Alan Smith: Welcome back to another episode of the Machine is a Message. I'm one of your co-hosts, Alan Smith, and with me today is...
Patrick Keenan: Hello, I'm Patrick Keenan and I'm excited to dive in to all things tech, AI design, Toronto, et cetera.
Alan Smith: I don't know, that wasn't very convincing, dude. Have you had enough coffee yet today? What's going on?
Patrick Keenan: I've probably had too much coffee. I'm like past peak. But yeah, let's go. Let's go.
Alan Smith: Oh, okay. All right. Let's get fired up. All right. So last time I wanted to talk about a few different things. We talked about product strategy in the AI world. Actually, I have an addendum to that as well. And this is maybe a small thing. And then I also wanted to talk about this. The second thing I want to talk about was the machine is the message and the medium is the message. Marshall McLuhan. And then lastly, I wanted to go through all the links that we didn't have a chance to go through last time. And I looked at a couple of them and I was feeling super freaked out. was like, dude, it's going too fast. I'm freaking out, man. And since then I got a message messages from a bunch of devs that I know who are like, I don't sleep anymore. I just make stuff with cloud code all the time because I can make so much stuff. And so that feels like the energy of the moment. If you're listening to this in a couple of years, that's the vibe that we're living in today.
Patrick Keenan: Yeah, feels like something that's changing every day. the only way you can keep up is if you're unemployed. This is basically where you're
Alan Smith: And actually some of the guys who were said to be stuff were like in between things and just like so stoked to not have a job and to be able to just build their own stuff every day. ⁓ Honestly, not much different than being a developer in the past. Like this just want to deaf, right? Like it's a huge part of it. ⁓ Okay, so first thing I wanted to share was I was thinking about this product idea of, ⁓ okay, so you know when you're like old software, you would have to like log in and then you would click on things and it would kind of like bring you information. was basically like a view on a database, right? We talked about this last time. Old software was for people. New software is people. There's this kind of space in between. Like I was watching, you know, my wife use ChatGPT and I'm stoked for her. She's doing really well. So like, just like finally onto it. I'm seeing what she's doing. I'm like, okay, okay. That's like, that's cool. I never thought of that use case. And she's doing things I wouldn't have expected. And then she closes ChatGPT and it's kind of done. So there's this like, then you think about the open-claw people who are like, yeah, I'm buying as many Mac minis as I can and spinning up like 40,000, you know, open-claw agents to become my army to do all this stuff while I'm sleeping. Like, that's different, right? Like if you and me are working on a team, when we like close a meeting, I'm still working at the company and you're still working at the company, right?
Patrick Keenan: Mm-hmm.
Alan Smith: Whereas with chat GPT, it's like you kind of, or like most of the AI chat tools, there's AI magic there, but it's not an AI agent that's working when you're not there. And I think that's a really big distinction in terms of where people are at today. Like a lot of people just haven't got to using agents that are running long-term beyond just waiting like a couple seconds for a chat here or there, or maybe like a deep research or something on perplexity. What are your thoughts? You noticed this gap too between people?
Patrick Keenan: ⁓ I noticed it in myself. Like I have to move to a place where agents are working when I leave the computer. I have like an open clause set up. My default is to go to a chat box. I think there's kind of like, I think that's one component of it, which is like the chatting piece. And I think what open claw has added is like a couple other components. So like one of them is web hooks or hooks. So like when something happens, I want you to pay attention to it. which is also what somebody does when they're like a human and not just like a computer waiting for you. Like right now it's very reactive. I get a message in my inbox. That's an event. It's a notification for me. I open my, in my inbox, I deal with it. I think the interesting bit about the web hook is like now my agent has an inbox. So that's a new thing and it can respond to that inbox. The other thing is just a heartbeat. So yeah, it's weird to think about like your chat app opening up and
Alan Smith: Mm-hmm. Mm.
Patrick Keenan: saying what's going on right now. But that's how OpenClaw works. It's got like a timer and like depending on how many tokens you want to spend, you can wake it up every half an hour, four times a day. But the notion that it would wake up, be conscious of like the things that are going on, the new threads, like information from the internet, and then choose to act or not based on that ⁓ information at the moment is a new thing. Yeah. And then I think the always on nature of it, which is that Now, when I chat with it, I know these things. So like, I know that now it's going to wake up and I know that it's going to get these web hooks. I'm not there yet. Like I don't, that's not how I use my open claw yet. I'm still like at the point where I'm like, click on a button on the screen and send me a screenshot, but I can see how it fits. Like I still do the chat thing. Still going to have the chat, GBT open up a session, ask a question. ⁓ But to be like, and we were talking about this before, but to have like a theory of mind for how my agent is working, that's going to be learning thing for me, for sure.
Alan Smith: Same here, I've seen other people further ahead on this. And I think it's partially because it's mostly with dev tasks, where you can get this longer term ⁓ runtime, of uninterrupted, hey, initially it was an agent could do something for 10 seconds ⁓ before the task quality would start to break down. And then it was, it's 30 and then it was a minute and then it was five minutes and then it was 10 minutes and now it's like four hours, right? And you project out the curve and it'll be like doing human work weeks unassisted, like go do this really hard task and it's essentially gonna like one shot it or you're gonna be able to, you know, give it some edits and it'll clean it up in a couple hours, you know, at the end of the week, right? And so think about that in terms of as a manager being able to show up and, you know, check in on someone's work. once a week, like that's a pretty high value or like high agency team member, you know, if you don't have to check in on them too often and you can give them like a week's worth of work or like a month's worth of work or a year worth of work, like that, think about it. That's crazy. ⁓ Especially if they can just be getting information from other areas that are kind of like course redirecting them as time goes on. It's like that old, what's that thing about like redirection where If you're flying from New York to LA and you only look at the dashboard like three times, it's going to take you, like you have to do these massive course corrections. Whereas if you're constantly doing like little updates, ⁓ you almost go in a straight line to get there. It feels like there's some metaphor there with like what we're doing today.
Patrick Keenan: Yeah, Drift, guess. Yeah, interesting. Well, I mean, we, yeah, we have that. And so I think that like, you know, chat, GPT 3.5 comes out and everybody's like, wow, this is cool. But like, ⁓ it's hallucinating. Right. And that was like a big thing. Like, ⁓ some of the answers it gives you sound true, but are not. And so like, then we had to adjust and like double check and add in sources and do a bunch of stuff to fix that. I think that same level of Drift is going to happen with these always on models.
Alan Smith: Yeah, Drift is the idea.
Patrick Keenan: But it's going to happen like a week and like you come back in a week and it's like, yeah, I made the dating site that you wanted. You're like, what the hell? was working on a SaaS like project. So there's just going to be like an incredible amount of drift without that like human kind of course correcting and we'll have to like debug a bunch. I haven't like, haven't seen that workflow in action yet. I've seen really good stuff that it works for maybe like 10 minutes and it comes out with a good result. But again, it's done a bunch of stuff where I'm like, you're to have to rework this. So it's something we have to kind of play with, feel the edges, like it will be different types of planning around it. But I think one thing that will be a challenge is if you work for a week unsupervised, that's like a lot of tokens. And there is a point where you're like, I am spending as much tokens as it would be to hire a software engineer. So I kind of wonder like, What is that kind of trade off of like how much human and how much machine should we have involved in any given task? But again, we'll play with it and find it.
Alan Smith: Yeah, I wonder in like the price of a token, is it going to go down? Is the token usage going to improve? Probably, you know, where does that net out? I think we can, I'm going to assume it nets out on the side of like, it's cheaper still to use AI to do most code and having a developer managing a bunch of like kind of like junior AIs. And the real question is how cheap does a junior AI get compared to, you know, a more senior dev who becomes the manager of this team, right? ⁓ So I've been thinking about this idea too, in terms of just like building products and trying to think through like workflow of, okay, if what we're primarily doing, like right now there's a moment and this kind of like loops into, my sort of like next big idea. Let's talk about Marshall McLuhan, right? And you know, the idea that, you he shared with us, we talked about this a little bit last time. But I think we're at this moment with AI where it's like when TV came out and people just put plays on TV, right? Like it was not a moment where we knew what TV was yet. TV became something different. And then the internet showed up and we were putting brochures online, ⁓ phone number, addresses, photos of shops, owners, cat pictures, right? Like it was brochure aware. That was like a whisper of what the internet actually became. And so this idea that the medium is the message, it takes a while using the medium to figure out what the message of that medium is. And I think we're still there, man. We're in the really early days. And so the first thing that we're doing, like our TV, or sorry, our plays on TV, or our brochures on the internet, before we understood what those things are today, I think are just like, taking all the human jobs and workflow and automating all those workflows. Like that's what we're doing. And that's what we think AI is for today. And I think that that's a world changing starting point. Like there is a lot of juice to squeeze in that on its own. Don't get me wrong. It solves some massive problems. It creates some real value. It may create some massive problems and it may destroy some real value as well. But I think that we're at that moment and it's still only a sliver.
Patrick Keenan: Which is to say, which is to say like when somebody put like a theater show on the TV, that was awesome. Like people at that moment were like, this is nuts. I'm not at the shit. Like this is so much cheaper than going to theater and this is like in London and I can see it. It was just like mind blowing and so exciting. And so clearly like, wow, this is the power of this medium. Little did they know.
Alan Smith: you Yes, they were getting just like a hint of what it was actually, right? And if that pattern is true for like what we're feeling already today, it's like, dude, like we're just gonna be like, like just like, it's gonna be lot of right. So I've been thinking about this and okay, so what is it right about where we're going next? What is this medium of AI really wanna become? And I don't think we can get there from a whiteboard. ⁓ I don't think we can get there from a podcast kind of like pontificating on what it's going to be. We're not going to predict it from a distance. We can only discover it by building, by being in the products, by trying to make stuff and letting it reveal itself to us kind of like piece by piece. And if McLuhan said the medium is the message, then I think you know, the motion and the creation reveals the message over time. You know, this is the idea that clarity follows action and direction comes from movement. So I think that, you know, our job as we build these things isn't to forecast like kind of like the final form of what AI is going to be with precision. It's to just jump into the meaning, jump into the medium, work with the best idea we have today, which is essentially taking human labor and trying to automate it, which again, like may create some serious problems. But yeah, basically like, you know, where we begin isn't going to be where we end up on this, but we just need to begin. Like we need to start, we need to be working on stuff. And I think one iteration at a time, one customer conversation at a time, one signal at a time, that's how we get there. We can't see around the corner. We just got to walk around it. You know what I'm saying?
Patrick Keenan: Yes, that this stuff is discovered through building. And I think you can see that at the model companies, right? So like the model companies ⁓ are all scrambling for use cases. And like you can see Anthropic doubling down like on the coding thing. And that's how they got where they were, which was the right intuition, right? And now all the other model companies are like, coding unlocks all these other possibilities. We shouldn't have started with like X, Y or Z. ⁓ We got to catch up with Anthropic. ⁓ So
Alan Smith: Mm-hmm.
Patrick Keenan: Yeah, the question is like, where do you start? ⁓ And that's from the model perspective. But then on top of the model, you have applications, right? And you're building out some kind of a wrapper, I guess you could say, or front end for this model. And the decisions you make there are going to like, yeah, there's like a path dependency where you're going to make some decisions that are going to fit you into a corner. And so I think every, every company out there that is thinking about AI is like, ⁓ we can be the wrapper. for this thing, like we'll be like the front end. What, like if you were starting from nothing, because like if you start with an existing business, you're like, obviously I'm going make it work for, you know, I'll make it work for construction workers. I'm working on like city services or I'll make it work for her salesmen. But if you were starting from nothing and you were saying like, how can I explore this new medium in a way that I will more likely reveal the inherent
Alan Smith: Yeah.
Patrick Keenan: benefits of it, like in the same way, like with a TV, you do closeups. Like that was a really big deal. It's like, ⁓ you can do a closeup. That's crazy. We never thought about that before. like jump cuts or like fades or like establishing shots. Like all this kind of language had to be established. I think that's what me and you get excited about is like, what are the new patterns and pieces of the language that need to be established to like really push this to the next level? ⁓ Where, what use case would you start with if you were, if you wanted to explore that stuff? Or how would you start on that?
Alan Smith: I don't know if it comes so directly, like the connection between the end of where we're going to end up and what we're doing in the short term. In the short term, in terms of where to start, I think that's a really interesting question for every small business, medium business, large business to do is to say, okay, what can we automate here, right? ⁓ One approach to that, and there's different strategies, so different people have different AI strategies, right? One approach is you say, okay, well, what's the thing that we spend the most money on today and ⁓ is the most automatable? So you start to say, okay, well, do you need a body to do this? If it's keyboard, video, and mouse, then you can just do it. ⁓ And we have a lot of people doing this, so it's going to be expensive. Let's say you have a call center. Your number one expense is the people sitting there answering calls. So you start to say, okay, well, can we build an agent that can answer the calls? Right? No brainer. Let's say that you're ⁓ a lawyer and a huge amount of your staff is juniors who are paid to stay up all night reviewing case docs. It's like ⁓ to prepare something for you for the next day. It's like, well, it's not really efficient. Are we getting the best value here? This is a good use case for AI. So you start to put it in there. And I think technically there's generally two places to add it. Either you go front of house and kind of make it more, this is the interface with the customer. This is, ⁓ or back of house, like bookkeeping and stuff. Or delivery, right? So you've got kind of like sales and marketing, ops, delivery. And if you break up those like few pieces, delivery is always like, what are you as a business? Are you an agency? Are you a ⁓ construction company? What is the thing that you do and that you sell as a core? And then there's all these enabling functions around that essentially. So in my experience, going towards the core works when the core is really expensive to deliver and highly automatable keyboard, video mouse, Are the core ways in which you actually deliver that. Otherwise you say, okay, well, let's go. Where's the bottleneck in the system? So that's like strategy number two. First one. What's expensive? Let's do that. What do we spend a lot of time doing and what's expensive? Second one, where's the bottleneck in the system? Are we demand constrained or are we supply constrained? So that's option number two. Option number three is just like, well, what is it that we're not doing today that we wish we could be doing? Let's add that to the business. Let's let the rest of the business run the way it should, but let's do the thing that we're not doing. I'll give you an example. We have this AI that we're building for plumbers and like all different home services businesses. All of them have a task, which is after the job is done and the technician leaves your house, they want to call up and say, hey, how did they do? You know, did they take off their shoes? Did everything feel like it was fixed? Did it feel like it was friendly? Did they leave the place clean? Would you leave us a Google review? You know, could I ask you to give us a testimonial here on the phone? I'll type it in for you, et cetera. ⁓ That's called a happy call. Nobody does happy calls because they're just too busy. They can never get around to it. So now you've got this use case of, okay, we could do that now. It's a good part of the business. The rest of the business is kind of running okay. Like we're not like dying in the rest of it. Let's just add this and then lift the whole value of the business in some way overall. So those are the three ways that I know in terms of deciding what to do next. What's expensive? Where's the bottleneck? And then just what aren't we doing that we would like to be doing that we might be able to automate and add in?
Patrick Keenan: I guess with all three of those, you're kind of like looking at an existing industry where maybe those things are like more quantifiable, right? So like you can look at a business, you can look at his P and L and you can be like, this is the costs. Um, that makes sense. I think, you know, I guess going back to the metaphor of TV, um, I'm curious now, like was there a director who was like, I mean, there was, what's the Rosebud movie, Citizen Kane? Like that was like, yeah, that was a really big breakthrough in terms of filmmaking. And he chose to do it about like, kind of like this, you know, this historical robber baron kind of guy, Hurst, I guess. And so it was a kind of a...
Alan Smith: ⁓ Susan K. and Orson Welles. Mm-hmm.
Patrick Keenan: a dramatic movie. You know, that's where they pushed the medium forward. I guess with brochureware, it's like what were the first companies to adopt like the internet? ⁓ don't know what were they, just nerds. I don't know.
Alan Smith: It was nerds, man. It was nerds. But then everybody needed a website all of a sudden. It was like you had to get your like GeoCities website and every small business like just like local businesses would be like setting up a website and then they would add like a www to their business card. Like remember that? Like that time when that was like a thing where it's like, hey, we have a website. What's on your website? ⁓ nothing really. It's our hours. Yeah, pretty much. There's like...
Patrick Keenan: Mm-hmm. The same stuff as the business card. Yeah.
Alan Smith: a carousel of pictures you can scroll through it. Whoa, cool.
Patrick Keenan: But then, then people started building web applications, which is like the Amazon or the Facebook or whatever, Google.
Alan Smith: Exactly. And we just didn't know that that's what it was going to be. The websites were going to be more, or the internet was going to be more about applications than it was about static messages. Right? And I think the same thing where I think we're looking at AI a lot now in terms of its intelligence, and let's take that intelligence and just apply it to human workflows. It's like, okay, well, that's one thing we could do, but there might be something, there will be something different, something much deeper than that.
Patrick Keenan: Mm.
Alan Smith: as opposed to just adding a bunch of synthetic humans to the workforce. Like, don't get me wrong, we're doing that, that is happening. There's an industry, there are tens of thousands of companies being born on that and hundreds of thousands of jobs are gonna be lost on that one idea. That's not where we're gonna end up though. That's not the end of the story.
Patrick Keenan: Right. Right. And I think that's what we're seeking out on this podcast. So if you have any ideas, please let us know and hit us up for what you think the future metaphor will be and what what AI is truly meant to do.
Alan Smith: Take a guess, take a wild guess. ⁓ I'm not gonna take a guess here today. Let's talk about some of the stuff that's going on in the world more specifically. What are people thinking about? What are people writing about? What are you seeing? Why don't you share some links with us?
Patrick Keenan: Yes. Why don't I share some links and these will all be ⁓ in the notes. So I'll share this window here. Okay, so let's start here. ⁓ this was a kind of tweet that I thought was really good. And I've been thinking about this since, I don't know, ⁓ two days ago. So my biggest current workflow. for making iOS apps is extracting out all my core functionality into a standalone Swift package, which is to say into just functions, no UI. Making it into a CLI, which is the command line interface, giving Cloud access to that CLI and going to town. If you do any kind of on-device ML or have lots of headless services, this is extremely helpful and a better way than just waiting for rebuilds. So this is like a very ⁓ technical way to say like, just want to test it with a CLI, but I think there's something else here, which is maybe the CLI is the future. So what we're seeing is like, I guess a couple of CLIs is the big one that, you know, is the breakthrough is Claude. Now, like everybody who's never touched a terminal is using ⁓ Claude CLI, which is you just type in Claude and then there's a bunch of magic words you can say to Claude via the CLI, right? So
Alan Smith: Mm-hmm.
Patrick Keenan: That's all it feels like 1970s computing, right? But what's interesting about it is that in a world where ⁓ computers ultimately run on a terminal and the fastest way to get something, whether that's a URL or like run a process, like convert this set of images to a video or like basically anything you do on the computer, the fastest way to do that is in a terminal. AI knows all the commands and then this is perhaps the base on where your UI is derived from. So like maybe your UI is just like built on the fly based on the CLI that you're using. So that was like one ⁓ thought and it kind of brings me back to what we were just talking about, which is like how do you kind of design for this stuff or like what is it meant to be used for? Like maybe this whole internet UI ⁓ path was just like a side quest. Like maybe all the UIs that we're seeing on the internet are just kind of like. A silly side quest will look back on and be like, everybody made their own UI. This is weird. Like look at all these old documents and people put so much effort into these like business cards and stuff. Why didn't they just use, you know, NFC? So I think it'll be like that. That's one world where, where you don't make custom UI. Instead, it's like built for you on the fly using these command line interfaces. And as a developer or dev or designer,
Alan Smith: ⁓
Patrick Keenan: you're actually just defining what functions you want to provide to people. Like I want to provide the ability to respond to appointments and book appointments. I'd like to get these fields and then when they're done, I want to send them over here to this human to check that the appointment is on the right calendar or something. ⁓ So yeah, that was one thought.
Alan Smith: So I love the term headless. Let's define this in this context a little bit. ⁓ You have lots of headless services on device ML. What's the context here for those use cases? Or just help me understand what Tyler means when he talks about that. Because I'm very interested in especially headless services.
Patrick Keenan: Yeah, I the service. It means a service without UI. it means so like you can run Chrome, for example, ⁓ in the background and you can have it like click on things, but it's never actually rendering to a screen. So headless means like I'm running this service and I just need the output from the service. So yeah, it would like in this example on your phone, you might take a screen like a photo. And it's like, find if there's like a hot dog in here. And so now it goes to this headless service. That's going to like take the image, do some edge detecting, like put that into an algorithm, figure out if it's a hot dog or not, and then spit it back. Right. And so like that's the headless service is like the hot dog or not hot dog. ⁓ and there's all kinds of stuff like that on your phone. The classic challenge with engineering is like the UI always gets mixed up with these like backend services, but
Alan Smith: Mmm.
Patrick Keenan: And so what has happened is like, we've kind of like gone to a UI first development approach, which is like, show me the flows, like make up the mocks. Like I want to see where the buttons are. How does the login screen work? What if I need to reset my password? We design all those things and then we implement it. But perhaps the future is just like, you actually just define what your CLI is going to do. And then you tell the AI to make a UI.
Alan Smith: Yeah. On the fly, yeah. So actually this is really, really interesting. Okay, dude, I'm feeling this. This is pretty exciting. So something I've been complaining about a lot recently is internally. So we're building 10 products at a time, right? Like every week, I get to look at 10 products and how they're kind of progressing. And the one thing I'm noticing across all of them is they feel like old software to some extent, right? Like we all know everybody on the team is well aware of where we want to go. We want to move towards this, this magic, you know, want type scenario where, you know, it's the team member that's working with you and you know, your team member doesn't necessarily have a dashboard or like, you know, a sidebar menu, right? Like that's just, those are parts of old software. Like those are old, like artifacts that don't necessarily apply to what we're trying to do in the future. and to try to show some of these ideas to a customer without building the whole thing and actually making all the functionality and the AI work, which is still like a non-zero cost, right? Like for some of the stuff we're trying to do, like we would need access to data and then we'd have to like run it a bunch of times. Like it wouldn't just work. like, these use cases aren't one shotable, quote unquote, yet. ⁓ So what I've noticed is that it feels like a lot of the demos that we're giving feel like old software. They don't feel like this future approach. a lot of like, if you use Replet, it pushes out kind of these old interfaces. And it was when you use Cloud Code, it's trained on that. So it's giving you these interfaces. Yet I feel like we're being pulled towards something new. And every time something comes out, I'm like, Do we need this screen? Can we strip this away somehow? Can we just get back to what's the thing they're asking for and how do we get it to them quicker? How could there be less ⁓ visible here? How could this feel more like it's a shop and you parked your car in it and then when you got your car it was fixed? You weren't following an interface that gave you a list of everything that was happening while it was happening and trying to direct the mechanic on what they were doing, right? Like this is the people want the outcome like you said. And so sorry to try to bring my point down. I'm kind of rambling a little bit. It feels like none of us are quite sure from an interface perspective how to come up with software ideas and demo them and validate them in a really cheap way without actually building the whole software yet in this kind of new style. And also what does the new style look like? Like there's kind of like two or three moving targets here in.
Patrick Keenan: in
Alan Smith: product development for me today.
Patrick Keenan: Yeah. And I think that one, ⁓ like misstep for me is like, maybe the future is not more UI, right? Like maybe it's not like a cooler UI. Like I love the Apple glass thing, but like, maybe that's not the future. that I'm maybe that's part of the future, but like, there's just this other world I I'm thinking about with this CLI conversation is like, I'm seeing people use command line interfaces and I'm thinking. this should be an auto-complete. I shouldn't be typing it, I should be talking it. And I should know what the inputs could be because somebody's telling me. And those fields should just manifest, they should just appear. And so I think that there's two business models that interfere with the step change, which means that it's an opportunity for people starting out. So one is, if you're not looking at it, I can't show you ads.
Alan Smith: Yeah.
Patrick Keenan: Right? Like most of the internet is funded by ads. But if you're using a command line interface, it's very hard to put an ad into that thing because I do a command and I expect an output. And like maybe there's an ad appended to the video that I processed or something, but it's much harder than just like showing you something because our whole internet is designed to like, to show ads before stuff. So that's one piece. The other thing is if you're using a CLI,
Alan Smith: Yeah.
Patrick Keenan: I can't really charge you. Like it's harder to do charity. You can do authenticated sessions and stuff, but ⁓ it's kind of harder to charge people for something that's a command line because it's very easy for me to remake your product. So ⁓ I have to look this up. So Figma ⁓ made in a weekend. There was this dude, this is a funny story. ⁓
Alan Smith: No, no, no, no chance.
Patrick Keenan: I don't know if I can find it. man, blocked. There was this dude who ⁓ had a plugin for Figma and it would take your Figma code and it would like build it into React and Tailwind and stuff. ⁓ But then Figma blocked him. ⁓ And then he was like, cool, you blocked my plugin, I'll just remake Figma. So he, in a weekend, he just like...
Alan Smith: ⁓ my god.
Patrick Keenan: Remade figma. I'll find the link and Yeah, this he built his own the figma in three days. So I haven't used this ⁓ But ⁓ it's open pencil. Okay, cool. So this relates to the other topics I want to talk about but yeah, this is the response So that's what I mean about the CLI is like, yeah cool. You're selling a CLI I Can read first engineer that in a weekend because there's no UI in it. It's not that complex, right? It's like text in text out or binary in and binary out. So those are like the two business models that are, think, preventing us from thinking about what the future might be because we need people to use these UIs, right? We need people to like log into these UIs.
Alan Smith: Mm. First, that's wild. Like the fact that the dude rebuilt Figma and you I remember a couple years ago, us talking about Figma and being like, this is such an engineering masterpiece that could only be done by a company that was founded by like, you know, a 0.01 % engineer. Like that's what this is. And the fact that now, that engineer is available to all of us through clogged code, essentially, like he's on all of our team. Like that is no longer an advantage to be a 0.001 % engineer. And like when they did that, like Figma, that was like... the entire Adobe Creative Suite or at least a good portion of it that was relevant to essentially like UI and web design first, I guess, right? Like the fact that they could take that and put that into a browser, I just remember being flabbergasted and that that was like a thing and how much money they must have spent to be able to get that done. And the fact that one guy could do that in a weekend. Dude, what timeline is this?
Patrick Keenan: mean, and it's really fast. Like, and are there collaborators? I don't even know. ⁓ It's so nuts. But I just love the checks, because I love the checks and balances of technology, which is if it sucks, and you're being a gatekeeper, I will just remake it. Like, I just love that check on things. And so this is like a check.
Alan Smith: It's like pure market forces in a way, right? Like, you know, lot of the, we talk about this in terms of moats a lot and like, you know, how can you build some sort of moat and how can you build some sort of competitive advantage? And I think that you shared a link or, you know, I saw one recently, which was around, you know, there used to be two product strategies. One was first, be first. The second one was be better. And it's like, there's no point in being first anymore because anybody could copy you at any given time.
Patrick Keenan: Yeah.
Alan Smith: And so you just need to be really, really good. Like that's the only product strategy moving forward is you have to understand what people want and then you have to give it to them as well as possible. Now I'm going to push back on that a little bit. I think that that's a little bit myopic and that comes from a product person. If you're thinking about the business overall, you have a lot more levers to pull in terms of how to be a better business. You have a brand, have customer service, you have distribution, right? Like there's all these components that product people sometimes forget about. I think that he's probably right just in terms of the product strategy piece of your product needs to be better. But as far as business strategy, it's maybe a little myopic.
Patrick Keenan: mean, so notably he made Figma. I think he was like the main designer that made Figma. ⁓ And I think, yeah, it is a product mindset. Well, would you say that brand and marketing are part of product or is that just like a product-centric way to view the world?
Alan Smith: It's definitely product-centric way to view it, but everybody stands where they stand, right? So, know, accounting is gonna say, well, product and brand are part of accounting. So, it's like, we add up the numbers without you guys, or without us, there's no business. It's like, oh, well, it's kind of true, right? So, everybody stands where they stand. are points of view. I think that brand is what people think about you when you're not in the room, and do they think that you're...
Patrick Keenan: Yeah. Mm-hmm.
Alan Smith: really interesting and great and they have loyalty to you or do they think that you're completely replaceable and they're ready to you know switch to anyone else at the drop of a hat you know there's there's aspects to that on its own the product can of course deliver something on that brand like it can be a part of the brand building component or the brand is built in part by you know the product if you want to put that first at the end of the day it doesn't matter they're They're very intertwined. They're part of the same thing.
Patrick Keenan: Mm-hmm. Yeah, but it's, it is a timely piece of advice. There, there used to be two playbots for commercial software. ⁓ so that's good. He frames it as commercial software. Be first to market or make the best product. Being the first was important yet so many software companies operate this way. We must ship by this time or next month or we'll lose. ⁓ is this shallow way to build now with the AI tools we've gone from LOL, nice JavaScript try again, ⁓ to damn it's nice C program. take the wheel. So I think he, yeah, he's basically saying the, the ability to go to the mark first is like, now everybody can do that. So like, I think the point here is that there was two strategies. Now there's kind of one because going to first to market is not going to put you ahead of your competitors because everybody can be first to market. So like, what are you going to spend time on? Yeah. What do think about that?
Alan Smith: Again, in the product centric perspective, I think he's right. I think you do need to have a really, really great product. The bar is always going to be high, and people are going to try to work around whatever else you have in terms of a moat. So there's this war for context that we talked about last time. How do you add more context to a product? ⁓ You have a lot of, if you're in a position where you can aggregate contacts. So let's say that I'm running a business that is front desk and I have access to all of the docs and all of the systems and I have unique access to just all of the phone calls, right? I have contacts that maybe some of the other systems don't have. And so I hold onto that, which allows me to be more useful and more like put my intelligence to more work, right? is going from you took the smartest person in the room and you put them in a black box and you ask them to solve a problem versus you help them understand complete situational awareness and what's going on in a business, right? So, sorry, I don't mean to get off track here. What I'm trying to say is that context is so important to being able to build a good product. And oftentimes that comes from how you position what you're going to be doing, that there are certain tasks that you can do. that if you do those, maybe in addition to the thing you were trying to do originally, you can collect more context. Therefore, you can make the original thing you were doing even better and build up a bit of competitive advantage there. And if you're doing that and someone else can't do it, then they can't do some of the other things that you could start to do because you hold that context, right? So it gives you a competitive advantage on the product development itself when you have that context. So I think it... context is really, really important here. It's not just about developing stuff. It's about where does it fit into the business and what do you have access to to be able to add value to these businesses. And I'm generally thinking business to business here, but I think the same concept probably applies with B2C products where if you're the AI pin that theoretically OpenAI and Johnnie I've are developing that just records everything I'm doing, and has ultimate context into everything in my life. Like it's heard and seen everything I've heard and seen. Like how much more context do you want? Right? Like you're going to be able to build whatever you want from having that.
Patrick Keenan: Yeah. I think that's right. ⁓ Increasingly, however, I think that's about now that I have options, who do I want to give that context to? So with that said, this is a link which is switched to Claude without starting over. now, like what we thought was the the moat ⁓ is now a commodity. So I can import from Gemini, I can import from Chachbt, I can import from, you know,
Alan Smith: Oooooh!
Patrick Keenan: any other service and now all that context comes with me, right? Which is a good move for consumers. ⁓ But ⁓ to your point about like the special context you have, you're going to have to think about like, is it context that's not the users? So like you could imagine I have an inventory of a bunch of electronical components and like that's my context is that I have this industry knowledge that you can't just export. But increasingly, I think
Alan Smith: Nice.
Patrick Keenan: anything AI can touch will be importable. I think it's just an interesting move is like the portability between these chatbots.
Alan Smith: We see so much, especially in, there's a lot of existing systems, like especially systems of record. Like if you pick any industry, right, like pick dental, you know, there's a bunch of systems of record there and they're all very protective of their data because they know as soon as they start letting this stuff flow everywhere, they're cropped, right? Like somebody's going to take it all out and stop using their software because it sucks. Like, because that's what I would do, right? Like I don't want to use this crap. It's like, let's build. a better system of record and they're like, nah, we're just going to kind of hold on to everything. Or we're not going to play nice. We're not going to allow this data out and give you access to it. Epic, all these like healthcare, ⁓ are they? ⁓ EMRs or EHRs, like ⁓ electronic medical records or electronic health records. These systems, and then look at anywhere else like SAP, there's always these systems of record. that suck, they're so old, they're legacy, and they've built these walled gardens around the data, right? Now we're seeing kind of like a system of intelligence being built on top. And that's where the opportunity where a lot of people are jumping in of being like, well, I'm not gonna be able to get this data out, but I can via API work with the data to accomplish a lot of things. And this idea of the system of intelligence is one that we're looking at a lot of, okay, well, that's a door until... all this data starts to flow everywhere, or you can rebuild it just by being the system of intelligence on top for long enough, right? Where you've just, you've seen it all at some point in time.
Patrick Keenan: Yeah. Or you could just run this prompt. Like I think it's funny that this whole product is actually just this prompt you put in here that says, I'm moving to another service. I need to export all my data, list all the memory stored about me, add dates to it. So like this is like the exploit ⁓ is just put that into your thing and then paste it into cloud. So these like the, again, the barrier, the technical barrier to do things is so much lower, which means you have to work so much harder. to have people want their context in your system or want to use your system ⁓ on a daily. ⁓ So with that, mean, I know we've got a lot of links and low on time, but I want to like run through some options of ⁓ Figma alternatives. So we looked at Open Pencil, which was made in a weekend. ⁓ There is also this one called pencil.dev, which is ⁓ a live canvas that like
Alan Smith: Mm.
Patrick Keenan: As you're building, it's also making the code. This is this, this by the way, is like my dream. Like Framer, Figma, I always hoped that they would do this, but they haven't done it because everybody's got to focus. But this is like my dream is like the thing I'm designing is code.
Alan Smith: Yeah. Dude, very first job out of school, I remember at blast radius, I was making mock-ups in Photoshop and I was like, this is for editing photos. Why can't I just make the website like here, right? Like this is so silly. And then it had to go to these HTML, like I had to slice up all these images and I had to send it to the HTML guys. And the HTML guys had to write all this code and then they had to like go check it in the browser and it was such a mess. And I'm glad that all that workflow has collapsed now too.
Patrick Keenan: Hahaha
Alan Smith: I have a few different ideas and what I love about pencil is it gives you essentially design alternatives at the same time. I think that there's a gap here for me in terms of brand of saying, well, like, look, you're just giving me a bunch of different ideas maybe for new projects where you haven't defined that stuff, it's good. ⁓ But the idea of, it's great, it's great. I love it, it's great. I'm not going to critique it. ⁓ This is a dream come true, we're there.
Patrick Keenan: Yeah. To like, ⁓ full disclosure, I haven't tried this yet. I mean, I've opened it up and pushed stuff around, but I haven't actually tried it for anything. ⁓ But again, it's like easy import port, like design as code. This is the future. So here's another one. I don't know if you've heard of this one, paper.design. So it's a similar thing.
Alan Smith: I like one is called pencil, the other one is called paper.
Patrick Keenan: And the other one's called open pencil. it's like, yeah, we're stuck in a metaphor. ⁓ So yeah, it's a similar thing. It's just kind of like, it's you, got your, you got your code buddy ⁓ at your side. So it's this one, I think is much more like open source ⁓ and like integrated with the AI terminal. But again, just like a different answer to how can I see the thing I built and then build the thing and like have that go back into AI. And then I just go back to our point about the CLI, like is the future move, like it's becoming so easy to make these boxes and arrows. Like, should we even be doing that? Like I know that there is like a, an attention to the detail there, but AI also has like a good attention to detail. Like will we be making, will we be pushing boxes around? I don't know. Or is this like the temporary moving around ⁓ chairs on the Titanic? I don't know.
Alan Smith: I also don't know. I struggled between going between, all we need is a command line interface and all we need is outputs. I don't know if I'm sold on the idea that the easiest way to interact with something is the command line interface. It's like saying, you know, for books, all we needed was text end to end on every page. Like that is one specific interface and it doesn't necessarily do everything we need it to do. I'm, without getting into it too far, I'm cautiously optimistic that we'll get more on the fly interfaces that do what we want when we want it and don't need it if we, if we can exert our preference here ⁓ as we move forward is what I hope.
Patrick Keenan: Yep. Yep. And it's the, we'll only find out by, ⁓ by playing with stuff. How much time we got left here?
Alan Smith: We're at 48 minutes, let's do 10. We got 10 left, so you got eight links, a minute each or something.
Patrick Keenan: Okay. Yeah, so we spoke about ⁓ design tools, talking about CLIs, talked about like future craft and software. This is another ⁓ chain in our Darwinian evolution of AI, which is like maybe it's hardware. So this ⁓ is looking at ⁓ an assistant that runs OpenClaw on an ESP, basically an open, or no, this is a Pi Zero. So like pretty cheap piece of hardware.
Alan Smith: Yeah.
Patrick Keenan: and you just talk to it. like maybe, and it gives you some visual feedback, but maybe the future is actually much cheaper and smaller than we think it is. Like maybe we won't be using these computers. Maybe it is like just like a really cheap commodified piece of hardware that connects me with the agent I want to be talking to. So that's one thing.
Alan Smith: Mmm. I think that that is, this is shades of what we can expect from Johnny Ive, OpenAI, always on assistant type thing. I think that this is like. Somebody with way less money, way less time, just giving it a shot. Like, hey, I've kind of got this idea. And then the idea is also being channeled through the ether, through that massive team with like $70 trillion to go build it. And it's just like, these are the two outcomes from the same idea that's just coming from the ether of the world through two different conduits at the same time, like how electricity was invented by. multiple people at the same time, like Edison and Tesla, right? I think relativity was even potentially found almost at the same time by another guy as Einstein. Evolution was discovered twice, right? So I think to me, this feels like a second idea, or sorry, this is one idea, and we're gonna see this expressed multiple times in the next X number of years, and someone's gonna get credit.
Patrick Keenan: Hmm. Yeah. And part of it is compounding factors, which is like the more AI you have, the more you can use AI to design stuff cheaper, faster. And the more you do that, the more AI you have. So I think there's that, but this is kind of like the hardware angle on things. This is somebody who used RayBan Meta and OpenClaw to like look around the world and route their vision into their OpenClaw. So, you know, we're seeing, so that's that's a hardware piece. And then this is open claw, but.
Alan Smith: So sorry, is it just building a ton of context then? Like everything that this person sees is like recording in some way? Like what's going on? What's the use case?
Patrick Keenan: So it's just kind of like having this AI companion that can see what's in your glasses. So there's an SDK you can use and you can see, you can take a photo. And so, yeah, it's like just taking photos and hands-free kind of way to do it, as well as there's a screen there. And yeah, you can kind of chat with it on WhatsApp on your glasses and it can like look at what you're doing. So that's kind of, that is like the wearable version of this thing.
Alan Smith: Okay. All right. Rad.
Patrick Keenan: And then this is controlling ⁓ old phones with open clause. So this guy ⁓ is able to turn the flashlight on or just do different things on the phone, turn the volume out, vibrate. And it's got root level access to these phones. You can see it's navigating itself. this ⁓ is kind of a pattern you would use as an automated test suite for many phones. You could see it in a lab where you're just testing this. app on a bunch of stuff. think in some of the comments there might be some of that. But like this kind of thing where you're like, yeah, maybe it's for testing. Maybe it's for like just bot farms. But ⁓ the interesting bit is like maybe this becomes kind of a consumer approach, which is like, I just want to control all my devices without touching them. ⁓ And you can just control your phone from afar. Yeah. So so that's like these are alternate
Alan Smith: Yeah. awesome.
Patrick Keenan: futures, you know, like, what does it look like? Will we be controlling our phones with tappity taps?
Alan Smith: Well, the second big idea in here is like computers learning how to use existing interfaces, right? So it's showing us what's happening on the interface, but is it working? Like if it opens up an app, it's using the app, not like the CLI and like function level under the UI, right? Like it's working through the UI. Okay, so that's cool too. And like, this is, think something that people forget about is, you know,
Patrick Keenan: Correct. Yeah. Yeah. Yeah.
Alan Smith: as devs do we need to build something that connects to every API and only works in that route? It's like, no, no, no, there's another route, which is all the interfaces we built already. Just teach the computer how to use those and click through them all. ⁓ That's a track. There's wild bot farming stuff going on, especially the top of funnel MarTech stack right now is so messed up. Dude, it's LinkedIn. Like X all these things are like turning into digital wastelands because of the amount of bots that are on there in the incentive structure to be able to just try to grab attention and then funnel a small portion of that attention towards whatever it is that you're selling that's related to what that person is so it to me it's like I Don't know where that's gonna end up like what the end game is there like it's a constant battle like I've heard about these guys who rent hundreds of thousands of computers to build essentially or to sort of like click on a bunch of LinkedIn things to just be able to extract all the data out of LinkedIn and then build lists of people to essentially market to, right? ⁓ And that's LinkedIn's data. And they're just like using hundreds of thousands of bots and computers like giant botnets to just get it out. Right. Like we talked about extracting data. It's like, well, we're just going to steal it. ⁓ Nobody's.
Patrick Keenan: and Yeah.
Alan Smith: We're going to make it so they don't really know and we're going to make these things really good at clicking on LinkedIn the way a human would and just extract all this info. ⁓ So anyway, I think that's hilarious. And computers learning how to use interfaces that were built for humans is like a funny moment. It's funny.
Patrick Keenan: Mm-hmm. Mm-hmm. Yeah, it is like a side journey. It's like, okay, cool. Now we can go back to using the command line. I think the other piece is like, I don't know if you've heard of YouTube DLP.
Alan Smith: Nah, what's that?
Patrick Keenan: It's this library, ⁓ YTDLP. It's a download script for YouTube. So you basically put any like ⁓ YouTube video ID in there. And so like,
Alan Smith: ⁓ sick. Or like a channel or whatever, and then you can download it and...
Patrick Keenan: Totally, yeah. And you can choose like you want audio or like the format, how many versions. ⁓ This exists for other stuff too.
Alan Smith: So what are people using it for? Are people using this for learning purposes, or is it more like marketing purposes, like going backwards and trying to reverse engineer channels and figure out topic clusters and stuff like that? What are people doing with it?
Patrick Keenan: That's a good question. I think some people just use it to like, you see a YouTube video and you want to use it in something like a presentation or something. Maybe they grab it, but it's kind of illegal. you know, ⁓ if you.
Alan Smith: Does it? Yeah, we're just, we're stealing all the content from the platform. So apparently that's what we do now.
Patrick Keenan: Yeah, and I think the last one kind of got blocked like this is a fork of the older one ⁓ and so it's similar to like the torrent world where if you're like nerdy enough and want to deal with all the hoops ⁓ You can do this stuff, but I do think in the future this will be much more accessible this kind of like command line utility to avoid using websites and I just mentioned that because like
Alan Smith: Yeah.
Patrick Keenan: Yeah, the AI is clicking around all these UIs, but like, there's only a very small step to be like, I'm doing this 10 times. I'm now going to make this into a CLI instead. And like, why would I run a bot when I can just kind of like programmatically do this? ⁓ And then the other thing is you have web MCP, which is basically voluntarily turning your website into an AI service. And so we'll just have all this stuff colliding so that like,
Alan Smith: Yeah, nice. Yeah.
Patrick Keenan: you design for agents rather than users primarily. And then what does that mean for the user interface? ⁓ Is it voice? Is it text? Does it change? Yeah. So it's very exciting.
Alan Smith: Yeah. There's a lot of incentives for people to do that. There's positive reasons to make your stuff more accessible to agents. And we haven't talked about SEO versus GEO or AI SEO, right? ⁓ This is a huge new ball game. Nobody really knows a lot about it yet, exactly how SEO works within ⁓ agents. But to first, as a first step, make your website more accessible to other agents, If there's more searches showing up in agents, that's where you want to be. So there is a very positive incentive structure for lot of businesses or products and services to show up in that way. And so to make your stuff more ⁓ open, that's cool. Yeah, I like that.
Patrick Keenan: Yeah. I mean, the other links, I'll put them in the notes, but they're really just about OpenClaw ⁓ and then some notes about design and AI. ⁓ I think the last one I'll just say is somebody reverse engineered the Apple neural engine. ⁓ And so I think increasingly our, our compute will be local, you know, and that's like another, it's like basically deflationary for AI tokens. If you can run everything on your local machine.
Alan Smith: Dude. Right, yeah, it's like, yeah, you don't have to, if you already have a computer sitting around that has a bunch of cycles in it that you're not using, that's money in the bank, right? Like, it's a, yeah, and we all have it.
Patrick Keenan: Mm-hmm. Yeah. Yeah. Yeah. Like, will we look back and be like, why were we not using our computers when we weren't at our desk? It's so crazy.
Alan Smith: Yeah, it was like remember SETI search for extraterrestrial intelligence and how it you would install the thing and it was like Doing this huge job of like searching the entire sky for aliens and all these little like code chunks that you would download and you know that I feel like to me that was the first big distributed computing project and The fact that okay is coming back again, you know distribute computer
Patrick Keenan: Yes, yeah, yeah.
Alan Smith: So you're just adding yourself to a botnet of like, you know, this, yeah, okay, cool. Yeah, yeah, it's true, right? It was all a cover story. Yeah, we'll make sure your house stays at 21 and a half degrees, no problem. All right, man, thanks for educating me this week. It was always super fun to talk about this stuff. We'll see you next week. Thanks everyone for learning in public with us.
Patrick Keenan: Hello, I'm Patrick Keenan and I'm excited to dive in to all things tech, AI design, Toronto, et cetera.
Alan Smith: I don't know, that wasn't very convincing, dude. Have you had enough coffee yet today? What's going on?
Patrick Keenan: I've probably had too much coffee. I'm like past peak. But yeah, let's go. Let's go.
Alan Smith: Oh, okay. All right. Let's get fired up. All right. So last time I wanted to talk about a few different things. We talked about product strategy in the AI world. Actually, I have an addendum to that as well. And this is maybe a small thing. And then I also wanted to talk about this. The second thing I want to talk about was the machine is the message and the medium is the message. Marshall McLuhan. And then lastly, I wanted to go through all the links that we didn't have a chance to go through last time. And I looked at a couple of them and I was feeling super freaked out. was like, dude, it's going too fast. I'm freaking out, man. And since then I got a message messages from a bunch of devs that I know who are like, I don't sleep anymore. I just make stuff with cloud code all the time because I can make so much stuff. And so that feels like the energy of the moment. If you're listening to this in a couple of years, that's the vibe that we're living in today.
Patrick Keenan: Yeah, feels like something that's changing every day. the only way you can keep up is if you're unemployed. This is basically where you're
Alan Smith: And actually some of the guys who were said to be stuff were like in between things and just like so stoked to not have a job and to be able to just build their own stuff every day. ⁓ Honestly, not much different than being a developer in the past. Like this just want to deaf, right? Like it's a huge part of it. ⁓ Okay, so first thing I wanted to share was I was thinking about this product idea of, ⁓ okay, so you know when you're like old software, you would have to like log in and then you would click on things and it would kind of like bring you information. was basically like a view on a database, right? We talked about this last time. Old software was for people. New software is people. There's this kind of space in between. Like I was watching, you know, my wife use ChatGPT and I'm stoked for her. She's doing really well. So like, just like finally onto it. I'm seeing what she's doing. I'm like, okay, okay. That's like, that's cool. I never thought of that use case. And she's doing things I wouldn't have expected. And then she closes ChatGPT and it's kind of done. So there's this like, then you think about the open-claw people who are like, yeah, I'm buying as many Mac minis as I can and spinning up like 40,000, you know, open-claw agents to become my army to do all this stuff while I'm sleeping. Like, that's different, right? Like if you and me are working on a team, when we like close a meeting, I'm still working at the company and you're still working at the company, right?
Patrick Keenan: Mm-hmm.
Alan Smith: Whereas with chat GPT, it's like you kind of, or like most of the AI chat tools, there's AI magic there, but it's not an AI agent that's working when you're not there. And I think that's a really big distinction in terms of where people are at today. Like a lot of people just haven't got to using agents that are running long-term beyond just waiting like a couple seconds for a chat here or there, or maybe like a deep research or something on perplexity. What are your thoughts? You noticed this gap too between people?
Patrick Keenan: ⁓ I noticed it in myself. Like I have to move to a place where agents are working when I leave the computer. I have like an open clause set up. My default is to go to a chat box. I think there's kind of like, I think that's one component of it, which is like the chatting piece. And I think what open claw has added is like a couple other components. So like one of them is web hooks or hooks. So like when something happens, I want you to pay attention to it. which is also what somebody does when they're like a human and not just like a computer waiting for you. Like right now it's very reactive. I get a message in my inbox. That's an event. It's a notification for me. I open my, in my inbox, I deal with it. I think the interesting bit about the web hook is like now my agent has an inbox. So that's a new thing and it can respond to that inbox. The other thing is just a heartbeat. So yeah, it's weird to think about like your chat app opening up and
Alan Smith: Mm-hmm. Mm.
Patrick Keenan: saying what's going on right now. But that's how OpenClaw works. It's got like a timer and like depending on how many tokens you want to spend, you can wake it up every half an hour, four times a day. But the notion that it would wake up, be conscious of like the things that are going on, the new threads, like information from the internet, and then choose to act or not based on that ⁓ information at the moment is a new thing. Yeah. And then I think the always on nature of it, which is that Now, when I chat with it, I know these things. So like, I know that now it's going to wake up and I know that it's going to get these web hooks. I'm not there yet. Like I don't, that's not how I use my open claw yet. I'm still like at the point where I'm like, click on a button on the screen and send me a screenshot, but I can see how it fits. Like I still do the chat thing. Still going to have the chat, GBT open up a session, ask a question. ⁓ But to be like, and we were talking about this before, but to have like a theory of mind for how my agent is working, that's going to be learning thing for me, for sure.
Alan Smith: Same here, I've seen other people further ahead on this. And I think it's partially because it's mostly with dev tasks, where you can get this longer term ⁓ runtime, of uninterrupted, hey, initially it was an agent could do something for 10 seconds ⁓ before the task quality would start to break down. And then it was, it's 30 and then it was a minute and then it was five minutes and then it was 10 minutes and now it's like four hours, right? And you project out the curve and it'll be like doing human work weeks unassisted, like go do this really hard task and it's essentially gonna like one shot it or you're gonna be able to, you know, give it some edits and it'll clean it up in a couple hours, you know, at the end of the week, right? And so think about that in terms of as a manager being able to show up and, you know, check in on someone's work. once a week, like that's a pretty high value or like high agency team member, you know, if you don't have to check in on them too often and you can give them like a week's worth of work or like a month's worth of work or a year worth of work, like that, think about it. That's crazy. ⁓ Especially if they can just be getting information from other areas that are kind of like course redirecting them as time goes on. It's like that old, what's that thing about like redirection where If you're flying from New York to LA and you only look at the dashboard like three times, it's going to take you, like you have to do these massive course corrections. Whereas if you're constantly doing like little updates, ⁓ you almost go in a straight line to get there. It feels like there's some metaphor there with like what we're doing today.
Patrick Keenan: Yeah, Drift, guess. Yeah, interesting. Well, I mean, we, yeah, we have that. And so I think that like, you know, chat, GPT 3.5 comes out and everybody's like, wow, this is cool. But like, ⁓ it's hallucinating. Right. And that was like a big thing. Like, ⁓ some of the answers it gives you sound true, but are not. And so like, then we had to adjust and like double check and add in sources and do a bunch of stuff to fix that. I think that same level of Drift is going to happen with these always on models.
Alan Smith: Yeah, Drift is the idea.
Patrick Keenan: But it's going to happen like a week and like you come back in a week and it's like, yeah, I made the dating site that you wanted. You're like, what the hell? was working on a SaaS like project. So there's just going to be like an incredible amount of drift without that like human kind of course correcting and we'll have to like debug a bunch. I haven't like, haven't seen that workflow in action yet. I've seen really good stuff that it works for maybe like 10 minutes and it comes out with a good result. But again, it's done a bunch of stuff where I'm like, you're to have to rework this. So it's something we have to kind of play with, feel the edges, like it will be different types of planning around it. But I think one thing that will be a challenge is if you work for a week unsupervised, that's like a lot of tokens. And there is a point where you're like, I am spending as much tokens as it would be to hire a software engineer. So I kind of wonder like, What is that kind of trade off of like how much human and how much machine should we have involved in any given task? But again, we'll play with it and find it.
Alan Smith: Yeah, I wonder in like the price of a token, is it going to go down? Is the token usage going to improve? Probably, you know, where does that net out? I think we can, I'm going to assume it nets out on the side of like, it's cheaper still to use AI to do most code and having a developer managing a bunch of like kind of like junior AIs. And the real question is how cheap does a junior AI get compared to, you know, a more senior dev who becomes the manager of this team, right? ⁓ So I've been thinking about this idea too, in terms of just like building products and trying to think through like workflow of, okay, if what we're primarily doing, like right now there's a moment and this kind of like loops into, my sort of like next big idea. Let's talk about Marshall McLuhan, right? And you know, the idea that, you he shared with us, we talked about this a little bit last time. But I think we're at this moment with AI where it's like when TV came out and people just put plays on TV, right? Like it was not a moment where we knew what TV was yet. TV became something different. And then the internet showed up and we were putting brochures online, ⁓ phone number, addresses, photos of shops, owners, cat pictures, right? Like it was brochure aware. That was like a whisper of what the internet actually became. And so this idea that the medium is the message, it takes a while using the medium to figure out what the message of that medium is. And I think we're still there, man. We're in the really early days. And so the first thing that we're doing, like our TV, or sorry, our plays on TV, or our brochures on the internet, before we understood what those things are today, I think are just like, taking all the human jobs and workflow and automating all those workflows. Like that's what we're doing. And that's what we think AI is for today. And I think that that's a world changing starting point. Like there is a lot of juice to squeeze in that on its own. Don't get me wrong. It solves some massive problems. It creates some real value. It may create some massive problems and it may destroy some real value as well. But I think that we're at that moment and it's still only a sliver.
Patrick Keenan: Which is to say, which is to say like when somebody put like a theater show on the TV, that was awesome. Like people at that moment were like, this is nuts. I'm not at the shit. Like this is so much cheaper than going to theater and this is like in London and I can see it. It was just like mind blowing and so exciting. And so clearly like, wow, this is the power of this medium. Little did they know.
Alan Smith: you Yes, they were getting just like a hint of what it was actually, right? And if that pattern is true for like what we're feeling already today, it's like, dude, like we're just gonna be like, like just like, it's gonna be lot of right. So I've been thinking about this and okay, so what is it right about where we're going next? What is this medium of AI really wanna become? And I don't think we can get there from a whiteboard. ⁓ I don't think we can get there from a podcast kind of like pontificating on what it's going to be. We're not going to predict it from a distance. We can only discover it by building, by being in the products, by trying to make stuff and letting it reveal itself to us kind of like piece by piece. And if McLuhan said the medium is the message, then I think you know, the motion and the creation reveals the message over time. You know, this is the idea that clarity follows action and direction comes from movement. So I think that, you know, our job as we build these things isn't to forecast like kind of like the final form of what AI is going to be with precision. It's to just jump into the meaning, jump into the medium, work with the best idea we have today, which is essentially taking human labor and trying to automate it, which again, like may create some serious problems. But yeah, basically like, you know, where we begin isn't going to be where we end up on this, but we just need to begin. Like we need to start, we need to be working on stuff. And I think one iteration at a time, one customer conversation at a time, one signal at a time, that's how we get there. We can't see around the corner. We just got to walk around it. You know what I'm saying?
Patrick Keenan: Yes, that this stuff is discovered through building. And I think you can see that at the model companies, right? So like the model companies ⁓ are all scrambling for use cases. And like you can see Anthropic doubling down like on the coding thing. And that's how they got where they were, which was the right intuition, right? And now all the other model companies are like, coding unlocks all these other possibilities. We shouldn't have started with like X, Y or Z. ⁓ We got to catch up with Anthropic. ⁓ So
Alan Smith: Mm-hmm.
Patrick Keenan: Yeah, the question is like, where do you start? ⁓ And that's from the model perspective. But then on top of the model, you have applications, right? And you're building out some kind of a wrapper, I guess you could say, or front end for this model. And the decisions you make there are going to like, yeah, there's like a path dependency where you're going to make some decisions that are going to fit you into a corner. And so I think every, every company out there that is thinking about AI is like, ⁓ we can be the wrapper. for this thing, like we'll be like the front end. What, like if you were starting from nothing, because like if you start with an existing business, you're like, obviously I'm going make it work for, you know, I'll make it work for construction workers. I'm working on like city services or I'll make it work for her salesmen. But if you were starting from nothing and you were saying like, how can I explore this new medium in a way that I will more likely reveal the inherent
Alan Smith: Yeah.
Patrick Keenan: benefits of it, like in the same way, like with a TV, you do closeups. Like that was a really big deal. It's like, ⁓ you can do a closeup. That's crazy. We never thought about that before. like jump cuts or like fades or like establishing shots. Like all this kind of language had to be established. I think that's what me and you get excited about is like, what are the new patterns and pieces of the language that need to be established to like really push this to the next level? ⁓ Where, what use case would you start with if you were, if you wanted to explore that stuff? Or how would you start on that?
Alan Smith: I don't know if it comes so directly, like the connection between the end of where we're going to end up and what we're doing in the short term. In the short term, in terms of where to start, I think that's a really interesting question for every small business, medium business, large business to do is to say, okay, what can we automate here, right? ⁓ One approach to that, and there's different strategies, so different people have different AI strategies, right? One approach is you say, okay, well, what's the thing that we spend the most money on today and ⁓ is the most automatable? So you start to say, okay, well, do you need a body to do this? If it's keyboard, video, and mouse, then you can just do it. ⁓ And we have a lot of people doing this, so it's going to be expensive. Let's say you have a call center. Your number one expense is the people sitting there answering calls. So you start to say, okay, well, can we build an agent that can answer the calls? Right? No brainer. Let's say that you're ⁓ a lawyer and a huge amount of your staff is juniors who are paid to stay up all night reviewing case docs. It's like ⁓ to prepare something for you for the next day. It's like, well, it's not really efficient. Are we getting the best value here? This is a good use case for AI. So you start to put it in there. And I think technically there's generally two places to add it. Either you go front of house and kind of make it more, this is the interface with the customer. This is, ⁓ or back of house, like bookkeeping and stuff. Or delivery, right? So you've got kind of like sales and marketing, ops, delivery. And if you break up those like few pieces, delivery is always like, what are you as a business? Are you an agency? Are you a ⁓ construction company? What is the thing that you do and that you sell as a core? And then there's all these enabling functions around that essentially. So in my experience, going towards the core works when the core is really expensive to deliver and highly automatable keyboard, video mouse, Are the core ways in which you actually deliver that. Otherwise you say, okay, well, let's go. Where's the bottleneck in the system? So that's like strategy number two. First one. What's expensive? Let's do that. What do we spend a lot of time doing and what's expensive? Second one, where's the bottleneck in the system? Are we demand constrained or are we supply constrained? So that's option number two. Option number three is just like, well, what is it that we're not doing today that we wish we could be doing? Let's add that to the business. Let's let the rest of the business run the way it should, but let's do the thing that we're not doing. I'll give you an example. We have this AI that we're building for plumbers and like all different home services businesses. All of them have a task, which is after the job is done and the technician leaves your house, they want to call up and say, hey, how did they do? You know, did they take off their shoes? Did everything feel like it was fixed? Did it feel like it was friendly? Did they leave the place clean? Would you leave us a Google review? You know, could I ask you to give us a testimonial here on the phone? I'll type it in for you, et cetera. ⁓ That's called a happy call. Nobody does happy calls because they're just too busy. They can never get around to it. So now you've got this use case of, okay, we could do that now. It's a good part of the business. The rest of the business is kind of running okay. Like we're not like dying in the rest of it. Let's just add this and then lift the whole value of the business in some way overall. So those are the three ways that I know in terms of deciding what to do next. What's expensive? Where's the bottleneck? And then just what aren't we doing that we would like to be doing that we might be able to automate and add in?
Patrick Keenan: I guess with all three of those, you're kind of like looking at an existing industry where maybe those things are like more quantifiable, right? So like you can look at a business, you can look at his P and L and you can be like, this is the costs. Um, that makes sense. I think, you know, I guess going back to the metaphor of TV, um, I'm curious now, like was there a director who was like, I mean, there was, what's the Rosebud movie, Citizen Kane? Like that was like, yeah, that was a really big breakthrough in terms of filmmaking. And he chose to do it about like, kind of like this, you know, this historical robber baron kind of guy, Hurst, I guess. And so it was a kind of a...
Alan Smith: ⁓ Susan K. and Orson Welles. Mm-hmm.
Patrick Keenan: a dramatic movie. You know, that's where they pushed the medium forward. I guess with brochureware, it's like what were the first companies to adopt like the internet? ⁓ don't know what were they, just nerds. I don't know.
Alan Smith: It was nerds, man. It was nerds. But then everybody needed a website all of a sudden. It was like you had to get your like GeoCities website and every small business like just like local businesses would be like setting up a website and then they would add like a www to their business card. Like remember that? Like that time when that was like a thing where it's like, hey, we have a website. What's on your website? ⁓ nothing really. It's our hours. Yeah, pretty much. There's like...
Patrick Keenan: Mm-hmm. The same stuff as the business card. Yeah.
Alan Smith: a carousel of pictures you can scroll through it. Whoa, cool.
Patrick Keenan: But then, then people started building web applications, which is like the Amazon or the Facebook or whatever, Google.
Alan Smith: Exactly. And we just didn't know that that's what it was going to be. The websites were going to be more, or the internet was going to be more about applications than it was about static messages. Right? And I think the same thing where I think we're looking at AI a lot now in terms of its intelligence, and let's take that intelligence and just apply it to human workflows. It's like, okay, well, that's one thing we could do, but there might be something, there will be something different, something much deeper than that.
Patrick Keenan: Mm.
Alan Smith: as opposed to just adding a bunch of synthetic humans to the workforce. Like, don't get me wrong, we're doing that, that is happening. There's an industry, there are tens of thousands of companies being born on that and hundreds of thousands of jobs are gonna be lost on that one idea. That's not where we're gonna end up though. That's not the end of the story.
Patrick Keenan: Right. Right. And I think that's what we're seeking out on this podcast. So if you have any ideas, please let us know and hit us up for what you think the future metaphor will be and what what AI is truly meant to do.
Alan Smith: Take a guess, take a wild guess. ⁓ I'm not gonna take a guess here today. Let's talk about some of the stuff that's going on in the world more specifically. What are people thinking about? What are people writing about? What are you seeing? Why don't you share some links with us?
Patrick Keenan: Yes. Why don't I share some links and these will all be ⁓ in the notes. So I'll share this window here. Okay, so let's start here. ⁓ this was a kind of tweet that I thought was really good. And I've been thinking about this since, I don't know, ⁓ two days ago. So my biggest current workflow. for making iOS apps is extracting out all my core functionality into a standalone Swift package, which is to say into just functions, no UI. Making it into a CLI, which is the command line interface, giving Cloud access to that CLI and going to town. If you do any kind of on-device ML or have lots of headless services, this is extremely helpful and a better way than just waiting for rebuilds. So this is like a very ⁓ technical way to say like, just want to test it with a CLI, but I think there's something else here, which is maybe the CLI is the future. So what we're seeing is like, I guess a couple of CLIs is the big one that, you know, is the breakthrough is Claude. Now, like everybody who's never touched a terminal is using ⁓ Claude CLI, which is you just type in Claude and then there's a bunch of magic words you can say to Claude via the CLI, right? So
Alan Smith: Mm-hmm.
Patrick Keenan: That's all it feels like 1970s computing, right? But what's interesting about it is that in a world where ⁓ computers ultimately run on a terminal and the fastest way to get something, whether that's a URL or like run a process, like convert this set of images to a video or like basically anything you do on the computer, the fastest way to do that is in a terminal. AI knows all the commands and then this is perhaps the base on where your UI is derived from. So like maybe your UI is just like built on the fly based on the CLI that you're using. So that was like one ⁓ thought and it kind of brings me back to what we were just talking about, which is like how do you kind of design for this stuff or like what is it meant to be used for? Like maybe this whole internet UI ⁓ path was just like a side quest. Like maybe all the UIs that we're seeing on the internet are just kind of like. A silly side quest will look back on and be like, everybody made their own UI. This is weird. Like look at all these old documents and people put so much effort into these like business cards and stuff. Why didn't they just use, you know, NFC? So I think it'll be like that. That's one world where, where you don't make custom UI. Instead, it's like built for you on the fly using these command line interfaces. And as a developer or dev or designer,
Alan Smith: ⁓
Patrick Keenan: you're actually just defining what functions you want to provide to people. Like I want to provide the ability to respond to appointments and book appointments. I'd like to get these fields and then when they're done, I want to send them over here to this human to check that the appointment is on the right calendar or something. ⁓ So yeah, that was one thought.
Alan Smith: So I love the term headless. Let's define this in this context a little bit. ⁓ You have lots of headless services on device ML. What's the context here for those use cases? Or just help me understand what Tyler means when he talks about that. Because I'm very interested in especially headless services.
Patrick Keenan: Yeah, I the service. It means a service without UI. it means so like you can run Chrome, for example, ⁓ in the background and you can have it like click on things, but it's never actually rendering to a screen. So headless means like I'm running this service and I just need the output from the service. So yeah, it would like in this example on your phone, you might take a screen like a photo. And it's like, find if there's like a hot dog in here. And so now it goes to this headless service. That's going to like take the image, do some edge detecting, like put that into an algorithm, figure out if it's a hot dog or not, and then spit it back. Right. And so like that's the headless service is like the hot dog or not hot dog. ⁓ and there's all kinds of stuff like that on your phone. The classic challenge with engineering is like the UI always gets mixed up with these like backend services, but
Alan Smith: Mmm.
Patrick Keenan: And so what has happened is like, we've kind of like gone to a UI first development approach, which is like, show me the flows, like make up the mocks. Like I want to see where the buttons are. How does the login screen work? What if I need to reset my password? We design all those things and then we implement it. But perhaps the future is just like, you actually just define what your CLI is going to do. And then you tell the AI to make a UI.
Alan Smith: Yeah. On the fly, yeah. So actually this is really, really interesting. Okay, dude, I'm feeling this. This is pretty exciting. So something I've been complaining about a lot recently is internally. So we're building 10 products at a time, right? Like every week, I get to look at 10 products and how they're kind of progressing. And the one thing I'm noticing across all of them is they feel like old software to some extent, right? Like we all know everybody on the team is well aware of where we want to go. We want to move towards this, this magic, you know, want type scenario where, you know, it's the team member that's working with you and you know, your team member doesn't necessarily have a dashboard or like, you know, a sidebar menu, right? Like that's just, those are parts of old software. Like those are old, like artifacts that don't necessarily apply to what we're trying to do in the future. and to try to show some of these ideas to a customer without building the whole thing and actually making all the functionality and the AI work, which is still like a non-zero cost, right? Like for some of the stuff we're trying to do, like we would need access to data and then we'd have to like run it a bunch of times. Like it wouldn't just work. like, these use cases aren't one shotable, quote unquote, yet. ⁓ So what I've noticed is that it feels like a lot of the demos that we're giving feel like old software. They don't feel like this future approach. a lot of like, if you use Replet, it pushes out kind of these old interfaces. And it was when you use Cloud Code, it's trained on that. So it's giving you these interfaces. Yet I feel like we're being pulled towards something new. And every time something comes out, I'm like, Do we need this screen? Can we strip this away somehow? Can we just get back to what's the thing they're asking for and how do we get it to them quicker? How could there be less ⁓ visible here? How could this feel more like it's a shop and you parked your car in it and then when you got your car it was fixed? You weren't following an interface that gave you a list of everything that was happening while it was happening and trying to direct the mechanic on what they were doing, right? Like this is the people want the outcome like you said. And so sorry to try to bring my point down. I'm kind of rambling a little bit. It feels like none of us are quite sure from an interface perspective how to come up with software ideas and demo them and validate them in a really cheap way without actually building the whole software yet in this kind of new style. And also what does the new style look like? Like there's kind of like two or three moving targets here in.
Patrick Keenan: in
Alan Smith: product development for me today.
Patrick Keenan: Yeah. And I think that one, ⁓ like misstep for me is like, maybe the future is not more UI, right? Like maybe it's not like a cooler UI. Like I love the Apple glass thing, but like, maybe that's not the future. that I'm maybe that's part of the future, but like, there's just this other world I I'm thinking about with this CLI conversation is like, I'm seeing people use command line interfaces and I'm thinking. this should be an auto-complete. I shouldn't be typing it, I should be talking it. And I should know what the inputs could be because somebody's telling me. And those fields should just manifest, they should just appear. And so I think that there's two business models that interfere with the step change, which means that it's an opportunity for people starting out. So one is, if you're not looking at it, I can't show you ads.
Alan Smith: Yeah.
Patrick Keenan: Right? Like most of the internet is funded by ads. But if you're using a command line interface, it's very hard to put an ad into that thing because I do a command and I expect an output. And like maybe there's an ad appended to the video that I processed or something, but it's much harder than just like showing you something because our whole internet is designed to like, to show ads before stuff. So that's one piece. The other thing is if you're using a CLI,
Alan Smith: Yeah.
Patrick Keenan: I can't really charge you. Like it's harder to do charity. You can do authenticated sessions and stuff, but ⁓ it's kind of harder to charge people for something that's a command line because it's very easy for me to remake your product. So ⁓ I have to look this up. So Figma ⁓ made in a weekend. There was this dude, this is a funny story. ⁓
Alan Smith: No, no, no, no chance.
Patrick Keenan: I don't know if I can find it. man, blocked. There was this dude who ⁓ had a plugin for Figma and it would take your Figma code and it would like build it into React and Tailwind and stuff. ⁓ But then Figma blocked him. ⁓ And then he was like, cool, you blocked my plugin, I'll just remake Figma. So he, in a weekend, he just like...
Alan Smith: ⁓ my god.
Patrick Keenan: Remade figma. I'll find the link and Yeah, this he built his own the figma in three days. So I haven't used this ⁓ But ⁓ it's open pencil. Okay, cool. So this relates to the other topics I want to talk about but yeah, this is the response So that's what I mean about the CLI is like, yeah cool. You're selling a CLI I Can read first engineer that in a weekend because there's no UI in it. It's not that complex, right? It's like text in text out or binary in and binary out. So those are like the two business models that are, think, preventing us from thinking about what the future might be because we need people to use these UIs, right? We need people to like log into these UIs.
Alan Smith: Mm. First, that's wild. Like the fact that the dude rebuilt Figma and you I remember a couple years ago, us talking about Figma and being like, this is such an engineering masterpiece that could only be done by a company that was founded by like, you know, a 0.01 % engineer. Like that's what this is. And the fact that now, that engineer is available to all of us through clogged code, essentially, like he's on all of our team. Like that is no longer an advantage to be a 0.001 % engineer. And like when they did that, like Figma, that was like... the entire Adobe Creative Suite or at least a good portion of it that was relevant to essentially like UI and web design first, I guess, right? Like the fact that they could take that and put that into a browser, I just remember being flabbergasted and that that was like a thing and how much money they must have spent to be able to get that done. And the fact that one guy could do that in a weekend. Dude, what timeline is this?
Patrick Keenan: mean, and it's really fast. Like, and are there collaborators? I don't even know. ⁓ It's so nuts. But I just love the checks, because I love the checks and balances of technology, which is if it sucks, and you're being a gatekeeper, I will just remake it. Like, I just love that check on things. And so this is like a check.
Alan Smith: It's like pure market forces in a way, right? Like, you know, lot of the, we talk about this in terms of moats a lot and like, you know, how can you build some sort of moat and how can you build some sort of competitive advantage? And I think that you shared a link or, you know, I saw one recently, which was around, you know, there used to be two product strategies. One was first, be first. The second one was be better. And it's like, there's no point in being first anymore because anybody could copy you at any given time.
Patrick Keenan: Yeah.
Alan Smith: And so you just need to be really, really good. Like that's the only product strategy moving forward is you have to understand what people want and then you have to give it to them as well as possible. Now I'm going to push back on that a little bit. I think that that's a little bit myopic and that comes from a product person. If you're thinking about the business overall, you have a lot more levers to pull in terms of how to be a better business. You have a brand, have customer service, you have distribution, right? Like there's all these components that product people sometimes forget about. I think that he's probably right just in terms of the product strategy piece of your product needs to be better. But as far as business strategy, it's maybe a little myopic.
Patrick Keenan: mean, so notably he made Figma. I think he was like the main designer that made Figma. ⁓ And I think, yeah, it is a product mindset. Well, would you say that brand and marketing are part of product or is that just like a product-centric way to view the world?
Alan Smith: It's definitely product-centric way to view it, but everybody stands where they stand, right? So, know, accounting is gonna say, well, product and brand are part of accounting. So, it's like, we add up the numbers without you guys, or without us, there's no business. It's like, oh, well, it's kind of true, right? So, everybody stands where they stand. are points of view. I think that brand is what people think about you when you're not in the room, and do they think that you're...
Patrick Keenan: Yeah. Mm-hmm.
Alan Smith: really interesting and great and they have loyalty to you or do they think that you're completely replaceable and they're ready to you know switch to anyone else at the drop of a hat you know there's there's aspects to that on its own the product can of course deliver something on that brand like it can be a part of the brand building component or the brand is built in part by you know the product if you want to put that first at the end of the day it doesn't matter they're They're very intertwined. They're part of the same thing.
Patrick Keenan: Mm-hmm. Yeah, but it's, it is a timely piece of advice. There, there used to be two playbots for commercial software. ⁓ so that's good. He frames it as commercial software. Be first to market or make the best product. Being the first was important yet so many software companies operate this way. We must ship by this time or next month or we'll lose. ⁓ is this shallow way to build now with the AI tools we've gone from LOL, nice JavaScript try again, ⁓ to damn it's nice C program. take the wheel. So I think he, yeah, he's basically saying the, the ability to go to the mark first is like, now everybody can do that. So like, I think the point here is that there was two strategies. Now there's kind of one because going to first to market is not going to put you ahead of your competitors because everybody can be first to market. So like, what are you going to spend time on? Yeah. What do think about that?
Alan Smith: Again, in the product centric perspective, I think he's right. I think you do need to have a really, really great product. The bar is always going to be high, and people are going to try to work around whatever else you have in terms of a moat. So there's this war for context that we talked about last time. How do you add more context to a product? ⁓ You have a lot of, if you're in a position where you can aggregate contacts. So let's say that I'm running a business that is front desk and I have access to all of the docs and all of the systems and I have unique access to just all of the phone calls, right? I have contacts that maybe some of the other systems don't have. And so I hold onto that, which allows me to be more useful and more like put my intelligence to more work, right? is going from you took the smartest person in the room and you put them in a black box and you ask them to solve a problem versus you help them understand complete situational awareness and what's going on in a business, right? So, sorry, I don't mean to get off track here. What I'm trying to say is that context is so important to being able to build a good product. And oftentimes that comes from how you position what you're going to be doing, that there are certain tasks that you can do. that if you do those, maybe in addition to the thing you were trying to do originally, you can collect more context. Therefore, you can make the original thing you were doing even better and build up a bit of competitive advantage there. And if you're doing that and someone else can't do it, then they can't do some of the other things that you could start to do because you hold that context, right? So it gives you a competitive advantage on the product development itself when you have that context. So I think it... context is really, really important here. It's not just about developing stuff. It's about where does it fit into the business and what do you have access to to be able to add value to these businesses. And I'm generally thinking business to business here, but I think the same concept probably applies with B2C products where if you're the AI pin that theoretically OpenAI and Johnnie I've are developing that just records everything I'm doing, and has ultimate context into everything in my life. Like it's heard and seen everything I've heard and seen. Like how much more context do you want? Right? Like you're going to be able to build whatever you want from having that.
Patrick Keenan: Yeah. I think that's right. ⁓ Increasingly, however, I think that's about now that I have options, who do I want to give that context to? So with that said, this is a link which is switched to Claude without starting over. now, like what we thought was the the moat ⁓ is now a commodity. So I can import from Gemini, I can import from Chachbt, I can import from, you know,
Alan Smith: Oooooh!
Patrick Keenan: any other service and now all that context comes with me, right? Which is a good move for consumers. ⁓ But ⁓ to your point about like the special context you have, you're going to have to think about like, is it context that's not the users? So like you could imagine I have an inventory of a bunch of electronical components and like that's my context is that I have this industry knowledge that you can't just export. But increasingly, I think
Alan Smith: Nice.
Patrick Keenan: anything AI can touch will be importable. I think it's just an interesting move is like the portability between these chatbots.
Alan Smith: We see so much, especially in, there's a lot of existing systems, like especially systems of record. Like if you pick any industry, right, like pick dental, you know, there's a bunch of systems of record there and they're all very protective of their data because they know as soon as they start letting this stuff flow everywhere, they're cropped, right? Like somebody's going to take it all out and stop using their software because it sucks. Like, because that's what I would do, right? Like I don't want to use this crap. It's like, let's build. a better system of record and they're like, nah, we're just going to kind of hold on to everything. Or we're not going to play nice. We're not going to allow this data out and give you access to it. Epic, all these like healthcare, ⁓ are they? ⁓ EMRs or EHRs, like ⁓ electronic medical records or electronic health records. These systems, and then look at anywhere else like SAP, there's always these systems of record. that suck, they're so old, they're legacy, and they've built these walled gardens around the data, right? Now we're seeing kind of like a system of intelligence being built on top. And that's where the opportunity where a lot of people are jumping in of being like, well, I'm not gonna be able to get this data out, but I can via API work with the data to accomplish a lot of things. And this idea of the system of intelligence is one that we're looking at a lot of, okay, well, that's a door until... all this data starts to flow everywhere, or you can rebuild it just by being the system of intelligence on top for long enough, right? Where you've just, you've seen it all at some point in time.
Patrick Keenan: Yeah. Or you could just run this prompt. Like I think it's funny that this whole product is actually just this prompt you put in here that says, I'm moving to another service. I need to export all my data, list all the memory stored about me, add dates to it. So like this is like the exploit ⁓ is just put that into your thing and then paste it into cloud. So these like the, again, the barrier, the technical barrier to do things is so much lower, which means you have to work so much harder. to have people want their context in your system or want to use your system ⁓ on a daily. ⁓ So with that, mean, I know we've got a lot of links and low on time, but I want to like run through some options of ⁓ Figma alternatives. So we looked at Open Pencil, which was made in a weekend. ⁓ There is also this one called pencil.dev, which is ⁓ a live canvas that like
Alan Smith: Mm.
Patrick Keenan: As you're building, it's also making the code. This is this, this by the way, is like my dream. Like Framer, Figma, I always hoped that they would do this, but they haven't done it because everybody's got to focus. But this is like my dream is like the thing I'm designing is code.
Alan Smith: Yeah. Dude, very first job out of school, I remember at blast radius, I was making mock-ups in Photoshop and I was like, this is for editing photos. Why can't I just make the website like here, right? Like this is so silly. And then it had to go to these HTML, like I had to slice up all these images and I had to send it to the HTML guys. And the HTML guys had to write all this code and then they had to like go check it in the browser and it was such a mess. And I'm glad that all that workflow has collapsed now too.
Patrick Keenan: Hahaha
Alan Smith: I have a few different ideas and what I love about pencil is it gives you essentially design alternatives at the same time. I think that there's a gap here for me in terms of brand of saying, well, like, look, you're just giving me a bunch of different ideas maybe for new projects where you haven't defined that stuff, it's good. ⁓ But the idea of, it's great, it's great. I love it, it's great. I'm not going to critique it. ⁓ This is a dream come true, we're there.
Patrick Keenan: Yeah. To like, ⁓ full disclosure, I haven't tried this yet. I mean, I've opened it up and pushed stuff around, but I haven't actually tried it for anything. ⁓ But again, it's like easy import port, like design as code. This is the future. So here's another one. I don't know if you've heard of this one, paper.design. So it's a similar thing.
Alan Smith: I like one is called pencil, the other one is called paper.
Patrick Keenan: And the other one's called open pencil. it's like, yeah, we're stuck in a metaphor. ⁓ So yeah, it's a similar thing. It's just kind of like, it's you, got your, you got your code buddy ⁓ at your side. So it's this one, I think is much more like open source ⁓ and like integrated with the AI terminal. But again, just like a different answer to how can I see the thing I built and then build the thing and like have that go back into AI. And then I just go back to our point about the CLI, like is the future move, like it's becoming so easy to make these boxes and arrows. Like, should we even be doing that? Like I know that there is like a, an attention to the detail there, but AI also has like a good attention to detail. Like will we be making, will we be pushing boxes around? I don't know. Or is this like the temporary moving around ⁓ chairs on the Titanic? I don't know.
Alan Smith: I also don't know. I struggled between going between, all we need is a command line interface and all we need is outputs. I don't know if I'm sold on the idea that the easiest way to interact with something is the command line interface. It's like saying, you know, for books, all we needed was text end to end on every page. Like that is one specific interface and it doesn't necessarily do everything we need it to do. I'm, without getting into it too far, I'm cautiously optimistic that we'll get more on the fly interfaces that do what we want when we want it and don't need it if we, if we can exert our preference here ⁓ as we move forward is what I hope.
Patrick Keenan: Yep. Yep. And it's the, we'll only find out by, ⁓ by playing with stuff. How much time we got left here?
Alan Smith: We're at 48 minutes, let's do 10. We got 10 left, so you got eight links, a minute each or something.
Patrick Keenan: Okay. Yeah, so we spoke about ⁓ design tools, talking about CLIs, talked about like future craft and software. This is another ⁓ chain in our Darwinian evolution of AI, which is like maybe it's hardware. So this ⁓ is looking at ⁓ an assistant that runs OpenClaw on an ESP, basically an open, or no, this is a Pi Zero. So like pretty cheap piece of hardware.
Alan Smith: Yeah.
Patrick Keenan: and you just talk to it. like maybe, and it gives you some visual feedback, but maybe the future is actually much cheaper and smaller than we think it is. Like maybe we won't be using these computers. Maybe it is like just like a really cheap commodified piece of hardware that connects me with the agent I want to be talking to. So that's one thing.
Alan Smith: Mmm. I think that that is, this is shades of what we can expect from Johnny Ive, OpenAI, always on assistant type thing. I think that this is like. Somebody with way less money, way less time, just giving it a shot. Like, hey, I've kind of got this idea. And then the idea is also being channeled through the ether, through that massive team with like $70 trillion to go build it. And it's just like, these are the two outcomes from the same idea that's just coming from the ether of the world through two different conduits at the same time, like how electricity was invented by. multiple people at the same time, like Edison and Tesla, right? I think relativity was even potentially found almost at the same time by another guy as Einstein. Evolution was discovered twice, right? So I think to me, this feels like a second idea, or sorry, this is one idea, and we're gonna see this expressed multiple times in the next X number of years, and someone's gonna get credit.
Patrick Keenan: Hmm. Yeah. And part of it is compounding factors, which is like the more AI you have, the more you can use AI to design stuff cheaper, faster. And the more you do that, the more AI you have. So I think there's that, but this is kind of like the hardware angle on things. This is somebody who used RayBan Meta and OpenClaw to like look around the world and route their vision into their OpenClaw. So, you know, we're seeing, so that's that's a hardware piece. And then this is open claw, but.
Alan Smith: So sorry, is it just building a ton of context then? Like everything that this person sees is like recording in some way? Like what's going on? What's the use case?
Patrick Keenan: So it's just kind of like having this AI companion that can see what's in your glasses. So there's an SDK you can use and you can see, you can take a photo. And so, yeah, it's like just taking photos and hands-free kind of way to do it, as well as there's a screen there. And yeah, you can kind of chat with it on WhatsApp on your glasses and it can like look at what you're doing. So that's kind of, that is like the wearable version of this thing.
Alan Smith: Okay. All right. Rad.
Patrick Keenan: And then this is controlling ⁓ old phones with open clause. So this guy ⁓ is able to turn the flashlight on or just do different things on the phone, turn the volume out, vibrate. And it's got root level access to these phones. You can see it's navigating itself. this ⁓ is kind of a pattern you would use as an automated test suite for many phones. You could see it in a lab where you're just testing this. app on a bunch of stuff. think in some of the comments there might be some of that. But like this kind of thing where you're like, yeah, maybe it's for testing. Maybe it's for like just bot farms. But ⁓ the interesting bit is like maybe this becomes kind of a consumer approach, which is like, I just want to control all my devices without touching them. ⁓ And you can just control your phone from afar. Yeah. So so that's like these are alternate
Alan Smith: Yeah. awesome.
Patrick Keenan: futures, you know, like, what does it look like? Will we be controlling our phones with tappity taps?
Alan Smith: Well, the second big idea in here is like computers learning how to use existing interfaces, right? So it's showing us what's happening on the interface, but is it working? Like if it opens up an app, it's using the app, not like the CLI and like function level under the UI, right? Like it's working through the UI. Okay, so that's cool too. And like, this is, think something that people forget about is, you know,
Patrick Keenan: Correct. Yeah. Yeah. Yeah.
Alan Smith: as devs do we need to build something that connects to every API and only works in that route? It's like, no, no, no, there's another route, which is all the interfaces we built already. Just teach the computer how to use those and click through them all. ⁓ That's a track. There's wild bot farming stuff going on, especially the top of funnel MarTech stack right now is so messed up. Dude, it's LinkedIn. Like X all these things are like turning into digital wastelands because of the amount of bots that are on there in the incentive structure to be able to just try to grab attention and then funnel a small portion of that attention towards whatever it is that you're selling that's related to what that person is so it to me it's like I Don't know where that's gonna end up like what the end game is there like it's a constant battle like I've heard about these guys who rent hundreds of thousands of computers to build essentially or to sort of like click on a bunch of LinkedIn things to just be able to extract all the data out of LinkedIn and then build lists of people to essentially market to, right? ⁓ And that's LinkedIn's data. And they're just like using hundreds of thousands of bots and computers like giant botnets to just get it out. Right. Like we talked about extracting data. It's like, well, we're just going to steal it. ⁓ Nobody's.
Patrick Keenan: and Yeah.
Alan Smith: We're going to make it so they don't really know and we're going to make these things really good at clicking on LinkedIn the way a human would and just extract all this info. ⁓ So anyway, I think that's hilarious. And computers learning how to use interfaces that were built for humans is like a funny moment. It's funny.
Patrick Keenan: Mm-hmm. Mm-hmm. Yeah, it is like a side journey. It's like, okay, cool. Now we can go back to using the command line. I think the other piece is like, I don't know if you've heard of YouTube DLP.
Alan Smith: Nah, what's that?
Patrick Keenan: It's this library, ⁓ YTDLP. It's a download script for YouTube. So you basically put any like ⁓ YouTube video ID in there. And so like,
Alan Smith: ⁓ sick. Or like a channel or whatever, and then you can download it and...
Patrick Keenan: Totally, yeah. And you can choose like you want audio or like the format, how many versions. ⁓ This exists for other stuff too.
Alan Smith: So what are people using it for? Are people using this for learning purposes, or is it more like marketing purposes, like going backwards and trying to reverse engineer channels and figure out topic clusters and stuff like that? What are people doing with it?
Patrick Keenan: That's a good question. I think some people just use it to like, you see a YouTube video and you want to use it in something like a presentation or something. Maybe they grab it, but it's kind of illegal. you know, ⁓ if you.
Alan Smith: Does it? Yeah, we're just, we're stealing all the content from the platform. So apparently that's what we do now.
Patrick Keenan: Yeah, and I think the last one kind of got blocked like this is a fork of the older one ⁓ and so it's similar to like the torrent world where if you're like nerdy enough and want to deal with all the hoops ⁓ You can do this stuff, but I do think in the future this will be much more accessible this kind of like command line utility to avoid using websites and I just mentioned that because like
Alan Smith: Yeah.
Patrick Keenan: Yeah, the AI is clicking around all these UIs, but like, there's only a very small step to be like, I'm doing this 10 times. I'm now going to make this into a CLI instead. And like, why would I run a bot when I can just kind of like programmatically do this? ⁓ And then the other thing is you have web MCP, which is basically voluntarily turning your website into an AI service. And so we'll just have all this stuff colliding so that like,
Alan Smith: Yeah, nice. Yeah.
Patrick Keenan: you design for agents rather than users primarily. And then what does that mean for the user interface? ⁓ Is it voice? Is it text? Does it change? Yeah. So it's very exciting.
Alan Smith: Yeah. There's a lot of incentives for people to do that. There's positive reasons to make your stuff more accessible to agents. And we haven't talked about SEO versus GEO or AI SEO, right? ⁓ This is a huge new ball game. Nobody really knows a lot about it yet, exactly how SEO works within ⁓ agents. But to first, as a first step, make your website more accessible to other agents, If there's more searches showing up in agents, that's where you want to be. So there is a very positive incentive structure for lot of businesses or products and services to show up in that way. And so to make your stuff more ⁓ open, that's cool. Yeah, I like that.
Patrick Keenan: Yeah. I mean, the other links, I'll put them in the notes, but they're really just about OpenClaw ⁓ and then some notes about design and AI. ⁓ I think the last one I'll just say is somebody reverse engineered the Apple neural engine. ⁓ And so I think increasingly our, our compute will be local, you know, and that's like another, it's like basically deflationary for AI tokens. If you can run everything on your local machine.
Alan Smith: Dude. Right, yeah, it's like, yeah, you don't have to, if you already have a computer sitting around that has a bunch of cycles in it that you're not using, that's money in the bank, right? Like, it's a, yeah, and we all have it.
Patrick Keenan: Mm-hmm. Yeah. Yeah. Yeah. Like, will we look back and be like, why were we not using our computers when we weren't at our desk? It's so crazy.
Alan Smith: Yeah, it was like remember SETI search for extraterrestrial intelligence and how it you would install the thing and it was like Doing this huge job of like searching the entire sky for aliens and all these little like code chunks that you would download and you know that I feel like to me that was the first big distributed computing project and The fact that okay is coming back again, you know distribute computer
Patrick Keenan: Yes, yeah, yeah.
Alan Smith: So you're just adding yourself to a botnet of like, you know, this, yeah, okay, cool. Yeah, yeah, it's true, right? It was all a cover story. Yeah, we'll make sure your house stays at 21 and a half degrees, no problem. All right, man, thanks for educating me this week. It was always super fun to talk about this stuff. We'll see you next week. Thanks everyone for learning in public with us.
