You Can't Prompt Your Way to a Flying Car | Tim Berglund

Speaker A: Because for most of my career, it was a field that seemed like it would always have a bright future, meaning never a bright present. Right. It was always a promissory note. December 2025, my Claude code moment. I'm like, okay, this is real. And I love that you can't prompt a flying car. It's real hard to make one.

Speaker B: Most of the people telling you AI changes everything discovered it about 15 minutes ago, maybe three years tops. And they'll swear that they always believed it would change everything, even though they did not. But my guest has been not believing in AI for more like 20 years. His whole career, this field kept promising. Human level intelligence, always 10 years. No, five years, maybe even faster. There was a summer conference at Dartmouth back in the 1950s, and the founders figured they'd nail that whole human mind thing in like a couple of months.

Speaker C: They were off by a bit.

Speaker B: But then last December, Atul finally did something he couldn't argue with. And when that guy changes his mind, I want to know exactly what he saw. Tim Bergeland runs developer relations at IBM. And skeptic or not, he's one of the best explainers in tech. He thinks the tools are a big deal, and he's still not convinced there's

Speaker C: a mind in there.

Speaker B: So let's dig back through the old hype cycles and find out whether this one's actually different. Here's Tim.

Speaker A: Thanks for having me. Thanks for having me. Great to be here. And yes, very recently joined IBM via acquisition. They acquired Confluent. And so, um, um, I'm in there running the larger DevRel program there. And yeah, I've always been at arm's length from AI just because for most of my career, it was a field that seemed like it would always have a bright future, meaning never a bright present. Right. It was always a. A promissory note.

Speaker C: The AI sun will come out in the AI tomorrow.

Speaker A: It's gonna really, bro. Just. We need five more years. Just five more years, bro. That. That kind of thing. And it turns out in looking into the history in a little bit more detail, just because that's a thing I like to do. That's. It's been that way for a long time. And it started that way in the early 50s with a conference at Dartmouth University of between 10 and 12 leading mathematicians and computer scientists of the day, kind of in and out throughout the summer. They're like, I think we just need a couple months and we can develop a complete understanding of human intelligence and creativity and language and everything we can. We can nail that in two months. Maybe they were just trying to get funded and saying crazy things. But honestly I think there was a lot of optimism there from the start and kind of went that way for a long time until recently. And I can, I mean I can, I can fill in more of that story if you think it's interesting or

Speaker C: wherever you want to go. Well, uh, I do want to dive into it. I don't want to like steal your whole talk. So you still do have to come down to Puerto Rico and give the whole thing. But I look back at that time like in the 1950s, when there was this belief if you could understand like every movement of a fruit fly that led it to its current point, then you could mathematically extrapolate like what that future looked like, go back and read anything by B.F. skinner.

Speaker A: Skinner would be a great example in psychology of.

Speaker C: Yeah, and the belief was, because it was kind of in the zeitgeist, that if you could do that with psychology and you had a powerful enough computer, you could kind of figure it all out and that everything was not really probabilistic, that it was just a matter of understanding all the variables.

Speaker A: Yeah, that's mechanistic, really. That's just a machine. You just have to know how to set all the dials and you'll be able to predict it.

Speaker C: Yeah, I was this close to predicting confluent's purchase by IBM, but I was predicting $9 billion. To be clear, what I was missing in my algorithm was the $2 billion Tim Berglund print.

Speaker A: That was it.

Speaker C: Yeah, if I'd known you and the team.

Speaker A: Try not to talk about that publicly,

Speaker C: but I'm sure I don't have this on good authority. I have it on, I just made it up right now, but I'm pretty sure it's completely accurate.

Speaker A: That's the best way to know.

Speaker C: So when you talk about that journey of there's the technology side, which is how it has evolved over the last, well, hell, since the dawn of civilization. But in the last hundred years you can argue back to Turing, back to the Dartmouth, back to, I don't know, the invention of the modern semiconductor, the transition from everything as logic and rule based to I don't know, maybe it's like probabilistic. But in addition to the tech journey, you have like the human journey where developers have to just like technology. We have to discover, we have to understand, we have to play, we have to build, then we have to advocate. If it sounds like I'm stealing, that I am Directly, not stealing.

Speaker A: That's not, that's being influenced by. That's, that's something should be.

Speaker C: You're an awesome influencer. Yeah. What is the Reagan line? There's no limit to what you can accomplish if you don't care who gets the credit. There you go. So you've, you've clearly influenced me. So explain that developer journey that I just encapsulated in five words.

Speaker A: Yeah, yeah. So that's, that's what I do these days. For the last 10 years I've been in, in developer relations roles. That's how I, it's kind of the foundation of how I think about the discipline. Um, I like to try to center my thinking on DevRel. And this, this applies to an AI technology. Applies to any technology. Before we were all focused on AI. This is, this is how I thought about things. It encourages you to think about who is out there adopting a technology. There's a developer like you and they encounter some new thing and you have to know what that thing is. So you bump into the, ah, Hermes. That's a nice, timely agent harness. So Hermes. Okay, there's a thing called Hermes. What kind of a thing is it? If I just hear the name, I might not know if it's um, language or some kind of. Is it a new model? A, uh, foundation model? Oh, it's a, it's an agent harness.

Speaker C: Have the gods return? Possibly.

Speaker A: Exactly. And this is heralding some new thing and Hermes shows up, little winged feet. Okay. It's like a coding agent. So I need to understand what it is. I need to have some kind of mental model. Sometimes that's hard. Right. Sometimes it's some new thing that really does require you to think differently and you might need to spend a while with it. Sometimes it's just uh, an analogy and it's, it's like, it's like Uber but for left aligned strings or whatever. It's. But you need, once you encounter it, you need to have a mental model of it. And then if it's interesting, if it's a thing that just seems cool or seems cool and maybe it's a thing you need, we have a tendency to want to play around with it. Right. I. This is beginning to sound a little old fashioned, but I download the quick Start or I pull the docker images and I, I want to go through the readme and, and just see does this thing work the way I think. And up to this point AI can be playing a part in these or not. I could ask an AI to help explain A new technology. I could ask an AI to code me up a sample or I, or not, I could still get something off of a website that's, that's still a thing people do. But after that, after I've played around with it, if it looks satisfactory, then I, I, I might use it. Right? I'm going to build things with it, I'm going to solve my own problems with it. And I, that can be hacking for an afternoon. That can be asking your coding agent to give you a prototype. That can be a whole season of your career. Right? That could be just a long time if you really get into it. Uh, now you're just using the technology. And the goal of the kind of, the primary first goal of a developer relations team is to present that technology to the world in a way that's easy to understand and the people for whom it solves a problem can understand it and then build stuff with it. You want to help people build stuff with it. Some of those people are going to love it so much that they're going to help you advocate for it. They're going to go to that final stage of they want to go give a conference, talk, make a demo, write a blog post. Whatever it is they do, they want to go do that because it's become a part of their personality. And I think in the build stage that's like, that's AI now you're doing that agentically pretty much. And because AI makes content creation so cheap, I think the value of people who are in love with a technology and just want to talk about it as people. Yeah, that, that, that's more important now that we can develop some kind of trust in the human being who is saying this thing's good, you should try it. And that, that could be an asymmetric relationship or a parasocial relationship. If it's a, a person making stuff online, maybe, maybe they don't know you, but you kind of know and trust them. And they're saying this is a good thing. That's more important now because that, that, that could, somebody could have prompted out an AI avatar saying it's a good idea or crapped out a LinkedIn post with emojis in it all over that, that advocates for the technology. So I think the role that people have to play in the process, and I'm, I'm completely on board with every bit of agentic engineering that could be happening along the way there. I'm no longer skeptical of that. Um, but the, the trust we place in people, I think matters. More.

Speaker C: So let's, let's talk about that from an AI standpoint. I didn't expect to talk about AI and love today.

Speaker A: But here we are part of.

Speaker C: But here we are. Let's, let's go down this path. Uh, when you go down that path from discovering until really just being passionately advocating for it. If AI begins to take over some portions of that, like if it takes away the build part because now we're by coding agentically building out what it is. If it takes away the playing because it just jumps you to the point of well, here's exactly what you should do. And now I'm happy to do that for you. Do we see that point of getting them to be an advocate broken because the AI is stepping in and trying to help? Or do we see it accelerating? Or is there, this is where it gets weird. Is there a way to make AI help us love whatever it is we're playing around with?

Speaker A: Yeah, I think that I understand the question to be one of what really is the shape of agentic engineering? What is the activity I am participating in? What is the agent doing? What am I doing? And at the true vibe, coding end of things where I just don't really know. I'm just trying to get it to do stuff. Yes, that process, that tool is going to interfere with me loving a technology and wanting it, making it a part of my personality that's just, that's going to get shortcut or short circuited. I'll give an example. I mean I do a lot of my day job. I'm on teams calls and occasionally they let me out of the box to go give talks. It's great. On the weekend I hack on stuff. I do some firmware and some hardware and plenty of agentic, uh, software engineering just for fun and some of that, like I'll do some, some front end things and I'm, I'm not a great front end person, but I'm fine with typescript and react and whatever's going on, I can be a participant in that. Um, I did a Kotlin backend for something like, I'm fine, I'm here with you. Here's postgres. You know, still hate your cli, but you're the database for me. It's, it's like we're all here. And then I thought I got mad at Lightroom and just mad at Adobe one day and I'm like, what? Screw this. I bet I can build a good enough Lightroom for myself and it'll just be better I've never done an iOS app. Right or not. IOS OS, Mac, Mac OS. Um, can I kind of read some Swift? Yes, sure. But there's a ecosystem of stuff I don't know. And it's more like I was flying blind. And that project has kind of. It's also a little ambitious, but it's kind of stalled because I am more like uh, an actual vibe coder there than an agentically enhanced software engineer. And I really like, I'm glad I started the project because it was just a great lesson to me that um, when I'm more disconnected from the underlying technologies and it's just AI doing stuff like yeah, that's. How am I going to know what to fall in love with? How is something going to impact me some part of that stack when it's really just making things and I'm testing the ui. Uh, compared to. I have an opinion about postgres, I have an opinion about Kafka, you know, maybe my Kafka opinion. Both those opinions are pre AI, but, but still a participant as an engineer with those technologies. And so there, there's this weird thing about us where we decide there's a technology that we just love and it's

Speaker C: a part of our.

Speaker A: Defines our whole personality now they're still going to have an opportunity to find that part of me. Mhm.

Speaker C: Yeah. I think we're struggling with the question you were asking, which is where does AI fit in to this entire journey? And I have to say it gets me so much faster from the idea to the point where I not only have an mvp, but I'm showing it to people. And then not only am I falling in love with it because it wasn't like a month of December drudgery getting from point A to point Z. It was, I had an idea and then I'm up and running with it and then you show it to someone else and they go, I totally get what you're talking about. So you don't have to explain it with your hands and sketch it out on a napkin somewhere, which you were thinking. And that somehow makes it more real. Like you're getting it at that moment where you're excited about it. So maybe that's it. Maybe the middle of that journey in some ways was exciting, but in some ways it was long and maybe it compresses it so you could just stay. I don't know that endorphin hit from um. I have an idea to. I'm rolling it out to look at the cool thing. Let me go tell the world about it.

Speaker A: Yeah, yeah. Of actually creating a usable product. Simple though it might be cost of doing that especially under the subscription pricing and not usage based per token pricing that can get spendy faster but it's so much lower and I just, I love that. I gotta tell you, I think it's wonderful.

Speaker B: Yeah.

Speaker C: Uh, I think we might enter the world not of fewer developers, but possibly of everyone becomes a developer or without having to learn. I don't want to explain to my wife what React is, but she always wants to make a new iOS app and wants it to just be responsive and.

Speaker A: Yeah, yeah, yeah, yeah, you probably don't. If she's not a developer, you don't want to explain. React native. Uh, yes, there are people who are professionals in that and when we, we love them.

Speaker C: Well, so let's talk about to the professional side. As you mentioned, you do once in a while get out of the cage. They let you go public speak. So in 2022 you gave a keynote where you said there is no modern data stack. There are many. Could you explain what you mean by that?

Speaker A: Yeah, that was now four years ago. So inasmuch as I remember, what I was really trying to get at there is that if at the time, and I think the few years before, the modern data stack term was a popular one, and I think what you'd see is somebody would go and give a talk at a conference about the modern data stack and the diagram was the diagram of the thing they had most recently built. So here is a big, complex, well engineered, useful data system, met business goals, got promoted, whatever. There doesn't have to be a single thing bad about it. It's just the one I made and that's the modern data stack and those are the pieces. And the argument I was trying to make was there is a palette of colors, there are a bunch of tools and that tool set itself is a, is a moving target. New tools come in, tools age out, tools try to make it and bubble up for a little while and then it's, it's a, it's a dynamic kind of marketplace, but there's a collection of tools, a general pattern of problems people are trying to solve. Now go put the tools together and solve the problem that you're solving. Uh, in the context that you're in, that's your data stack. There's not like the one modern data stack that is some canonical thing. And the things that people were presenting as canonical were just people doing what I was arguing they should be doing. It's just less Canonical than you might think.

Speaker C: Well that's what I'm thinking is they might firmly believe that we've solved it at any given point.

Speaker A: Because they did.

Speaker C: Because yeah, whatever problem they were running into, they figured out this thing to solve it, not realizing that more problems would then happen. Like you, I get out and speak at a lot of conferences and it always amuses me that every five to ten years someone takes the word data, follows it with a different common noun and declares that to be the modern new way of doing things like engineering. So if you go like way back. Yeah, well, we were like, we started off, oh, we're doing databases, now we're doing data marts, now we're doing data warehouses, then we're doing data lakes, then data lake houses, which I'm still not entirely sure I understand to your point. I'm a data scientist, so maybe I shouldn't even joke about this entire thing. Now we've got agentic AI or either data is no longer important because of AI or the only thing that's valuable now is data. I'm so completely confused. Should uh, the answer just simply be when we're trying to figure out whatever that modern stack is right now, just use the best possible technology, use the easiest technology, maybe don't care about the technology. Is it just let AI do whatever AI wants to do nowadays?

Speaker A: I think, I mean, so no. And by the way, I, I, I remember the first time I encountered Data Lake House and I, I feel like I, you know, I have my mind wrapped around what that means these days. But the first time I saw it I'm like, okay, we're just, we're going too far. And I, I was inspired. I tweeted like, okay, Lake House, how about Lake Pier, Lake Wharf, Lake Jetty, Lake Slough, Lake Isthmus or Data Isthmus? Data Jetty.

Speaker C: You know, I, I think we should have like a data lakehouse with a dock that's big enough to hold three boats. But potentially we're going to need it to have the two story Data Lake House but with a subdivision that's in the next county over just to handle power issues. Of course the name's too long.

Speaker A: Data Data Marina. Yes, let's just Data Bass fishing boat. All makes sense. But I guess to your actual question, I don't think AI solves all of that. I think engineers still build those systems and now we have an extremely helpful tool. So a lot of the question that the modern data stack and sort of my point of view on it, uh, in 2022 is getting at is just one of architecture. There's this company and they have a problem and you want to build a tool that makes that problem less bad and probably has something to do with data. And so now let's look around at what the tools are that exist right now that are in common use. That's good to know. And how do we wire them up and what is the thing that we build to create the results that the business wants. You're making a tool or a platform on which to deliver tools, a tool making tool. And so design that and to understand the components. Sometimes there might be a part of that that you just don't know real well. Like I have to integrate with this weird old database in my mainframe. Well, that's a bad example. There's probably going to be like one vendor solution for that, but you have some part of it that you just don't know well. I mean you can ask AI's questions like that. I'm trying to do this thing. What are the common technologies in use for that? So in my developer journey that's the discover phase. You're asking an AI to help you discover technologies and then as you actually build a system with them, um, um, hopefully the engineers that are executing on that vision are using agentic tools in whatever way is appropriate. But a person is still being the engineer, making the trade offs and again ideally being assisted in whatever way is useful by Model M and harness of their choice. But I don't think AI takes all that away. I think that's just not the nature of the tools that we have now. There isn't a harness where you're going to say build me my data lakehouse in my company. It's too big. Uh, you got to break that down.

Speaker C: Yeah. So we talked briefly at the beginning about kind of AI and its semi modern incarnation back in the 50s. And then we have even the most average person knows about the Turing Test, although they don't know that we passed it a few years ago. At this point that's done. It's kind of passe. So uh, what is that? You've been watching it. What is that through line? From the days of the Turing Test and the Perceptron to what everyone nowadays is saying, I'm going to go ask chat. What is that through line that connects it all?

Speaker A: Well there's a through line of a debate between separate approaches to how we build artificial intelligence. Uh, and those were the symbolic and the connectionist schools. And that emerged right away in the 50s, some researchers, thinkers, they're almost being philosophers at this point. Said we need to think about the symbolic concepts that are involved in our own mental processes through introspection. Let's think about what those, what things are there in the universe and how do they relate and how do we mix and match these things when we do thinking. And the other one was, no, we don't really know. We're not going to build in our ideas like that. We're going to just train a system on a bunch of data that's built more like a brain. And that, that, that really was the first. Those were early neural networks intentionally inspired by what we knew about neurons and those connections. So it was. One was let's build something that's like a brain and train it from data. The other was let's think about how thought works and build systems that do that. And I won't spoil how that debate went. It was twists and turns and people not liking each other.

Speaker C: Ah, did AI kill us all? You'll have to tune in.

Speaker A: You'll have to tune in to find out. We don't know. It's not October yet, but certainly what

Speaker C: we have now, we'll have to wrap the conference up by 5:00 clock on October 10th, because at about 6:01. Well, it's going to be crazy. Let's leave it at that.

Speaker A: No, I think we know now that all of machine learning is neural network based. And so the connectionist approach has triumphed these days. That's not too hard. It's not a spoiler, but that's the through line that I see is the debate between those things. And, um, we right now have this really good connectionist tool. We've been able to scale neural networks in a way and come up with architectures of combining different networks together, that we've got a system that uses language really, really well. It uses language like we do. It's the first time you've got the. We've had Markov chains for a while and 25 years ago there was this funny website called the Postmodernism Generator that would randomly create like a graduate literary theory paper that was gibberish that you could probably get a good grade on if you turned it in. Um, that was all probabilistic and just a fun parlor trick. Now there's a thing you can have a conversation with and it sounds like it acts like a mind. Um, and I think that's what's different. This is the first time since late 22 that we've really had a machine that would do that. And I say machine. My own point of view, to be. To be clear about this, is that it's not a mind. It talks like one, but it's not one. And it's.

Speaker C: Yeah, it's incredibly useful someday. But yeah, yeah, uh, it's interesting. I was looking back over the history of AI and I saw somebody saying, like, oh, Eliza was this key step along the way, which you full well know was basically just taking whatever you said and asking it back to you in the form of a question. And that is not how we think. Right. I can code Eliza in basic, uh, in about 10 minutes.

Speaker A: But, yeah, it was like 400 lines of code was the original Eliza. And, uh, yeah, I talk about that more in the talk and where Eliza fits into my own personal journey there. There is. There is a, uh, part of that story there.

Speaker B: So.

Speaker C: Well, it is interesting. We'll go into it. Some point between now and the end of the day on October 10th. That. Is that a key step, like, in that AI journey? Is it? Ah, just, uh, understanding that a lot of what we say has patterns to it because there is a pattern aspect to that, and it's not a deterministic pattern. If we finished every sentence the same way, life would be freaking boring at that point. What is it? Free will is the ability to stand in the middle of a crowded fire and Yale theater. I think I.

Speaker B: And.

Speaker A: Oh, sorry, go ahead.

Speaker C: No, no, um, after you.

Speaker A: One of my favorite things about what's happening right now is, I m mean, the way software engineering is being revolutionized I think is incredibly exciting. And in my own personal story that started last December, I had a little time off at Christmas, and I played around with cursor about a year before that. I wasn't that impressed. It's kind of flaky. Couldn't make a working Next JS build. Come on, you got one job here. And then not unusual. Uh, December 2025. My Claude code moment. I'm like, okay, this is real. And I love that. Okay. I love what it's doing for engineering. I think it's. I think it's a good. It's a net. It's a net positive. But maybe my even more favorite thing about what LLMs are doing is forcing us to ask questions about what it means to be human. I mean, you just made a joke about free will, but that's. That's a. Interesting question. Now, what does it mean to be a mind or have a mind? Pick your framework and explore that. And is it true that Sonnet 5 isn't mental. There's no mental substance there. And what does it mean that there is one in me? And what even is one? These are suddenly things that seem much more interesting to everybody. And I love those questions. Mhm.

Speaker C: Well, there's a wonderful level of evolution where we maybe get to the point of self reproduction and constantly each generation generating the one after. Not a new concept. It's been theoretical for a while. It makes a big appearance in Hitchhiker's Guide to the Galaxy where the one computer says, I can't tell you the answer or even the question, but I can tell you the computer that will. We'll invent that one to get to some future point. What we have right now are computers that are showing signs. The AI is showing signs that it can build the next generation better. We're seeing the acceleration curve pick up right now. Now between models where it's not a couple of years and it's not even six months, it's speeding up within that. What I've noticed is even within the same model there's some level of evolution. And I don't want to get into is, does it have a mind, does it have a soul, Is it conscious AGI, anything like that. But earlier today I noticed that I use Claude code and I have about 12 agents running at the same time. And I tend to refer to those as, uh, sub agents, or I'll tell one agent there's another agent running. Opus 4.8 earlier today started on its own referring to the other agents as its siblings. And that terminology did not come from me. What's weird about it is that model's been out for a little while and it's never done this to me until today. But now I guess it has siblings that it's paying attention to. And now I'm sure it's in a memory and it'll remember it for forever. But somebody might look at that and think that it's evolving. But in reality, is it more just kind of how the patterns and the weights and all the math ends up playing out?

Speaker A: And is there something in the system prompt that's kind of nudged one little pathway in the weights and this neuron activates a little more because that concept is perfectly sensible. You didn't have to stop and think, what does this alien intellect mean by this? It was immediately obvious what that meant. And it's kind of cute and I like it. I mean, I'll put googly eyes on my 3D printer. I don't object to anthropomorphizing things. But there could be any kind of number of things. I got to hear from a guy over last weekend about explainability in particularly in anthropic and oh man, what a. What a fascinating thing to dig into how one of these incredibly complex machines work and find here's this group of neurons that's associated with sibling relationships and here's when they activate and it's cool.

Speaker C: Well, and people don't realize you brought it up, but maybe it's just a subtle change in a system prompt somewhere and someone added a period or changed a word and it affected the whole thing. When you remember this, when Google went back and looked at why their first image generator was generating non white Nazis, it's because somebody added a little change to a prompt that said the images should be more diverse and it went well. Why are all these people white? Well, because they were Nazis.

Speaker A: Nazis shouldn't be all white.

Speaker C: But yeah, well, yeah, okay, we're getting more diverse about this. Yeah, let's rewrite some history. But in some ways it's more subtle than that. They were talking about GPT, I can't remember, I want to say it was 4.5. Got really sycophantic for a bit because of a two word change to a system prompt which sends it down an entirely different weighting path. But when we actually see it, I don't know, starting to modify its own weights, generating results that truly seem unique, it does make people pause. I can see where an average person might go, do we have a mind? Do we have something actually out here?

Speaker A: And in fairness, again, there has never been a machine that has used language like this. The only things that use language are, as far as we know, other minds. And so to kind of create a theory of mind and think, oh, there's a person in there, as it were, because it talks like one. I think that's an error, right? I don't think that is an actual description of what is the case, but it's, uh, understandable as I have this debate with my wife who thinks a lot about sort of the, uh, for lack of a better term, the kind of the philosophy of mind behind all this. And we have these very different approaches to talking to AI. I am Mr. Nice Guy. I'm like, okay, that was good. Next, let's try it, uh, a little different. Like I'm giving feedback to a person like, okay, thank you for that. I need to tweak it a little. And she's got this, and she's not like this as a person when she's talking to other people. When she's talking to AI though, it's very directive and no, that's wrong. Do it like this. And like. You would never hear her talk that way to a human being. But we just. It's this new category of thing that we have to kind of work out new rules for how we're going to use language with it.

Speaker C: Yeah, I'm flipped around. I, in my normal life, tend to forget to say please and thank you and treat humans like humans, but I always say please and thank you to my AI, which is bizarre. But I have this belief there's a coming revolution. I don't want to be the first against the wall.

Speaker A: I want to be in the logs as that guy. Not that one.

Speaker C: He was good like him. He said please that one time and no one else actually did. What we're running into right now is AI is the greatest productivity enhancer in a short amount of time that I have ever seen. Now, I wasn't there the day they went, hey, maybe we should invent a computer and see how that works. But we. What we're getting is that AI is making the work of generating knowledge dramatically faster. Is it getting dramatically better, though? Or are we just doing, I don't know, mediocre work at a really high speed?

Speaker A: Yeah, good question. And if I knew that, and I don't even. We're kind of seven, eight months into the. We'll call it the Claude Code revolution, where agentic engineering is clearly, uh, a thing everybody has to do. Right? It's just, it's so much faster. You don't. Okay, back up on that for a second. There's projects on the bench behind me. Uh, there are some React and Node kind of network web components, and there are some firmware components. And the firmware components, that's all handcrafted, artisanal, small batch C written by me. And it's going to stay that way. That's like m making your own furniture out of wood. It's. If what you want is furniture, that's economically irrational. If the process of making the thing is kind of the consumption good, then it's great. And that, that's what it is for me. I like it. It's relaxing. But the rest of us, if you're doing it for money at work, it's, it's engineering is agentic now. And I forgot what your question was because that, uh, that, uh, that diversion took too long.

Speaker C: Are we doing knowledge work?

Speaker A: Is it better?

Speaker C: Dramatically faster? Are we just doing knowledge Work dramatically better. Are they completely independent?

Speaker A: Yeah. Let me just think aloud for a minute because I. We're seven months in and I'm not necessarily seeing everything. Like all of the products I use, like Jira is not good. All of a sudden. Teams, I don't, I'm not like teams didn't become nice. Confluence. The wiki is still what it is. Just to complain about a couple of common productivity tools and let that percolate down the stack to whatever kind of SaaS thing is a part of your life. Is it really better? Is it really getting new features? Uh, that's a good challenge because by now it seems like we should be able to, to see that I could talk about my hobby projects and the kind of productivity I've been able to realize. I've built stuff I just never would have built ever. Would never have taken the time. Like I have a family and I have friends and there's other stuff. I have a job. I don't have very much time to screw around at the workbench. But now there's all kinds of code I can build because I got an agent. So I don't think that is just limited to Tim's workbench hacks. I think that is a real reflection of the productivity we're able to realize. And yet at some point there's a where's my flying car? Kind of question that.

Speaker C: This, this and the metric system inside the United States.

Speaker A: Uh, no, I'm not going to give you that one. That's a different podcast.

Speaker C: But. Oh, okay.

Speaker A: It's. But like that, that should, that should be a thing. And maybe it's too soon, maybe I'm just not seeing it, but it's the kind of thing that it seems like we should all see at the same time, which tells us that the kind of pipeline of activities that we've had, the engineering and to some degree the coding, uh, as a subset of engineering, uh, has gotten faster. Like that's gotten cheaper. I think now we're able to see the true costs along the way and the true limiting factor of. It's hard to make things that don't suck. I've got a handwriting capture app. It's not public yet and when it is, I'm sure there will be literally dozens of people who want to use it. I wanted a thing where I could have my handwriting and I love it.

Speaker C: I will be one of your double digit users.

Speaker A: Awesome. Okay, I'll get on the waitlist. It's super cool. It does work. It just needs a Little bit of polish to be productized. Is that a great idea? And actually trying to build a uh, user interface that gets all the right concepts in the right order and in an obvious way. And I, it turns out I'm terrible at that and AI is not that good at that. And so there are basically product design things that um, I could get the code all done but somebody still needs to be good at making this be pleasant to use. And so I think what we're seeing is one of these tasks has gotten cheap and now relatively relative. The others look more expensive. If you were to kind of keep the horizontal or the vertical scale, depending on which way you're doing the bar constant. Now we see what the real costs are. And I don't think that it's still the case that you can't prompt a flying car. It's real hard to make one. And even in the world of the purely immaterial information SaaS app thing, there's still things in there that are hard. Coming up with a good idea. Still hard. Designing a product interface. Still hard. Um, and we get to use more of our minds on those problems. Now that's a win, not m pooh poohing AI. I'm just saying maybe the lack of the so called flying car at this point is a symptom of the fact that there's uh, just a whole bunch of other stuff in there. That this is not a super intelligence that, that figures all that out for us better than we could ever could and gives us a food pellet that we're happy with. That's. That's not the machine that we've made.

Speaker B: Yeah.

Speaker C: It always amuses me how late in civilization it took for us to build a bicycle.

Speaker A: Yeah.

Speaker C: Like a bicycle that you could actually ride with a chain connected to two wheels. It was invented after photography.

Speaker A: Right, Right.

Speaker C: And people don't get it. But it's kind of to your point. Just because you have an idea in your head that it'd be really cool to go that way doesn't mean that you have the tech. Other surrounding technologies have been built up to that point to actually support it.

Speaker A: Because there were metallurgy and machine tool innovations that had to happen to make that work. And somebody had to have the idea.

Speaker C: Yeah. Uh, someone had to figure out that you had to spoke the wheel otherwise the wheel. Because if you're trying to pedal on wagon wheels like the thing's never going to move. Plus you'll get remarkably tired. What do you do if there's no suspension? Yeah. Sometimes Just having an idea doesn't translate it. And sometimes you can translate it and just turns out to not really be as good idea as you thought it was. By M the thing I finally vibe coded. I've wanted to do this for years. This sounds silly. I wanted to use my webcam to tell if I was playing air guitar and play the notes that I was playing. And if I held like this, it'll play the air guitar. If I do this, it'll let me air piano. And actually I've been able to play Ode to Joy on my air piano.

Speaker A: Nice.

Speaker C: And if you do this, you could actually play air drums. I got it working. It's really, really wonderful. And I felt really satisfied and I went, I'm not going to do anything with this idea now. So I've got an application. You talk about your dozen people. It's like, I've now satisfied that itch. I've developed the air band machine at this point, but should it roll out like, I don't know.

Speaker A: Yeah, it's not going to help me know how to play any of those instruments if I air guitar.

Speaker C: What I discovered is when I air guitar, I'm playing totally the wrong notes. And when I think I'm air pianoing and air drumming, it's just like clank, clank, clank. So it doesn't help with that one. But the point being now we've, we've compressed that idea to proving the idea out, to realizing it wasn't quite as much fun as you actually thought it was. We were. We're going to talk just briefly before we wrap up about a couple places you're talking in the future. So I think you're talking at uh, Uber Comp in the next month. What are you talking about?

Speaker A: That is just one little talk. A very kind of for me these days, tactical thing. But I like it on openspec. That's the spectrum development tool that I've gravitated towards and plenty of people have. So giving a talk on Open Spec, there will certainly be live coding. I may just say, hey folks, what do you want to build? Let's build it. Okay, here's how Open Spec works. Um, I'll be speaking at Java Zone in Norway in uh, September and GOTO in Denmark, Copenhagen in. Remember, if it's late September or early October, it's around then. And of course I'll be at Tech Exchange, IBM's big tech conference late in October and current Confluence event based. That's all focused around everything streaming data related in early and I Hope I'm not forgetting anybody.

Speaker C: I will host you down here in Puerto Rico and show you around our beautiful island. October 9th cannot wait. At the Caribbean AI Summit. And it will be a highlight of the conference. I think a lot of people, they always seem to think that what's happening at the moment just occurred out of nowhere and doesn't realize all the things that had to fall into place to do that.

Speaker A: Yeah.

Speaker C: Practical question. Where. Where should people find you? Where should they follow you?

Speaker A: Sure.

Speaker C: Projects you're working on, they should be aware of.

Speaker A: Timberglund.com is a good, uh, place to go. I know maybe having, like, an actual blog is a little old school now, but here I am.

Speaker C: Links pre substack. You're like, retro. I like it.

Speaker A: But you know, also substack. There's a link there to substack and Twitter and LinkedIn and all that stuff. So. TL Bergland on Twitter, I've not yet begun saying X. I just, I find you can rename things.

Speaker C: I think we're like five or 10 years in and it's always X. Formerly known as Twitter, because otherwise I can't figure out what I'm doing on X. Am I tweeting? Well, no, that doesn't make any sense. Am I xing?

Speaker A: That doesn't make any sense. Yeah.

Speaker C: So implies like, I'm crossing something out.

Speaker A: Definitely there. And links to all that. If you go to timberglund.com, that's B, E R, G, L, U N D. Awesome.

Speaker C: Well, thank you so much for your time. You've been a wonderful guest. I will have you on in the future whenever I want somebody to just put it all into context. So thank you.

Speaker B: We asked a lot of good questions in that one. We talk about trust. AI made content cheap, so Tim figures that humans got more valuable. Anybody can prompt out an avatar that says, this product's great for free in about a minute. But the one thing you can't fake is somebody you'd actually believe putting their name on it. We also talked about the mind question. Tim's line was, it talks like a mind, but it's not one. It's the first machine that ever used

Speaker C: language the way we do.

Speaker B: So we're all making up the rules as we go. His wife gives her AI marching orders. She's lovely to humans.

Speaker C: He does swear.

Speaker B: I say please and thank you to mine.

Speaker C: No.

Speaker B: Since being first against the wall when the revolution comes, Talk about the flying car test. If AI were making knowledge work better, Jira would be good by now. It isn't coding got cheap and cheap code showed you that what was expensive all along is still kind of expensive. It's the idea, the design, what Tim calls making things up that don't suck. And, uh, now that's a win. We can find out if those ideas

Speaker C: are any good and then we get

Speaker B: to spend more of our minds on the hard parts. Because sometimes the best person to ask for the the magic is real is

Speaker C: the one who spent 20 years calling

Speaker B: it a trick and then figured out how to explain it all to us in a really nice, easy to understand way. Tim kept saying that's in the talk and I didn't push the full 75 years is his session on the history of AI at the Caribbean AI Summit October 9th and 10th here in San Juan, Puerto Rico. I'm hosting him all around the island and come hear how that story ends. I'm Edward Broski, and until next time, keep asking good questions.