One AI Bot or a Whole Team? The Context-Switching Trap

[00:06] Dave Pengelley: Uh gentlemen, gentlemen, what a week, what a week. Uh when I say that, at least I wasn't up at three AM this morning. I enjoy that part of today. I'm feeling a little bit more bright eyed and bushy tailed than last week. [00:18] Matt: Yeah. [00:18] Dave Pengelley: Uh how how are you all doing this week? [00:21] Matt: Uh [00:21] Richard Webbe: We're [00:21] Matt: I'm pretty [00:22] Richard Webbe: good. [00:22] Matt: good. But yeah, [00:23] Richard Webbe: Yeah, [00:23] Matt: that that [00:23] Richard Webbe: we're [00:23] Matt: was uh that [00:24] Richard Webbe: good. [00:24] Matt: was rough last week when you were so tired. Felt really bad [00:27] Dave Pengelley: I [00:27] Matt: for [00:28] Dave Pengelley: just [00:28] Matt: you. [00:28] Dave Pengelley: grumpy. It just made me grumpy. I'm still Grumpy, but I'm I'm I'm awake and grumpy this week. [00:33] Matt: No, it's all right. Richard. [00:36] Richard Webbe: I'm I'm awake. We've got daylight. When daylight savings kicks in, unless I'm getting up and cycling really early. It does give me a bit of jet lag, so I'm suffering a little bit of jet lag. [00:47] Dave Pengelley: Yeah, it is weird the crossover. We were we're camping and I didn't notice it camping because you're in a tent and it's weird timing and waking up and everything's uncomfortable anyway. But the last few days I think I'm I've felt it a little bit more. Um [01:01] Richard Webbe: I had [01:01] Dave Pengelley: Yeah. [01:01] Richard Webbe: I had I had to laugh at my brother speaking of AI and technology and that of course there's no more changing clocks, is there? Your phone changes automatically, your watch changes automatically, and I was staying at my brother's Airbnb, which he'd built down at Rye underneath his uh his beach house. And I was staying down there. I thought this would be good. And we'd had a nice dinner out the night before, and I knew I had to get up early for a golf comp. And I opened my eyes and looked and went, ah, he's Given us an AI, you know, sorry, a internet connected clock, so he'll have his time correct to go to golf. But as is my brother, I hear this calamity upstairs and this guy falling down the stairs and complaining and screaming. Every clock in the house was AI or internet connected, except for his. [01:51] Richard Webbe: But [01:51] Dave Pengelley: Uh [01:51] Richard Webbe: his [01:51] Dave Pengelley: yeah, [01:51] Richard Webbe: phone [01:52] Dave Pengelley: our [01:52] Richard Webbe: battery [01:52] Dave Pengelley: uh our microwave. [01:53] Richard Webbe: die. [01:53] Dave Pengelley: I've I've got update the microwave. My wife made me do a wall clock yesterday. Uh the oven I haven't done yet. There's a few manual clocks around the house. We do have one of the um Amazon wall clocks or links to the Echo, and that's great. Uh I don't think they sell them anymore. It was like a short-run product they made for a little brief window of time. But it's so good because whenever you set timers, it's in the kitchen. So you set a timer for 10 minutes and you get little little blue little link, little lights that light up and you can just see it count down. So you get this [02:20] Richard Webbe: Oh, [02:20] Dave Pengelley: like [02:20] Richard Webbe: that's [02:21] Dave Pengelley: visual [02:21] Richard Webbe: cool. [02:21] Dave Pengelley: on the analog clock of your kitchen timers. Just incredible. So that one updated automatically when we got home from camping. I was like, Oh, that clock's that clock's good. That's Nice. [02:30] Matt: That actually [02:31] Dave Pengelley: Um, [02:31] Matt: sounds like such a like we people talk about, you know, going from zero to one. Like if you either you don't have it and then you have a version of it, and then you already have a version and you go to that increment from like one to one point one. That sounds like an incremental upgrade that's super [02:46] Dave Pengelley: I'm if not sure you're not [02:46] Matt: worth [02:46] Dave Pengelley: gonna be [02:46] Matt: it. You know, like a lot of people talk about getting a really good pillow or [02:50] Dave Pengelley: a bit little more. [02:50] Matt: a really good bin. You know, I I think a smart wall clock like that sounds great. [02:55] Dave Pengelley: Well, especially because this was before they had the like the echo show range with the screens and so when we just had the vertical black speaker there was no visual affordance of what the timer was doing so you'd be like oh how much time's left on my timer three minutes 30. how much time's left on my timer seven minutes and twelve uh and so it was really annoying back then so now that they've got a lot more screens and we've got one on the other the echo show with the screen on it on the counter now that does show the timer so I mean it's less critical. Maybe that's why they phased it out because I went ah we want people to buy the ones with the screens anyway. But just as a you know that blended technology, that sort of steampunkish, as an analog clock with this digital layer on it, really nice. Great [03:38] Matt: Yeah, it's [03:39] Dave Pengelley: product. [03:39] Matt: cool. It's funny how like you mentioned the oven clock is one you didn't change. I was just thinking like every house I've been in, like the oven clock is such a a a pivotal thing. Like it's just one of those clocks that you always refer to. I don't know why. It's like [03:55] Dave Pengelley: Ours is an under-the-bench oven, so it's not like not an eyeline. So our microwaves and our eye line it's up higher we've got a higher microwave but our oven is down below the cooktop and so it's just one that we don't reference as often we don't use the oven timer because we use a um the echo device for [04:12] Matt: Yeah. [04:12] Dave Pengelley: our for all our oven timing and stuff. So yeah we we haven't got there. The microwave one we had to reset pretty quickly because we use that one a lot. Um and and reference uh we should we should make note that for our Queensland viewers we are now at eleven AM for you. No longer a midday show for the next few months. Uh [04:31] Richard Webbe: They're [04:31] Dave Pengelley: and [04:32] Richard Webbe: they're [04:32] Dave Pengelley: for [04:32] Richard Webbe: acutely [04:32] Dave Pengelley: the all [04:33] Richard Webbe: aware [04:33] Dave Pengelley: the states, [04:33] Richard Webbe: of such [04:33] Dave Pengelley: right? Um [04:34] Richard Webbe: things. [04:34] Dave Pengelley: Adelaide, does Adelaide have daylight savings? Adelaide's probably consistent. I think they DST, but Perth? [04:39] Richard Webbe: Earth doesn't. [04:40] Dave Pengelley: No, so Perth, we've gone from a uh what 10 [04:42] Richard Webbe: Nine [04:42] Dave Pengelley: a.m. [04:42] Richard Webbe: AM. [04:42] Dave Pengelley: show. We're back to a 9 a.m. show. Maybe we can get Mano to join us for a 9 a.m. show. [04:47] Matt: Maybe. Get a [04:49] Dave Pengelley: Doubt [04:49] Matt: book him in [04:49] Dave Pengelley: it. [04:50] Matt: like three weeks in advance, [04:51] Dave Pengelley: He's too [04:51] Matt: though. [04:51] Dave Pengelley: busy, too busy killing it. Uh well, speaking of Mano and shows, let's not forget that we are doing a show. [05:02] Dave Pengelley: Welcome to [05:13] Dave Pengelley: AI operators, the show that's like a pub chat with mates just talking about what's happening in the world of AI and technology, shooting the breeze, sharing our own experiences and learning from each other. And hopefully you can learn something from that as well. Jump in the chat, say hello. Drop any questions or comments you've got. We love audience participation. One of the reasons we do a live show, not just so you know we are not robot avatars, we are real people. [05:44] Dave Pengelley: Um, [05:46] Matt: It's true though, but any any questions that pop up, like it's actually really fun. So any random question you could ask how old this hoodie is and why I'm still wearing it. Um the what were your favorite uh [05:59] Dave Pengelley: I [05:59] Matt: you [05:59] Dave Pengelley: I [05:59] Matt: know. [06:00] Dave Pengelley: I I'm im you're implying that it's a that's an old hoodie. [06:03] Matt: It's [06:03] Dave Pengelley: Like just [06:03] Matt: ancient. [06:04] Dave Pengelley: Yeah. [06:05] Matt: Yeah. [06:05] Dave Pengelley: It it does look like fashion wise a good twenty years old at least. [06:09] Matt: Yeah. Did you ever have a a JJ's near you? [06:13] Dave Pengelley: Yeah. [06:14] Matt: So it was from JJ's Ringamall, which does not exist anymore. [06:18] Dave Pengelley: I think there's [06:18] Matt: Um [06:18] Dave Pengelley: a still the JJs in in Rouse Hill. [06:20] Matt: yeah, right. And yeah, I think it was like early, early high school. So yeah, [06:25] Dave Pengelley: Yeah. [06:25] Matt: it's [06:25] Dave Pengelley: But [06:25] Matt: a while. [06:25] Dave Pengelley: them they mostly do branded like um licensed merch now, JJs. Lot of licensed t shirts, jackets, jeans, everything. [06:33] Matt: Yeah, nice. [06:34] Dave Pengelley: Hmm. Yes, this is this is the AO operator's show, not the men's fashion. Although Richard always very fa very fashionable in the sports coat. [06:44] Matt: I have actually I've got a question for potential live viewers if there are any of you out there. Um let me know like which you're preferring more if you've tried both, either bots or dots. You know, the whole the the grockbot side of things or the open AI dots. outside of things [07:00] Richard Webbe: I've [07:00] Matt: because I'm [07:01] Richard Webbe: I found the launch of that very interesting. And I should say, Matt, thank you for introducing me to my second brain. I'm not sure if I'm using it properly, but I'm kind of and and Claude did some weird things and said it stored some stuff in the new environment you gave me called, is it called Edge? [07:21] Matt: well I I just call it the core. [07:22] Richard Webbe: Core, the core, sorry, that's the core. That's right, yes. So the core. But um I noticed uh, yeah, it's um Grok and uh dot or ChatGPT and Grok have taken two different approaches to personalized bots, haven't they? [07:38] Matt: Hmm. Yeah, it's [07:40] Richard Webbe: Explain [07:40] Matt: it's super [07:40] Richard Webbe: explain [07:41] Matt: interesting. [07:41] Richard Webbe: to the rest of us what that is. [07:44] Matt: What the two approaches are. [07:46] Richard Webbe: Yeah. [07:47] Matt: Oh, the simplest version, I guess, is the open AI one, you have a single entry point and on for now, yeah. And on the uh the grockbot side of things, you can have a whole roster. Yeah, I don't actually know what the maximum is. So yeah, one side you're building a team, and the other side you're building Javas. [08:08] Richard Webbe: Yeah, Java's good reference, yeah, from the uh the super superhero space. And [08:15] Dave Pengelley: Yeah, [08:15] Richard Webbe: so [08:15] Dave Pengelley: so [08:15] Richard Webbe: you [08:15] Dave Pengelley: that always on entity assistant that just knows what you're about and just generally pitches in and helps. Um, although, funnily enough, off the back of AI uh OpenAI launching their dot, the GrocBot came out with a new feature like 24. [08:29] Richard Webbe: can see that. [08:30] Dave Pengelley: Hours later going, we now have a primary bot. Here's your go-to bot. You can have additional bots, but here's your main one. Here's the one you should go to the most. And so I was [08:39] Matt: It's [08:39] Dave Pengelley: like, [08:39] Matt: weird [08:39] Dave Pengelley: okay, they're [08:40] Matt: because [08:40] Dave Pengelley: trying [08:40] Matt: like [08:40] Dave Pengelley: to simplify because people go, Oh, do I need one? Do I need many? Like, why can't I just have one? And it's this is what we talked about a little bit last week. I don't think they know their ideal customer [08:50] Matt: I'm not sure. [08:51] Dave Pengelley: profile. They're trying to create a new market for this thing. They don't know who the best user is for. They don't understand the best use cases. They're just putting stuff out there. Looking at the data, looking at the user feedback, and then just trying to shape it from there, I think. [09:06] Richard Webbe: I [09:06] Matt: Yeah, [09:06] Richard Webbe: think [09:06] Matt: that [09:07] Richard Webbe: I [09:07] Matt: there's [09:07] Richard Webbe: think [09:07] Matt: definitely that like opinionated approach, but then you know, trying to do it slightly differently because [09:13] Richard Webbe: a good idea. that's [09:13] Matt: it's their thing. Um, I think Grockbot kind of appeals to people that are used to teams and you know, speciality differentiation. Um, and it's funny with that primary agent thing because they since the launch of that tool have been specifying that you should have a a a you know like a chief of staff. You [09:34] Dave Pengelley: Yeah. [09:34] Matt: know you should always have one that manages the rest. So now they're like, okay, we'll just put it in there. [09:39] Richard Webbe: I think that's like the classic argument between marketing and engineering. If engineering puts a you know a voicemail uh hierarchy in for you, it'll only be one or two choices, but it'll go 30 deep. Um and if marketing puts in a uh a voicemail, you know, press one, two, or three hierarchy, it'll only Be too deep, but it'll be about 50 wide for the choices. And it's the two left and right brain ways of doing things. I I really enjoy when I'm doing my own research and preparing my own approaches to market with products and services. I do enjoy one butt to kick or one place to go. And I really enjoyed, I was doing some work. Um we put this down as a little wind, but I was doing some complex work on market analysis and value proposition and a few things. Excuse me, that should be on silent. And um I um I I found that the uh the LLM itself was reminding me of things that I'd forgotten, kind of like the second brain you introduced me to, Matt. Uh and uh it may have been pulled up from data there and it goes, Oh, Richard, you don't want to do that. We already researched that. This information's over here, come back there. And that's the value of having one butt to kick or one place to go. It's all in one spot. We don't have to find it. It finds [11:01] Matt: Yeah. I [11:02] Dave Pengelley: I've [11:02] Matt: think [11:04] Matt: I could go down a really com complex rabbit hole here, but I won't. [11:10] Matt: No, I want to listen to Dave's [11:11] Dave Pengelley: do got it. to [11:11] Matt: thing [11:12] Dave Pengelley: No, [11:12] Matt: first. [11:12] Dave Pengelley: I was gonna say, and Matt, you and I touched on this before we hit go uh on the on the live stream, is one of the issues with the one bot approach is that threads and context management and how do you differentiate different projects that you want to work on back to back? Because historically, you know, we talk about you know the clawed code users and stuff, they've got six terminals. Open and they're swapping between different terminals to do different projects and they're give that one a command and keep that one going and then get that next one going and do the next one and then do the next one. If you've got one bot with one thread, how do you do that? Because as soon as you give it a second instruction, it's now going to be context switching itself. Even if it can disaggregate and know that there's different threads in the background of memory and context that it's running on, if it's a bot that is basically doing one thing at one time, you can't start multi threading and multitasking. And doing multiple things all at once. It can do the one thing. And if you start trying to back to back and go, do this, now do this, now do this, now do this, that that gets really messy, both from a you managing, like where was I up to in that conversation, but then also for the robot to know what it's meant to be actually working on doing next with its limited set of resources and powers. So I think, yeah, I found that really early with like Hermes. I tried the telegram thing and I couldn't do it because the telegram session would start blending and blurring the Different things to you know, I didn't like it. [12:31] Matt: Yeah, it's a weird thing. I would liken it, you know, to imagine you're organ orchestrating a special event for your family or a friend or something like that, and you're speaking to one person just via SMS, and you know, you're you're messaging your auntie, and then your auntie's delegating things out to the rest of your family, and you you're saying, like, hang on, did did so and so finish that thing? And they go, Well, yeah, let me go find out. But you know, there's not that really tight feedback loop. You have to actually communicate. communicate yourself through those awkwardnesses, which is true. And like I'm I'm finding those things now. Like I my front end of this of choice now is actually T3 code, which is just a desktop application that aggregates all of my CLI agents together. So I'm talking the Codex agents, Claude Code, OpenCode, Grok, Cursor, Pi, you know, all of the various ones that I'm using right now in one interface, but it is still that same sidebar list of threads. And yeah, I'd bounce between them. So yeah, [13:34] Dave Pengelley: Yeah. [13:34] Matt: having one one point of int in um interaction is nice. It's it's nice to just be able to dump your thoughts into it. But then when you actually are in like a flow state of doing work in a specific area, you kind of need to be in that specific area. [13:50] Dave Pengelley: Yeah. [13:50] Richard Webbe: Divide and [13:51] Dave Pengelley: Like [13:51] Richard Webbe: conquer. Yeah, [13:52] Dave Pengelley: like [13:52] Richard Webbe: I find [13:52] Dave Pengelley: I'm [13:52] Richard Webbe: it [13:52] Dave Pengelley: I'm [13:53] Richard Webbe: sorry, [13:53] Dave Pengelley: I'm [13:53] Richard Webbe: go on, [13:53] Dave Pengelley: I'm [13:54] Richard Webbe: go. [13:54] Dave Pengelley: so yes, I'm I'm I'm developing a a video game which is fairly complex and so I I'm I talked about how I'm using Chat GPT, like the free chat stuff and integrating the project management with that. And then I've got a Hermes agent that's doing all the actual delivery and the writing of the code. But you know, that's one of those ones where I might be thinking, oh, okay, for the story, I've got this thing. So I can open a thread and start talking about story and lore or something. But then I go, Oh, hang on, I want to talk about the actual fight mechanics or something. So I'm gonna have a different thread where I talk about fight mechanics. And if I want to add something to the story, I don't put it in there. I go back to where we're talking about story so that way chat GPT knows we're talking about story here or we're talking about the mechanics, or here's Feedback on the latest um the PR that has submitted to GitHub. And so we're working on the active iterative piece of code while I'm also running one around potential future features. I'm running one around the marketing campaign for it. I'm running one around story and lore. If I just had one bot for managing that game, I couldn't be easily and effectively running all of those things all at the same time. Um clearly for me or for the robot. [14:54] Matt: I think it's more so for you. Like that's that's something that I think is um the real limiting factor because I feel that. Like if I go through like a single thread thing, I'm starting to feel friction and discomfort as far as but like how is it going? Can you give me feedback? You know, like where are we at? Like what is the current vision, goal set, task list? You know, what's the progress percentage? You know, like I want to be able to see and feel these things, whereas you kind of can when you're in like an individual thread, but the one centralized one, it just feels so disconnected. There's another kind of human interaction thing that I would like to use as a as a reference point here. Imagine you're at work and your boss comes up to you and you know you're already in the middle of something, and then they say, I need you to do this, and that quintessential, everyone's experienced this. They're like, which one of these priority ones do you want me to do first? So, and then your brain starts to get frazzled. They don't understand what they've previously told you. So they go, Well, this one. So it's kind of that same thing. [16:00] Richard Webbe: I think these points are really valid and and people work different ways. It is um focus versus communication versus confusion, and you've got to get them right. I mean, I have people say to me, I had someone say to me the other day, um, I I sent you a note. And I yeah, pick one of the 84 different ways that someone could send me a note and I'll go looking for that. And I've said this before, and I haven't written it myself, and I probably should, but if someone could come up with an app. That sits on my phone beyond the helpful AI we've got now that just says Matt sent you a note, he sent it to you on this messenger. Time to respond. Whereas you know, I get a note and it pops up on my screen and says, you know, Matt sent me a note, and then I go through all these different applications looking for what the note was and where it came in. And there's a lot of room to separate for focus and confusion, but to combine. For faster communications. [17:01] Dave Pengelley: That's one of the criticisms of OpenAI though at the moment. There's like eight different ways to interact with Chat GPT. Is it do through the web chat? Is it through their work thing? Is it through codec? Is it now through the dot? Is it like which one do I do where? And which one's using what usage? And how do I how do I manage my amount of things I can do and which is the right one? Because they all kind of do the same thing, but they do it differently. And some use token percentages and some don't. And it's that's like uh [17:28] Matt: I think it's based on the job. And yeah, that that you only really know why you should use a different numbered iron in golf when you are used to playing golf and you know which iron is actually better for which reason. Um otherwise you just use a driver for everything and say, bugger it. I'm getting out of this sandbox with the driver, get lost. [17:49] Richard Webbe: Yeah, [17:50] Matt: Um [17:50] Richard Webbe: not gonna [17:51] Matt: so [17:51] Richard Webbe: work. [17:51] Matt: yeah, for me, what Richard you're saying there is like a signal aggregator, like an external thing where you know it can proactively keep Keep track of things coming in. For me, I actually think that's what dots are really good at. And like the dot for me has been the most proactive by itself. So actually having initiative to say, hey, like literally to like I have a real example of this. I won't share my screen or anything because it's proprietary stuff. But I literally had it say, you've got a message from this client team from this Slack channel. They're asking you this thing. So I just told it at one point, you know, can you listen for mentions to me? Because I probably will miss them, especially in really busy Slack workspaces. And yeah, it came up saying, hey, this this team member has this issue. It's probably actually this thing that they're looking for. So it it actually went through and it found the stuff for me. So like, can I set GrocBot up to do that? Yeah, probably. But I I I I didn't even set the dot up to do that. It just sort of started to proactively start suggesting things. So I think all of these different surfaces that you mentioned, Dave, like each one should be like specialized for a different thing. So I mentioned Laurent Tan a few times, the creator of Pea Stack. She talks about having an external and an internal loop, where the external loop is is the one that you don't really focus on. It's the one that is that signal aggregator. And it can escalate things to you or it can forward things into your internal loop. And the internal loop is kind of like your codex sessions. You know, it would be Whatever you want it to be, I'm about to hook it up to use the MCP tooling inside T3 code so it can spin up any session [19:38] Dave Pengelley: Yeah. [19:38] Matt: from any agent from those external loops. [19:41] Dave Pengelley: Yeah, [19:41] Matt: And then yeah, [19:42] Dave Pengelley: nice. [19:42] Matt: that that that's how we're going. [19:44] Dave Pengelley: Yeah, I I suppose the other thing I'm trying to do with i is leverage the uh tokenomics at the moment of dot being free astra usage. And so trying to get it to not go and spin up codex threads and stuff. Cause you're right, if you've got I had a managing Threads, that's one thing, but I'm trying to get the outer loop to be the inner loop as well because it's it's free at [20:06] Matt: Yeah. [20:06] Dave Pengelley: the moment. And so trying to maximize that is the other other thing, like who's doing the work? And if it's just an orchestrator that's sort of that that aggregator layer, that's different than if it's also the doer. [20:18] Matt: You talk talking about tokenomics and freeness. I j I'm kind of blown away how at the cost efficiency of both Opus five five and um Sol six one. [20:29] Dave Pengelley: So Six one's amazing. [20:31] Matt: I don't understand like how we've made all of a sudden such a huge leap in in efficiency and like cost per task. Um I was not going to use up my codecs weekly. And I set something overnight to use Astro and I had an hourly refresh to be like, where are you at? Keep going. Where are you at? Keep going. And overnight I couldn't even burn through my Astro usage. So I don't really understand what's going on. [21:00] Dave Pengelley: Um because 6.1's been so efficient. And you know, because I was away and I had the weekend where I wasn't really doing stuff because I was camping, I did manage to get a little bit of remote access. And I said, you know, for my SaaS chart reporter, I haven't given that a full code review with the latest agents. So I told Astra. I said, Astra on like a high, maybe even a high Astra. So expensive Astra. I said, just do it. Just go double check that you know stats are calculating properly, the everything, like just do it and find gaps and so it came back with a whole laundry list of things I said cool go create github issues and then orchestrate soul 6.1 agents to execute it and so it's it did that um and uh you know a few loops and I'm about ready to merge [21:42] Matt: Okay. [21:42] Dave Pengelley: it all and push all the updates across into you know the three surfaces my marketing site the app itself and the documentation site but but uh yeah I mean if I was pure Astra and not orchestrating Sol 6.1 I wouldn't have got that done as it is but but Astra still definitely burnt Where I'm just if I'm just living on Sol six point one, quite a last for forever. It's it's amazing. [22:06] Matt: I I'm so uh quietly curious to know what they do with Astra 6-1, because like how they had to pull it back, they they hit some sort of regulation blocker and didn't release it, and that's kind of where Soul 6-1 came from. It's not actually like an incremental soul, it's more like a like a mini astra in a way. Um if if if that six-one astro ends up having Those huge efficiency increases as well, and is somewhat more pleasant to interact with. That's that's crazy. Like unfortunately, they've got such a huge bar to try and meet now with Opus 5.5 that like the jump between Opus 5 to 5.5 was immense. Um [22:49] Dave Pengelley: Yeah. [22:50] Matt: I haven't even touched Fables really at all since uh that Opus came out. [22:54] Dave Pengelley: Yeah. [22:55] Matt: And [22:55] Dave Pengelley: I I I haven't touched Claude since Fable first came out and then got pulled again. That's that was like my last two. weeks of of anthropic [23:01] Matt: Yep. [23:02] Dave Pengelley: subscriptions. So I'm I'm way behind on on having an opinion on Opus. [23:06] Richard Webbe: So [23:06] Matt: It's [23:06] Richard Webbe: let's [23:06] Matt: fine. [23:07] Richard Webbe: take oh sorry, let's take the AI conversation if I'm not hijacking the conversation, which I usually do. Let's take that in a different direction. There's been so much in the news. You meant to want you're on your news thing now, but about how everyone's freaking out about agents getting out of their sandboxes and out of their places and getting into different parts and places they shouldn't. And uh not being a political podcaster, I thought it was hilarious at how OpenAI and a few other organizations have announced that they've gotten into places they think they shouldn't have got there. And everyone's going, oh no, the agents are going to destroy us and take over the world, and it's the start of Skynet and all that. But really, all it is is lazy companies and lazy governments leaving their data openly available that a three-year-old with a web browser could get access to. What do you think about that? [23:58] Matt: I think you I think you're you're right. Um the the wording is is a bit harsh, but um yeah it I I can't remember who it was, whether it was like you Dave or Richard that mentioned i it's kind of like water flowing into recesses of a [24:14] Richard Webbe: Oh [24:14] Matt: of a [24:14] Richard Webbe: yeah. [24:14] Matt: chasm. You know, it will find the things [24:17] Richard Webbe: Oh [24:17] Dave Pengelley: Yeah. [24:17] Matt: because [24:17] Richard Webbe: I [24:17] Matt: yeah that that's [24:18] Dave Pengelley: Well [24:18] Matt: its goal. [24:20] Dave Pengelley: we we we brought up OpenAI's response to the whole Medicare thing last week, right, and discussed that. And and basically the finding was [24:27] Richard Webbe: a good idea. it's think [24:27] Dave Pengelley: they left the keys on the table. Um, and then got some and it wasn't even keys, they just left the folders out. But they, you know, they they they they they they they put the folders in the drawer and they they sort of closed the drawer but left it ajar. Like they didn't push the drawer shut, they didn't lock the drawer, they just left some folders sort of slightly visible through a crack in an in an ajar drawer, [24:47] Richard Webbe: I [24:47] Dave Pengelley: and [24:47] Richard Webbe: love [24:47] Dave Pengelley: then they're [24:47] Richard Webbe: your [24:48] Dave Pengelley: like, [24:48] Richard Webbe: metaphors, [24:48] Dave Pengelley: Oh, we [24:48] Richard Webbe: guys. [24:48] Dave Pengelley: got hacked, oh the a the AI found our secret documents that weren't public. Well, they kind of were public. You left them exposed, you didn't lock them away. It just it just found them. Like it with and like Matt, other people have scraped these things before. Like, why are they making such a big deal about it now? It's because it's to say, hey, we need to put more legislation in to keep you safe. It's like what Reagan say, the scariest words are I'm from the government and I'm here to help. [25:14] Richard Webbe: Absolutely. [25:14] Matt: Yeah. [25:14] Dave Pengelley: Like, [25:15] Matt: I [25:15] Richard Webbe: Yeah. [25:15] Matt: think you said [25:16] Richard Webbe: And [25:16] Matt: that [25:16] Richard Webbe: it's [25:16] Matt: last week. [25:16] Richard Webbe: just [25:17] Dave Pengelley: yeah, [25:17] Richard Webbe: insane. [25:17] Dave Pengelley: that's [25:18] Richard Webbe: It's [25:18] Dave Pengelley: a good [25:18] Richard Webbe: just [25:18] Dave Pengelley: quote. [25:18] Richard Webbe: insane. [25:18] Dave Pengelley: I'm gonna use it all over and over again, Matt. [25:21] Richard Webbe: But it's just insane that the government left the keys or the files on the table, as you said. I mean, you know, we've all had different security clearances. And some of them say, you know, things like I think Nv2, you know, need to know. So if a file is on a table and you read it and you don't need to read it, then you need to arrest yourself. But the trouble is you don't know you shouldn't read it until you start reading it. And that's the logic they're applying to all of this stuff. I mean, you know, the internet is the internet. Like we said, it's like water. And and you know, if I if I'm going to create a substantially good LLM or DL. Database or Wikipedia, I need to go out and find that information. And as Dave said, if you leave the keys on the table or you leave the file open, I'm going to read it. And I'll make a decision later on if I need that information or not. I think the AI companies were totally well well placed and not guilty of anything. And that they came out and said, look, we're so they were so good. They came out and said, we're so sorry that we read that file that you left out in the filing place on the street on the footpath for a Right [26:31] Matt: Yeah, [26:31] Richard Webbe: using [26:31] Matt: there [26:31] Richard Webbe: that's [26:31] Matt: was a really [26:32] Richard Webbe: metaphor. [26:32] Matt: good video. I think it was you, Dave, that shared this a few episodes ago. And it was, you know, you put the put the person in a room. [26:38] Dave Pengelley: Yeah, [26:38] Matt: And [26:39] Dave Pengelley: give me the keys. [26:40] Matt: yeah, it's like, all right, you need to you need to find out how to get out of this room. Um, but but yeah, you can't use the keys. And then like they put the key on the table and then they walk out. And so naturally the thing just grabs the keys and leaves because it's the easiest option. Like the the thing that people keep forgetting about is that like LLMs. And you know, through agent tooling, they will always go the least friction option, [27:05] Richard Webbe: Yeah. [27:06] Matt: like they'll go the fastest option, the shortcut. And if the shortcut is the wrong is the wrong thing, um, I didn't necessarily know that. Uh it's just the more efficient option. [27:16] Dave Pengelley: I mean you you say, Richard, these things are the sort of the the average, you get the average result, like it's the the the the net average from these things. And you know, I've been experimenting and getting a few of these new models to try and flip some websites and just say, what if you did what if you Build this thing or even for my own personal website. I just redid my whole portfolio website the other week. And it comes back with the generic, like what everything's the tans and the browns now. Everything's the browns and the rust colors. And it's like generic tan. And I was like, no, no. And I said, Oh, this this does look like an old website. I might talk to them as a potential uh client. Um, flip them and it gave me generic browns. And I'm like, that's not even their color scheme. Like they have a color scheme. Their website might not look real great, but they have a color scheme. Can you go use that? He goes, Oh yeah, sure. And I'll look, look, I'll do you a favor. I'll grab their logo while I'm there as well. I'm like, oh, really? You think we should copy their logo? Like, come on. I'm like, this is frontier models. This is this was Grokbot, like four Groc using 4.7 on the back end, which I know, Matt, you you're highly critical of 4.7, and I understand why. Um Grok [28:19] Matt: It's actually very [28:20] Dave Pengelley: models. [28:20] Matt: good on legal stuff. [28:21] Dave Pengelley: Grok models can be very lazy. That that's my experience from before they were agentic, like Grok 4.3 or whatever, before they really understood tool tool handling. Just lazy and they just try to push work back to you. They'll do some analysis and say, now go do it. And it's like, you do it. I'm giving you the job. [28:37] Matt: Yeah. It's um surprisingly very good on like discussing legal stuff. So, you know, if you need a like a in a pinch I don't know, just chat about contracts or agreements or how does this affect XYZ? [28:53] Richard Webbe: Loopholes, yeah. What's [28:54] Dave Pengelley: Yeah. [28:54] Richard Webbe: what's the codependent result of me changing these words? [28:58] Matt: Yeah, it's it's very good at That kind of thing for some [29:01] Richard Webbe: Unintended [29:01] Matt: reason. I don't understand [29:02] Richard Webbe: consequences. [29:02] Matt: why. [29:02] Richard Webbe: Yeah, [29:03] Matt: Um [29:03] Richard Webbe: yeah. [29:03] Matt: but yeah, it would be fascinating to see if they or I guess when they achieve their super intelligence as now Elon's uh forcing everyone to say, [29:13] Dave Pengelley: No. [29:14] Richard Webbe: S I, [29:14] Dave Pengelley: Why? [29:15] Richard Webbe: let's go for S I, [29:16] Dave Pengelley: That [29:16] Richard Webbe: not AI. [29:16] Dave Pengelley: yeah, [29:17] Richard Webbe: I love it [29:17] Dave Pengelley: that [29:17] Richard Webbe: that way. [29:17] Dave Pengelley: that's [29:17] Richard Webbe: I love how [29:18] Dave Pengelley: that's [29:18] Richard Webbe: he changed [29:18] Dave Pengelley: a [29:18] Richard Webbe: the poet laureate changes the language. [29:21] Dave Pengelley: I [29:21] Richard Webbe: Look, [29:21] Dave Pengelley: thought [29:21] Richard Webbe: it [29:21] Dave Pengelley: what [29:21] Richard Webbe: is the [29:22] Dave Pengelley: was did did Elon push that or did Trump push that? Because it felt like a Trump thing that that I saw originally and then obviously Elon and Trump like [29:29] Richard Webbe: one. [29:29] Matt: I'm I'm not sure what saying. [29:30] Dave Pengelley: In public they they fight, but in in private they're probably like best buddies [29:34] Richard Webbe: I [29:34] Dave Pengelley: still. [29:34] Richard Webbe: know [29:34] Matt: Yeah. [29:34] Richard Webbe: that fight that they had in public was on purpose to stop people burning Teslas. [29:38] Dave Pengelley: Yeah, exactly right. [29:40] Matt: So it's kind of like um the fight between Lauren and Tebow. I I feel like that is entirely marketing as well. Like it's [29:50] Richard Webbe: Yep. [29:50] Matt: not necessarily just like a a childish beef. I think they're doing it for engagement. [29:55] Richard Webbe: Yep. I think one hundred percent. Uh [30:00] Dave Pengelley: We'll we'll we'll we'll look back on any on any news uh topics and things, but I want to spend some time doing this. [30:18] Dave Pengelley: Wins of the week. The week. Uh who's got wins of the week? I see you've been um generating some marketing material. Materials, Richard. Um, [30:32] Richard Webbe: Well, [30:32] Dave Pengelley: which [30:32] Richard Webbe: like [30:33] Dave Pengelley: model is is generating these for you? [30:35] Richard Webbe: well, I'm not um they're not uh that's just ChatGPT doing an image for me. The articles are written by me, [30:41] Dave Pengelley: Yep. [30:41] Richard Webbe: but my win was actually having so the way I work with AI, and and people are are all very interested in this and ready to point the finger, you did that with AI, you didn't do it yourself. I usually find myself literally so I use it. I'm driving along and my brain, like you know, I've I have an ADHD brain, it's racing along at a million miles an hour, and I'm thinking, you know, butterflies, ducks, AI, you know, any random thought. [31:08] Dave Pengelley: Yeah, [31:08] Richard Webbe: And something comes [31:09] Dave Pengelley: I'm [31:09] Richard Webbe: into me, I go, [31:10] Dave Pengelley: not sure. [31:10] Richard Webbe: you know, that's a trap for the world. If that happened and that happened, that would cause that and that would do that. And so then I engage my um LLM by a voice while I'm driving, and I ask it all the right questions. Because we know in any problem solving with any intelligence, and and uh Socrates said it, uh, the definition of intelligence. is something that knows how much it doesn't know. So I start asking all these questions. Well, what's the percentage of that? How many people do that? And is this working? And and that builds my knowledge to say, you know what, I was right. That is a gap in the market and people need to know about that. And that's some value add I can do. And look, 90% comes from my experience. But a large part of those articles comes from the research I get to do more efficiently to double check that I'm not still in the dark ages of my views. So yeah, I I that's my win. In I'm I'm engaging with the LLM and the AI in a different way, a way that I knew was always appropriate, but it is more as a prosecutor of the LLM with distrust and then getting the value back. And like you said, though, several times I'll say, do this, it'll go, no, you do it. No, go, no, you do it. I asked him. So I've got to turn that question around so that I it comes back to me at the right angle. And not the popular angle, as we said, the subnet average of [32:29] Dave Pengelley: Yeah. [32:29] Richard Webbe: everyone. We were talking about how companies are interviewing people at the moment. And I've had some friends experience this recently. And one of the things that companies were doing is they run their people's CVs that are coming through a system called ATS, which is, I think it's automated talent [32:44] Dave Pengelley: App applicant [32:44] Richard Webbe: system. [32:45] Dave Pengelley: tracking applicant tracking [32:46] Richard Webbe: App [32:46] Dave Pengelley: system. [32:46] Richard Webbe: contractions, there you go. Um and whatever it is got a bit of AI in there, looks at who you are, and and they well, we had a resume from this person before, and are they a criminal and all of that? But then what they do is they do a physical Thing, and this is mainly the larger companies that AI would do. They then, like in I've got a friend of mine, very senior executive, uh, Chris. I work with her a lot, uh, very, very well respected. And she said to me once, Richard, if anyone's ever going for an interview at a company, there's no way they should interview with their peers. That's undermining you from day one, and it's the wrong way to do it. Uh, and I'll explain in a second. So, what these companies are doing now is telling applications. For jobs that they have to interview with everyone on the team. And it could be six or seven people. And it's like the UN Security Council. If someone doesn't like you, then they vote you out and you're gone. And I say, kind of like AI does a similar thing with fax, it creates a very average response. It is actually averaging out and making sure you're not getting very many intelligent responses because you're allowing one average. Person to say, Yeah, I don't like that, and off they [34:02] Dave Pengelley: It [34:02] Richard Webbe: go. [34:02] Dave Pengelley: it it's weird you mention that. I I was in an interview um a few months back for a a good looking um leadership role in a major multinational household name company that's on I I won't mention their name, but you know they're doing a lot in technology and computing space. Uh and it was for a leadership team of one of their pre sales groups. And in one of the interviews, [34:26] Dave Pengelley: yeah, so in one of the interviews, um it was a peer. Manager from one of the other pre sales teams that covered a different product area, which is fair. I mean management management, that's fine. But also someone I would be who would be reporting to me. And I thought that was that's really strange. Get get someone who I will be their boss of to assess whether they think I should be their boss. And I get they were very tenured, they'd been there fifteen, twenty years. They they probably had good wisdom and knowledge, but still it's a weird thing to say, pick your own boss [34:55] Richard Webbe: It's [34:55] Dave Pengelley: and give [34:55] Richard Webbe: not a [34:55] Dave Pengelley: feedback [34:55] Richard Webbe: weird thing. [34:56] Dave Pengelley: on [34:56] Richard Webbe: It actually creates uh confirmation [34:58] Dave Pengelley: it's a conflict of [34:59] Richard Webbe: bias. [34:59] Dave Pengelley: interest. [35:00] Richard Webbe: Yeah, it creates confirmation bias. It creates the wrong conflict. So, you know, if I'm in a if I were, you know, let's say you and I were you were interviewing for the that senior leadership pre-sales role, which we know you're very good at, in a company I was at, and they said, Oh, you know, we're gonna hire Dave Richard, would you go interview him? And I'd interview you, you'd interview very well, and I'd probably be thinking to myself, you know, I don't really want to hire this guy, because he'd be competition for the next uh promotion. So I think I'll give him a gong. [35:27] Dave Pengelley: Yeah. [35:27] Richard Webbe: And it's [35:27] Dave Pengelley: Yeah. [35:28] Richard Webbe: wrong. [35:30] Dave Pengelley: And and the pre-sales thing is it's interesting, there's a lot of contention um around you know th things like activity tracking and monitoring where your time goes, which I'm a big fan of. I think you really need to, even if I was back in as an individual, I think if the business is actually using that to report on impact and make sure time's going in places where it actually has an ROI, that's really important. But obviously most individuals they don't want to clock their calendar hours and do all that kind of stuff. And so, you know, in an interview situation where you talk about how do you make sure the team's being affecting. All of a sudden you're gonna you're not gonna you're not gonna win friends very quickly in that situation. So anyway, um yes, I I think the back back to the brainstorming thing. We we we went down a hole completely tangent, uh which will actually all overlap back into this career thing uh with one of my wins of the week. It's uh interesting. But uh yeah, I working on my game and doing storyboarding, and it's what and you we know that AI is very agreeable and tells you that every idea you have is the best idea and wants to integrate it all. But you know, when you're starting to deal with sort of Complex world building, whether that's for a business project, business plan, or fictional worlds or whatever. I actually did one session with whether it was out in the car that said, Look, I need you to actually be like an editor and tell me where we're generating conflicts. Not every idea is going to be perfect. I need you to think through the entire law and go, this doesn't work, this breaks this because of this. There's going to be a conflict with this one. And then actually give me some collision awareness of just because I'm soundboarding an idea. It doesn't mean that it's instantly something we should adopt because it might break or or conflict with something else. And so she was like, Cool, no worries. I'll I'll I'll do that. The whole tone of the conversation changed from like, yes, let's do that, and yes, let's do that to hang on, you said earlier you wanted X, Y, and Z to be true. That can't be true if you do this. And so I was like, okay, good, that's good to consider, and helped me brainstorm and work through possibilities. [37:26] Richard Webbe: I think that's critical prompting as they refer to it, but I I would just call At asking the right questions and getting the right angle. Confirmation bias within your own conversation and development inside an LNM to get a problem solved or a process sorted, I think is dangerous. And if you don't ask it to look at it from several different angles, the law of unintended consequences will take over, and you'll end up with something wrong or very average. [37:54] Matt: There's also like what Dave was saying, when they they lull you into a sense of security where you have Haven't actually gone down those pathways yet. And it just goes, Yeah, cool, everything's fine. Do you want me to start? Or it will just start doing stuff. And then you sort of have to rein it back and be like, wait a minute, we haven't actually talked through this at all yet. Um, if you go and do it now, we're gonna drift. [38:18] Dave Pengelley: Yeah. [38:18] Matt: So yeah, [38:18] Dave Pengelley: I [38:18] Matt: finding that. [38:20] Dave Pengelley: I I kick off some of my conversations going, This is just a soundboarding exercise. I'm just open brainstorming, do not record anything down, do not take any action. Are you ready? To just have a discussion. Yes, I'm ready to be a sample, Dave. I'm like, all right, let's begin. [38:34] Matt: I [38:34] Dave Pengelley: But [38:34] Matt: should [38:35] Dave Pengelley: you gotta [38:35] Matt: show [38:35] Dave Pengelley: preface [38:35] Matt: you my um [38:36] Dave Pengelley: it. [38:36] Matt: my discuss skill. I I've got a I've got something I've pinched from BMAD, um which is it's literally like putting your agent into plan mode, but it's just conversationally. And it will not stop asking you just discussion questions and bouncing ideas until you're like, all right, this sounds pretty good now. Um yeah, it's [38:58] Richard Webbe: Oh [38:58] Matt: just a nice little [39:00] Richard Webbe: I did um I was going into a negotiation recently, and so I told my LLM to be the other side. I gave them all the data I could, and I gave them what I thought their anger would be, but don't rely on that. I said, right now, let's go into negotiation and have a debate. Uh and I did it vo vocally, voice-wise, but it documented [39:19] Dave Pengelley: Yeah. [39:19] Richard Webbe: everything, and it was brilliant. And when I went into the negotiation, I was more than armed. [39:26] Dave Pengelley: I think I think that's the move actually. Um what you're saying there, Matt. And something I haven't done with mine when I go into soundboarding mode is add to the prompt, um ask me a follow-up question with each of your answers. And because often, you know, I'm thinking if we go in this direction, we can make this do that. And if that clicks on that one, then that happens. And they'll respond. But yes, that would confirm this mechanic. And if we do that, that, that, that, and sort of, you know, justifies and explains in more depth my idea, but then just could sort of stops. And then it's on to me to prompt do the next prompting thing. I almost need it to do that reverse prompting. And then like get it to go based on this, what didn't it clarify or what does it still have questions about and get it to prompt [40:07] Richard Webbe: What's [40:07] Dave Pengelley: me for [40:08] Richard Webbe: missing? [40:08] Dave Pengelley: the next question sometimes? [40:09] Matt: GPT [40:09] Richard Webbe: Yeah. [40:10] Matt: models are terrible at that. [40:11] Dave Pengelley: Ah, [40:11] Matt: They they [40:12] Dave Pengelley: brilliant. [40:12] Matt: will just they'll just output context. They'll just be like, here's a bunch of information. And you're there going, okay, like what's the next turn, you know? And like I I literally will respond quite often and I'll say, okay. [40:27] Dave Pengelley: Well, [40:27] Matt: And then [40:27] Dave Pengelley: I [40:27] Matt: the next one is they they like It would respond, got it. I'm like, got what? Like it's such a funny broken like conversation [40:37] Dave Pengelley: I [40:37] Matt: loop. [40:37] Richard Webbe: They [40:37] Dave Pengelley: I find [40:38] Richard Webbe: do [40:38] Dave Pengelley: myself [40:38] Richard Webbe: freeze [40:38] Dave Pengelley: responding [40:38] Richard Webbe: a lot. [40:39] Dave Pengelley: like [40:39] Richard Webbe: I look [40:39] Dave Pengelley: the [40:39] Richard Webbe: at them [40:39] Dave Pengelley: AI. [40:39] Richard Webbe: and I often say, and and [40:42] Dave Pengelley: Yeah. [40:42] Richard Webbe: and then [40:43] Dave Pengelley: I I I respond like the AI because I find often on the voice chat when you go blah blah blah blah blah, she'll start by going, Yes. And like so so she's like, Yes, she agrees with me and then she continues on to like uh expound or or or talk more about the thing. [40:59] Richard Webbe: can see that. you [41:00] Dave Pengelley: And so when she gives me a whole bunch of extra stuff, I'll go, Yes. That's correct. That's the direction we want to go in and building on that. And so I mean it does help me because I'm I'm a bit of an iterative builder and I connect dots. And so inevitably, when I have an idea and then the AI sort of builds and confirms that around existing other premises, then as she just repeats it to me, even if it's my own ideas in more depth, then that'll trigger the next idea for me to build on that and continue doing stuff until I get to a point where I'm like, cool, that's good. I think we can lock that in. Like that's that's that's dumb, but [41:32] Matt: This is like a another thing to that. Like if somebody else is hearing this and they're just missing this little extra piece, hopefully this might help. Like, I I uh I spent probably 80% [41:42] Dave Pengelley: I a good idea. that's think [41:42] Matt: of my time pushing back and just saying, I don't understand. Like, what are you talking about? Can you explain that again? Um, how does that relate to what I just asked you? You know, but uh just just being dumb, like just saying I just don't get it. Like, can you just just do it again? Again, like, you know, I'm still struggling, I'm a bit too tired, I've got cognitive load, I'm going across 10 different projects. Um, I don't even know what we're talking about anymore. So like that, when you constantly push back, you will get this more like that deeper comprehension that you normally would get if you weren't working with [42:18] Richard Webbe: And [42:18] Matt: AI. [42:18] Richard Webbe: this is why this is why your architecture with the core uh that you've created, Matt, um having the extra data available to mine, you know, independent. Of the LLM is so important because it [42:32] Matt: Yeah. [42:32] Richard Webbe: helps bring forward context and uh and and and the shape of what you're trying to do. I think it's really clever. [42:39] Matt: I just I kind of like I liken that to you know going through [42:43] Richard Webbe: Second [42:43] Matt: high school, [42:44] Richard Webbe: brain. [42:44] Matt: taking taking notes in your workbooks, or you're [42:46] Richard Webbe: Yeah. [42:46] Matt: going to conferences and you've you've brought a workbook and you're scribbling down stuff that the speaker's saying. But [42:50] Richard Webbe: Yeah. [42:51] Matt: you leave that event and you got no idea what you've written in your book. And like you could try and flick through it if you wanted to, but what what am I looking for? Like what's relevant and why? Like that's all it was meant to be for me at the beginning was just this idea of hey, I'm capturing notes, I have ideas, I have thoughts, I've got stuff I need to complete, they're open loops, I don't want to hold on to them. Like there's [43:12] Dave Pengelley: Yeah. [43:12] Matt: another like image I had was Dumbledore like pulling the thoughts out of his head and putting it into that bowl. You know, the it's the same concept. [43:19] Richard Webbe: Have you [43:19] Dave Pengelley: Yeah. [43:19] Richard Webbe: heard of a life scribe? Have you heard of a life scribe? [43:23] Dave Pengelley: Yes. [43:23] Matt: Maybe what is it? [43:24] Richard Webbe: A friend of mine who was a seed investor, I was at a conference in the US and uh she bought me a life Scribe pen. So it uses paper with special dots and it's a recording pen. Um, and so basically you write out, you take your notes, but it records your conversation as you're doing it and links it to as you take the notes, and so it creates a three-dimensional view of the conversation, what was happening, and then it triages that data into you know ASCII character text, so you can just cut and paste and send it back. It's like pre-wearable uh, you know, note-taking [43:58] Dave Pengelley: Yeah. [43:58] Richard Webbe: technology, [43:58] Dave Pengelley: But but [43:59] Richard Webbe: but [43:59] Dave Pengelley: they [43:59] Richard Webbe: it was [44:00] Dave Pengelley: Idea [44:00] Richard Webbe: Really [44:00] Dave Pengelley: was you could [44:00] Richard Webbe: clever. [44:00] Dave Pengelley: replay as you replay the conversation with show you the notes you were writing, or you [44:03] Richard Webbe: That's [44:03] Dave Pengelley: could click on a part of [44:04] Richard Webbe: it. [44:04] Dave Pengelley: the page and it would show you the voice note that happened as you wrote that particular part of the note. [44:08] Richard Webbe: As [44:08] Dave Pengelley: Yeah, it's [44:08] Richard Webbe: as [44:08] Dave Pengelley: 20 [44:08] Richard Webbe: they were [44:09] Dave Pengelley: year old [44:09] Richard Webbe: saying [44:09] Dave Pengelley: technology. [44:09] Richard Webbe: something, I could touch on it and you know you you'd hear what they were saying and I'm writing, I hate this person. You know, they go now I know what I hate them. [44:18] Matt: That's [44:18] Richard Webbe: I never [44:18] Matt: interesting. [44:19] Richard Webbe: said that. No, it's not [44:20] Matt: We [44:20] Richard Webbe: that [44:20] Matt: had [44:20] Richard Webbe: cool. [44:20] Matt: a we had a rep um at a company I used to work at, and he was he was part of the site management team, you know, like they owned all the buildings across the company anyway. He was in there and he had this this pen and it was this big fat chunk on the end of it. And I remember we're asking him what that pen was and it did something similar where he didn't have special paper or anything, but he was signed into this pen and it had accelerometers in it and [44:43] Dave Pengelley: Oh, [44:43] Matt: it would [44:44] Dave Pengelley: wow. [44:44] Matt: know what he was writing. And uh he was telling us stories of somebody who was just testing it out, like saying that you know, so-and-so boss is a Wally and you know that that kind of thing, and they got fired because like it it [44:56] Dave Pengelley: Wow. [44:56] Matt: detected it. So yeah, [44:58] Dave Pengelley: That [44:58] Matt: interesting. [44:58] Dave Pengelley: that that must have been a Newer version because the original ones yeah needed special paper so it could like work out where it was and had a little camera in the tip because I remember sitting there so it was actually visually like identifying where on the page on this special sort of grid identified. Um, and then it had a microphone in as well. And then you got the sort of the newer type, like low low-tech versions, like the rocket book, where it's just using an erasable pen on this sort of hard plastic-ish paper. Um, but every page has like a little QR code, so you can take a photo and there's a little like action icons so you can tick. Which destination you want to send your note to? And then the Apple sort of, you know, you take a photo through the Rocket Book app and all add it to your Dropbox or your G drive or whatever as a picture in a OCR document. [45:40] Richard Webbe: And [45:41] Matt: That's kind [45:41] Richard Webbe: and [45:41] Matt: of [45:41] Richard Webbe: the [45:41] Matt: cool. [45:42] Richard Webbe: the non-legally aware people come up with a whole well you were legally recorded their conversation and there was a test case in Australia and I won't use the name of the companies, but it was a particular bank in Australia that had stolen some businessperson's idea after lending them money to do it. And went off and did it themselves and came back and raided them and gave us our money [46:04] Matt: Wow. [46:04] Richard Webbe: back and took their idea. So the guy in the meeting had recorded it on his phone. But at the start of the meeting, as far as I know with the precedent, and someone will probably send a note in and say you're wrong or whatever, but it went something like this, and I got this from a lawyer, that at the start of the meeting, the the the founder said, Do you mind if I take notes? And they said, Of course, go for it. So he recorded it. When it was brought to be admitted in court. He said, I want to play the recording of the meeting, and they protested and said, We didn't approve of that. And the judge had a sidebar, and the guy played him the start of the recording, and he said at the start of the recording, Do you mind if I note down the details of this meeting? And the judge said to the um the uh plaintiffs or the defendant, he said, They told you that they were going to take the details of the meeting, whether it was recorded on an AI pen. A watch, you know, it doesn't matter. You were aware that everything you were saying was being recorded, and we allow notes to be admissed admitted in court, which are not very reliable, but an AI recording is much more reliable, and you can prove that. So I think the law in this case was great, it applied common sense to the logic, and so therefore I say to everyone: be careful what you say. [47:26] Matt: Yeah. Yep. [47:30] Matt: Same with any kind of surveillance y stuff though, like people freak out, like just don't do the wrong thing. Like [47:34] Dave Pengelley: Yeah. [47:34] Matt: why why [47:35] Richard Webbe: But [47:35] Matt: do you have [47:35] Richard Webbe: that's [47:35] Matt: to [47:36] Richard Webbe: it. [47:36] Matt: do the [47:36] Richard Webbe: If [47:36] Matt: wrong [47:36] Richard Webbe: you're not [47:36] Matt: thing? [47:36] Richard Webbe: going to be a crook, so what are these? They've got these new devices they're announcing now. You've got the pocket, you've got that. They're announcing that your your Apple Watch can now become a recording device, [47:47] Dave Pengelley: Yeah, [47:47] Richard Webbe: your [47:47] Dave Pengelley: granola [47:47] Richard Webbe: phone. [47:48] Dave Pengelley: granola's on your on your watch now. You can record meetings with your watch. Yeah. Look it's I [47:54] Richard Webbe: And [47:55] Dave Pengelley: mean [47:55] Richard Webbe: it all contributes to [47:56] Dave Pengelley: there [47:56] Richard Webbe: that [47:56] Dave Pengelley: is [47:56] Richard Webbe: database. [47:56] Dave Pengelley: the there is the nth degree. While I was camping, I I picked up an Uh a copy of a book I've been reading on and off for years now, which is George Orwell's nineteen eighty-four. Um, and so there is a nth degree of of constant surveillance and stuff on who has the data and what they're doing with it, um, which [48:12] Richard Webbe: Well [48:12] Dave Pengelley: is [48:12] Richard Webbe: that [48:12] Dave Pengelley: w [48:12] Richard Webbe: was [48:12] Dave Pengelley: it's [48:13] Richard Webbe: the [48:13] Dave Pengelley: reasonable [48:13] Richard Webbe: article. [48:13] Dave Pengelley: to be aware of. [48:14] Richard Webbe: Well that was the article I wrote a few weeks ago, and I I had a long think about this again on one of my drives, and uh I must drive a lot of places. And I was thinking forward to a world because of AI where information is not debated. It's just there. Do you know what I mean? Facts become totally democratized. So, you know, we hear about fake news and we hear about that never happened and that didn't happen. I believe we will get to the stage where there'll be enough checks and balances, enough recording devices, enough CCTV, enough human input, enough sensors and AI, assuming that no one's corrupting the core of those of that [48:55] Dave Pengelley: I mean [48:55] Richard Webbe: information. [48:55] Dave Pengelley: that that that that's the whole thing. Win Winston in nineteen eighty four, his job is to rewrite history. [49:00] Richard Webbe: Yes, [49:00] Dave Pengelley: Like [49:00] Richard Webbe: oh [49:00] Dave Pengelley: they [49:01] Richard Webbe: yeah, [49:01] Dave Pengelley: they [49:01] Richard Webbe: I [49:01] Dave Pengelley: destroy [49:01] Richard Webbe: remember. [49:01] Dave Pengelley: the originals [49:02] Richard Webbe: Yes. [49:02] Dave Pengelley: and they modify and rewrite hist [49:04] Richard Webbe: Yeah. So [49:04] Dave Pengelley: history. [49:05] Richard Webbe: that's that [49:05] Dave Pengelley: Like [49:05] Richard Webbe: yeah, [49:05] Dave Pengelley: that [49:05] Richard Webbe: but [49:06] Dave Pengelley: that [49:06] Richard Webbe: that's the [49:06] Dave Pengelley: that's [49:06] Richard Webbe: Hollywood, yeah, that's the Hollywood conspiracy theory. I mean, today, if you go through a red light and hit a car and hurt people, you're not going to challenge it, right? Because the technology is considered reliable. I know people have tried to do that, or they try and get the the speed sensor on a speed camera recalibrated. They're minor things. I think in general, if we look towards the future. Facts will be remain facts. They won't be debatable. They'll be very easy to bring it up, excluding the theory. And then it made me start to think: so what does that leave us with? And I think it leaves us with a really great opportunity, and that is to be totally creative about improving business and the human condition, if you want. So [49:51] Dave Pengelley: So [49:51] Richard Webbe: all [49:51] Dave Pengelley: long [49:51] Richard Webbe: that [49:52] Dave Pengelley: as people [49:52] Richard Webbe: other [49:52] Dave Pengelley: aren't getting [49:52] Richard Webbe: stuff [49:52] Dave Pengelley: creative [49:52] Richard Webbe: that [49:52] Dave Pengelley: with the facts, Richard, because I mean that that is the issue, right? As AI becomes more powerful and can dive into more systems and change more things and even video. The deep fakes and stuff, you've got to be aware of the fact that there's a risk of who owns the truth. Because you know, the the the recorded the recorded uh evidence of the truth might not be the actual recorded evidence. [50:14] Richard Webbe: Totally agree. And we've seen Hollywood and books and people replicate that and even in real life. But eventually the technology will go to a certain point where certain amounts of information are undisputable. And that's what I that's why I use the example of a radar. gun or a speed camera or several witnesses, it will be corruptible to a certain point, but the more there is, the less chance you have of corrupting it. And I just like to think ahead, if you are watching, I call it the supermarket trolley theory, which is a philosopher's reference, right? If you use a supermarket trolley in a car park to get your stuff to your car, the test of humanity is will you take the supermarket trolley back? And there's two types of people. There's one that go, you know, I don't need to and I'm in a hurry, but you know what, someone else on there, so they take the trolley back. But there's a certain number of people who put their needs above everyone else goes, you know what, it doesn't matter if I leave it here and I block cars or whatever, I've got to go, I'm in a hurry. And that's, do you know what I mean? That's the moral compass. If you think you're being watched everywhere, your moral compass will change, and maybe for the good. It could be good for humanity. [51:30] Dave Pengelley: Getting more deeply philosophical and borderline political than I'm comfortable with. Uh [51:34] Richard Webbe: All right. [51:34] Dave Pengelley: not that I'm not that I'm uncomfortable with it, just we don't have time for it. Uh we've got a limited time on the show, and there's uh specific uh um wins of the week I want to discuss. And we've been talking about agents and dots and bots and things. Uh and for those that missed it, I think it's worth showing an example of what a dot could do for me, and a big win of the week for me was last week's field notes. So cadence on the channel, we're trying to do now for our regular viewers is. Mondays, I'm trying to put up an operator story. So I'm trying to do at least one interview a week and build a bit of a pipeline of interviews. Um, so that's that's happening. Anchotts interview was put up on Monday, which is fantastic. I'm got another interview booked for Friday, which we'll put up next Monday with with a uh couple of partners are building an A AI business here in Australia, which I'm looking forward to. That chat Wednesdays is our live show, and then Friday is Field Notes. And obviously, Friday with the long weekend, I went camping and had limited signal and no computer. And no fancy cameras and set up like this. But I was like, oh, I want to keep the cadence up. Like, you know, you got to keep keep shipping. So I had my my chat GPT dot, and uh, this this was the chat with it. I'll show you the actual chat. [52:46] Dave Pengelley: Comes up. I have to make this week's field note video for the AI operators podcast, but I'm way camping with limited equipment bandwidth. Not sure. Uh I'm not sure what my learn tip for the week even is. Maybe something to do with dots and bots. Maybe the way I used you to set up remotion and build animated. Video. Um, and it's like that's a solid field note. Give your assistant production tool, then direct it like it's creative, blah, blah, blah, blah. And then I was like, well, hang on, if you fetch the assets from the podcast, I'll give you clips. You can edit the field note for me on brand integrating, showing [53:12] Richard Webbe: Oh, [53:12] Dave Pengelley: the uh the thing you built for me, and so on. So I ended up doing, you know, my classic like size by agents, demo in a demo in a demo kind of thing. So while I was doing the field note about video editing, I thought, well, I'll get it to actually edit the video. [53:28] Richard Webbe: I'm not sure. [53:34] Dave Pengelley: So this lower third, cutting to the B-roll. So I literally gave it, I think, about 12 different little clips of me doing these little like talking to cameras in little like little segments. Gave it three or four clips of the kangaroos, uh, and I gave it the opening bumper and the end bumper for it as well. [53:56] Dave Pengelley: So I gave I gave it these original assets. [54:00] Dave Pengelley: Um [54:18] Dave Pengelley: that that was my direction of the bot. There, I didn't later on say, when you edit this, make sure you put yourself in. I just said that in the transcript and he interpreted that and put his own little gif in and stuff. So pretty wild that The hardest thing was actually uploading the footage because in AI dots, you can't upload video files. So I had to go into the new Chat GPT spaces and create a folder in spaces and then work out how to upload my video files into the space. And it was a whole thing. The hardest bit was actually uploading the footage. But once I gave it the 12 video files, um, the first version cut too much of my thing and tried to be really condensed. I was like, dude, I gave you five minutes of footage. Why is there a two-minute video here? I was like, oh, sorry, I got too excited and gave me the extended version. But that was that was huge. That I gave it the video direction, the editing direction in the actual transcript, and it fully just nailed it with the animated graphics and the cuts to the kangaroos and the b-roll and all of that kind of stuff. Awesome. [55:14] Richard Webbe: I would [55:14] Matt: That's [55:14] Richard Webbe: I [55:14] Matt: pretty [55:15] Richard Webbe: that's very [55:15] Matt: cool. [55:15] Richard Webbe: impressive, Dive. [55:18] Dave Pengelley: So [55:18] Matt: I [55:18] Dave Pengelley: um [55:18] Matt: um I I I yeah, that's that's pretty awesome. I just my brain's still on Richard's conversation. I I would love a podcast where we could just go down weird rabbit holes. I think that'd be great. [55:31] Dave Pengelley: Maybe [55:31] Richard Webbe: That's [55:31] Dave Pengelley: maybe [55:31] Richard Webbe: here. [55:31] Dave Pengelley: maybe we need to do like a Friday night um Scotch cigars and uh philosophy [55:36] Richard Webbe: The Rabbit Hole [55:37] Dave Pengelley: sort [55:37] Richard Webbe: podcast, [55:37] Dave Pengelley: of [55:38] Richard Webbe: I like it. [55:39] Dave Pengelley: yes, yes. Um, but you know, we we talked about also with the these uh these dot bots and so on, how the moment they're basically the tokenomics is unlimited. Uh and so, you know, and I've been writing a game and watching a lot of uh YouTube content about game design and stuff. And one of the themes that comes up is sometimes games do have a message political. Or otherwise, that is a bit of a meta thing. It's not like directly in the game, but the game's throws ideas that are important to the author. And so I got thinking about, you know, all this job interviews and stuff. And so as a background thing, I've had Hamilton building a hearthstone style game, like a career RPG about interviewing, where you've got to prove your impact. So I'm going to like play my spreadsheet modeling card so that it knows that I can I can model spreadsheets. And if you can't answer the questions, it impacts your confidence. Like this is very Very super early alpha. But and you know, it's it's a bit too hard, sturdy, a bit too much, you know, um like weird, warping, uh like it needs to be more office-based, the the graphics and stuff. But you know, this is this is my dot bot has built this thing as a background task, is part of its included quota, and published up this whole sort of game and it's multiple iterations and feedback and stuff, but just it's pretty wild that and like you say, Richard, the average. Thing. Some of this is pretty average. Like it's very generic. The AI has done way too much of the heavy lifting. I've got to go through and actually keep working with it, trim it down, reduce it, make it a bit more human, make it a bit more sensible. Because it's just put lots of ideas in there that don't all make sense. But you know, it's something that these dots can do for you as a thing. It's that and and for this one, I was like, do not use codecs. Do this on your own computer. Go build this. Um, so yeah, like there's lots of stuff you can do. And like you said, Richard, creativity. Like all this sort of stuff, if you use it well, unlocks your creativity. If you can be creative with what matters to you, whether that's, you know, editing up videos, whether it's making games, whether it's building business software or tools for your team, or learning education, lots of people build like learning apps for their kids to help them learn maths or whatever it is. These things are there and available. You've just got to find the problem you need solved. [57:53] Richard Webbe: And most of my creativity comes to me not when I'm sitting at the computer. That's why I like turning [57:58] Dave Pengelley: Well, [57:58] Richard Webbe: on the [57:59] Dave Pengelley: I [57:59] Richard Webbe: voice. [57:59] Dave Pengelley: yeah, I I go My daily walks, like I talk about using voice mode a lot. It's generally when I'm out doing my 5k walk every day. [58:04] Richard Webbe: Yeah, or bike [58:05] Matt: So [58:05] Richard Webbe: ride or [58:05] Matt: you can [58:06] Richard Webbe: drive. Go on, [58:07] Matt: ring [58:07] Richard Webbe: Matt. [58:07] Matt: your dot. Have you managed [58:09] Dave Pengelley: I [58:09] Matt: to [58:09] Dave Pengelley: c [58:09] Matt: have a phone [58:10] Dave Pengelley: I can't. [58:10] Matt: call? [58:10] Dave Pengelley: It never never picks up. I tried. It never picks up. [58:14] Matt: Wow. Like you had a good you sat like with it ringing for a long time. [58:18] Dave Pengelley: Yeah. Yeah, just rings and rings and rings. Like, [58:20] Richard Webbe: Matt [58:20] Dave Pengelley: ah, [58:21] Richard Webbe: know [58:21] Dave Pengelley: forget [58:21] Richard Webbe: rang his [58:21] Dave Pengelley: this. [58:21] Richard Webbe: bot the other day and he joined our conference and contributed, and I was very impressed. [58:27] Matt: Yeah, I was I just quickly I joined [58:36] Dave Pengelley: Ha wild. Wild. So um yeah, lots lots going on. Um, Chat GPT, they got, you know, we discussed that a lot of questionable feedback after their dev day last week, and so they've come out and said we're gonna ship something new every day for 28 days that's helpful or an improvement, or give you a reset. I have a feeling we're not gonna see any resets. I think that's I've got enough stuff in their 28-day back catalogue, as much as I would like to see a bunch of resets. They wouldn't have done that. Unless I've strategically scheduled a couple resets in on you know the the Saturday or Sunday when most people are going to be resetting their seven days anyway. Um, seems to be the way. [59:15] Matt: Yeah, [59:15] Richard Webbe: I [59:16] Matt: uh [59:16] Richard Webbe: had a couple of other news things I wanted to raise with you two to get your opinion on, if that's okay. [59:30] Richard Webbe: Great segue, you're very good at [59:32] Dave Pengelley: We're [59:32] Richard Webbe: that [59:32] Dave Pengelley: we're [59:32] Richard Webbe: now. [59:32] Dave Pengelley: on the hour, so we we will have to make a quick [59:35] Richard Webbe: Just [59:35] Dave Pengelley: news. [59:35] Richard Webbe: very quickly, SpaceX has committed another $40 billion to buying chips from NVIDIA. And they see the key to the future is infrastructure. I wrote my data center article on the same thing, but I put more of a we need more power no matter what in the background, whatever that means for this country, because I know we're not doing well with our uh power story. And of course, Anthropic has opened up a lot of its little secrets. Secrets to a number of other organizations to accelerate cybersecurity and protection in the AI world and making uses of AI resources because certainly the bad guys are making use of AI resources. So we need more help on the other side. So I thought that was a good move by Anthropic and anyone who's done SQL cyber injection. They identified 129,000 verified software vulnerabilities, including more than 33,000 critical ones. Uh and that's dangerous when agents can find those in a microsecond. [01:00:37] Matt: So was this [01:00:38] Dave Pengelley: Yeah. [01:00:38] Matt: something that was new? Like I feel like that's what the whole mythos and glass wing and everything like that was meant [01:00:43] Dave Pengelley: And [01:00:43] Matt: to be [01:00:43] Dave Pengelley: and [01:00:43] Matt: about. [01:00:43] Dave Pengelley: OpenAI is I think they've released Astra 6.1 for the project daybreak or whatever, which is their version of that. They've they've given it to selected agencies up front for security and pen testing. [01:00:55] Matt: Yeah, so Daybreak Blue and Daybreak Red both were those ones that you could get access to. to they must have newer versions now. [01:01:02] Richard Webbe: Yeah. [01:01:02] Matt: Um yeah, I think a lot of that stuff, especially if it's a news article related, a lot of it is very much um it's positioning. It's it's it doesn't sound like it, but it's political positioning and it's about VC funding. [01:01:19] Richard Webbe: Yeah. Oh yeah, no, I I get that. Yeah, where's the money coming from and who's going to make money out of it? As we know, there's a lot of um how would I say, social justice, political posture. In the environment and in the media at the moment because only a few people are making a lot of money out of AI, and that's not making many people very happy. [01:01:40] Matt: Hmm. [01:01:42] Matt: And bit like they talk about like the bursting of the bubble, like when the VC money dries up and all that sort of thing. Um, there was a lot of talk about Dario pushing for the slowdown and the regulations because if the more regulated it is, the more funding they get. [01:01:57] Dave Pengelley: Yeah. [01:02:00] Richard Webbe: So years ago, when Amazon was just a bookseller and the funding side of things, people were talking about the only reason Amazon's still selling books is it's making money out of all the people investing its shares in it, right? But Amazon came up with a rule, or Vizos came up with a rule, no one can discuss the company's financials outside of the company. And so that conversation died down, so they kept their investment coming in. And now, of course, Amazon is very profitable and very successful. Successful and still getting lots of uh investment. And they sort of rode that chasm. And I see a lot of AI businesses will do the same thing. We can see we're going to make money down here. Don't worry about us kicking all this money in here. We'll get there. We'll cross that chasm. [01:02:45] Dave Pengelley: It's it's al always always back to who's getting paid. I mean, James Cameron fully backed the the big um game changers documentary documentary about the benefits of vegetarian diets over eating meat shortly after he invested in a giant. Giant pea protein fractionation business. So he was somewhat motivated to get people off meat and onto plant-based proteins. [01:03:08] Richard Webbe: Yes, there's a fair bit of that going on. Building the lily for the people who are going to benefit. [01:03:13] Dave Pengelley: Yes. The uh the the yeah, there's some there was some genuinely funny moments in that in that doc though though. It was it was a bit funny. Uh we won't we won't go into such things on the this stream. It's a it's a family show. Uh all right. Well, thank you, gentlemen. Uh we are we are at the hour. We will keep it there, wrap it tight. For our audience. Um, good chats, always more rabbit holes we could dive down, but limited time. Hopefully, uh, viewers, you learned something today. I learned something. We all learned something from each other. Please drop your chats in, drop your questions and comments in. If you're watching this on the replay, like, subscribe, tell your friends, and we'll see you next week. See, boys. [01:03:52] Matt: Yeah.