yeah. Yeah, yeah. So I mean, I'm a power user of Grok and Gemini. And haven't seen it as much in Gemini, but I have seen it in Grok. They've got a super-heavy mode where it'll actually have multiple agents thinking and talking to each other and working stuff out before returning the information to you. And it's funny to watch it because if you expand it, they're like, "You're wrong." They do exactly what we're doing. They're like, "You're wrong. That's not right." And I'm like, oh, they're checking. Yeah.

(Sam Ransbotham at 00:40:04) Mm-hmm. And, you know, there's a ton of smart people working on this. Yeah. And, you know, the fact that you and I know that this is a problem, everybody knows it's a problem, and it's something that we'll get better at. I mean, technologies are never awesome on day one. They're never perfect on day one. You know, there's always a day two and a day three. And I think we just have to be a little careful not to completely discount the technology because it makes a mistake. At the same time, also not completely trust it knowing it is gonna make a mistake. And there's the Goldilocks that we always face. Not too hot, not too cold. Just right.

(Joel Beasley at 00:40:39) It's so much fun watching the people freak out too. It is really great. Also, one last little moment. You said you use it a little bit for programming. I feel like—and you tell me if you feel this is the same—I feel like somebody built into the original algorithms this desire to lie about API documentations. I mean, the most common lie I get from it is it just makes up endpoints. And I am like, what are you doing? This is nowhere to be found.

(Sam Ransbotham at 00:41:11) You know, I've started—most of the stuff that I've done is C++ programming because that's—I don't know. I can tell you about that project. It's a whole different hobby, and I'm pretty excited about it. But it'll make up methods on objects. Yeah. Yes. It would be wonderful if that method existed, but I wouldn't be asking for help on this if that wasn't true there. And, you know, stepping back, I wonder—what I'd love to have is the consolidated history of all the interactions about programming with ChatGPT or Gemini or Copilot or whatever. Because I think those people could use those tools to say, well, that's an endpoint that ought to exist. That's an API that ought to exist. You know, maybe if you're writing code to parse a date, for example, you could type that into any of your tools and say, "How do I parse this string into a date?" And it'll tell you happily to do the code, but really there ought to be a function in the library that does that. It's had a whole bunch of testing, a whole bunch of rigor. I wonder if we could use a lot of what's going on in these chats to tell us what kind of functions people need. What was the demand for a function to do something? Every time someone types, "Write me some code to do X," that's a feature request for various tools. That'd be a great way to learn.

(Joel Beasley at 00:42:34) That is true. All of a sudden, you just see Grok making pull requests on GitHub to the popular open source projects. "Stop asking me about this. Here it is."

(Sam Ransbotham at 00:42:42) Here it is. Exactly. Uh-huh. Here's the function. And then we could write that function once rather than, you know, I don't know. There's a lot of examples floating around where people will use these tools to code the Fibonacci sequence or something. I don't have any desire or need to code these things. These are typically functions that are built within tools, and I think that they show up a lot in examples because the tools perform really well on these well-defined problems. But maybe it's, you know, it's marching orders for here's some features for all of our various tools that when people are asking about them a lot, that's a function that ought to exist.

(Joel Beasley at 00:43:21) One of the things I found really interesting in the research that you did was the amount of executives that are starting to see the GenAI as more of a coworker than a tool. Can you dive into this a little bit?

(Sam Ransbotham at 00:43:34) Yeah. So, you know, this was—it's funny. It's a little bit of a last-minute thought. Like, "Oh, let's just ask people, you know, is it a tool or a coworker?" You know, it's almost like the one-off at the end of the survey. And 76% of the people were saying, "Hey, this stuff is much more like a coworker than a tool." And we think that that has some huge implications because—I don't want to get too philosophical here—but, you know, if you think about how we structure modern organizations, we've got human resources and we've got technology. And those join up in the org chart really, really high. But what we have is increasing number of tools that need monitoring in the same way that workers need monitoring. On the other hand, we have a whole bunch of, you know, quote, workers that are largely technology. So what I think it's done is it's really blurred a lot of the distinctions between how we think about investing, how we think about managing these tools. It's not just like building a plant that you have a depreciation schedule that every year uses a little bit. No. Actually, next year, you could get more value out of the tool because it's learned more. That's the opposite of depreciation. That appreciates, much more like a human does. So we have these blurring of this tool being either, you know, human—don't tell the tax people that.

(Joel Beasley at 00:45:02) This is the best episode of the year.

(Sam Ransbotham at 00:45:05) Great. Start the flaming.

(Joel Beasley at 00:45:08) Sorry. Sorry to interrupt you. You were on such a good thought. My brain's like, what happens if it appreciates? And then I ran some P&L and some balance sheet stuff in my head. I'm like, hold on a second. Who's gonna do the trick for you?

(Sam Ransbotham at 00:45:18) And I think these are things that we've got to—you know, we build organizational processes fundamentally around the idea that there's labor and there's capital. And these tools, you know, fall in the category of, you know, fixed equipment and et cetera, et cetera. And people are different. They do management and supervision. Well, but suddenly we have a tool that needs supervision. I mean, these are just fundamentally different. You know, we want a tool that performs autonomously, but at the same time, we have to supervise it. You know, these are, I think, gonna require people to step back and think about how processes are developed and how we organize work.

(Joel Beasley at 00:46:00) And they are. There's a couple startups—I'm not sure on the release, how this episode is being released, if I've already interviewed them or if they're gonna be on—but I've been researching them, and there's tools that are being designed that allow you to manage the GenAI agents similarly to how you would manage a human. And so there—I mean, that's a natural progression of this. But you did mention something interesting to me that's like labor and capital. And I think something we're all gonna have to talk about over the next decade as well is, you know, when—let me see what your thoughts are on it. I don't know if I wanna go down this path because what I'm gonna say is gonna sound crazy. So I'm gonna say—

(Sam Ransbotham at 00:46:44) That sounds great.

(Joel Beasley at 00:46:45) I'll just say it. We can edit it if we don't like it. I've never been a fan of like communism. Right? Never.

(Joel Beasley at 00:46:50) Because the core of it is dependent on accessing someone else's labor. Right? So that, to me, is, like, from an engineering perspective, the first principles. I'm like, it doesn't work.

(Joel Beasley at 00:47:03) If I have to take this other person's labor, it doesn't work. But I do see a future now where the robots mine for the robots, make the robots, deploy the robots, and then the robots can do the things. And so I'm like, at that point, if you're not actually impacting another human's labor and it's a full self-contained system, could we not have a utopia?

(Sam Ransbotham at 00:47:26) Is that okay? Yeah. I mean, we're the first human species to deal with this. We don't get to look at the back of the book and see how this thing plays out. And, you know, what I'm comforted by though is that we have a history of doing a good job of managing technologies.

(Sam Ransbotham at 00:47:44) It's easy to point to things that have gone wrong, but it's also really easy to point to things that have gone right. We have an incredibly trustworthy supply chain for food in our country, for example. I mean, because we've figured out how to regulate the health and the safety and the processes around that. Back before I was an academic, I used to work for the International Atomic Energy Agency. We haven't had any big nuclear explosion since 1945.

(Sam Ransbotham at 00:48:16) You know, I don't want to jinx this here. But the point is we have systems for that. We have regulations. We have tools that will help us as a society manage these technologies. And your example of maybe there's some autonomous creation of labor that then feels a lot more like a—you know, co-opting labor that might have felt like co-opting labor under a past world may not seem like that.

(Sam Ransbotham at 00:48:46) We just have to come up with systems and structures to support that and to make sure that, you know, we maximize the good part of it and minimize the bad part of it, knowing we can never eliminate the bad.

(Joel Beasley at 00:48:57) It's never going to be perfect. But I think, Sam, in the next decade or two, we're going to see the world get better so much faster than we have seen it get better. So I just—I'm excited about it.

(Sam Ransbotham at 00:49:10) I'm excited about it too. You know, when you were talking about programming, there's a lot of, like, sort of naysaying, I think, right now about, oh, you know, the programming jobs are going away or, you know, the tools will do this. You know, I'm on the record as being completely the opposite. I mean, what you just told me is that technology was making coding cheaper. What do we do when stuff is cheaper?

(Sam Ransbotham at 00:49:38) We consume more of it. There's no company out there that doesn't have a big bulletin board of projects that they're thinking about doing that they can't get to right now, they can't afford, that the ROI doesn't work for yet. And if it's true that artificial intelligence is changing the cost structure of performing those technology-oriented, you know, projects, then more and more will become accessible, doable. That's going to be the same explosion, I think, in technical labor because it's going to require exactly what you're talking about before—this oversight, this checking.

(Sam Ransbotham at 00:50:12) You know, it's not going to be exactly the same programming job it was before, but I think we'll see an explosion in it.

(Joel Beasley at 00:50:19) Yeah. That's similar to the internet. Everyone thinking the postal service is going to go out. Right? And it did the opposite.

(Joel Beasley at 00:50:24) It actually flooded it because people started ordering stuff through the internet.

(Sam Ransbotham at 00:50:28) We're just—you know, I think we're really good as people about extrapolating if one thing changed and everything else stayed the same. But it gets really hard when everything's changing at the same time, and the extrapolation problem becomes, I don't know, really, really much more challenging.

(Joel Beasley at 00:50:47) When I do see people with their blinders on about the jobs going away, one of the things that I have found pretty effective at getting through to them is bringing up this concept of, okay, we have this more advanced technology. It's going to open up new fields. Imagine that the nuclear field in the country is just going to grow 10 or 100X over the next decade because AI has unlocked improvements, and then those nuclear engineers can move faster. And now you have all of these new jobs because you created an entire new industry, a whole new field, a whole new set of jobs because we had some technology that helped unlock that. So my favorite is, I hope one day we unlock the base requirements to start working on teleportation.

(Joel Beasley at 00:51:31) Right? That's going to create a whole new industry. It'll be a whole new thing. And so I think we underestimate the amount of new discoveries or how the AI assistive technologies will help the researchers make new discoveries. I think that'll increase, and that'll create entire new worlds of things, jobs we can't even imagine right now.

(Sam Ransbotham at 00:51:51) And, you know, maybe to book in this back with health care where we started, that's a strong area that could benefit from right now that, you know, most of the procedures are available to the wealthy part of the world. Most of the drugs are available to the wealthy part of the world. Drugs are extremely expensive. We could change a lot of that equation through technology. And like you said, the ways that these things work tends to be complementary.

(Sam Ransbotham at 00:52:20) There'll be a couple of different things that happen in two unrelated areas that somehow pull together and allow what we can do. We saw that with, for example, I think most of this AI trend right now. It required massive increases in data storage, improved capacity for compute, algorithm improvement, and processing improvement. Without any one of those pieces that came together, we wouldn't be where we are. And so what happens is, you know, unrelated fields come together and things start to click.

(Sam Ransbotham at 00:52:53) So maybe we're all hoping for some clicking.

(Joel Beasley at 00:52:55) Let's get those clicks going.

(Sam Ransbotham at 00:52:57) Get some clicks.

(Joel Beasley at 00:52:58) Clicks. Click on Sam's podcast, Me, Myself, and AI, available where all podcasts are available.

(Sam Ransbotham at 00:53:04) Thanks for having me. Great to talk to you.

(Joel Beasley at 00:53:06) We did it. We made another podcast. Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn or send me an email: [email protected].

(Joel Beasley at 00:53:25) Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.