Your AI Subscription Is Already Too Expensive. Here's What Comes Next.

[00:00] Dave Pengelley: There we go. Live on the interwebs. [00:05] Matt Slager: Hello. [00:07] Dave Pengelley: Hello, interwebs. [00:08] Matt Slager: Good morning, good evening, good night. [00:11] Dave Pengelley: Now often, often we jump on and we chat for five or 10 minutes before we actually hit live, but I was running late today, so we've actually done no banter. This is actually the legit pre-show during the show. [00:21] Matt Slager: Can you tell that I need a haircut yet? [00:23] Richard Webbe: I was about to say, Matt, I was about to say your beard's getting bigger [00:27] Matt Slager: Yeah, I've, um, I've been strapped for time. Freaking agents are flogging me. [00:31] Dave Pengelley: I th- I, I thought he was just happy to see you, Richard. [00:37] Matt Slager: I need to move you guys up here too so I'm not looking down weirdly. [00:40] Richard Webbe: Now we're, we're... I, I found an article I wanna bounce off you guys that you might be more attuned to this, Dave, that when they did their trip to China, Zuckerberg and Elon Musk and all of them, uh, apparently Musk and Zuckerberg convinced Donald Trump to not give his AI mandate order.[00:01:00] [01:00] Richard Webbe: And I don't even know what the mandate order was. [01:03] Dave Pengelley: I'm not across that specific news article. [01:06] Richard Webbe: Hmm. [01:06] Matt Slager: I'm not sure either. W- what was, what were they saying? That, that the Meta AI needed to do things for, for Trump or something? [01:15] Richard Webbe: So the, the headline was, right, and I... It's me, I was confused, so I thought you guys might know. [01:21] Richard Webbe: Zuckerberg and, and Musk convinced Trump to scrap AI executive order. Now, I don't, I can't even find out what that executive order was. [01:31] Matt Slager: Yeah, right. [01:32] Richard Webbe: I even use Google. Well, if [01:32] Matt Slager: it's scrapped, then I guess it's not a problem, but, um- [01:35] Dave Pengelley: Yeah, I mean, it was probably some kind of data sovereignty models, remote data center, China. [01:41] Dave Pengelley: China. China. [01:43] Richard Webbe: China. [01:44] Matt Slager: China [01:44] Richard Webbe: virus. China [01:45] Matt Slager: virus. [01:46] Richard Webbe: China. Yes. [01:49] Dave Pengelley: Uh, yeah, I, I don't know. Um, 'cause I, I know and I, and I, and, um, I saw the Bill Gates documentary on Netflix a few years back, uh, when Trump was in office the first time, [00:02:00] and I know he stopped Bill from doing some of the nuclear power generation research with Chinese companies. And so Bill was experimenting with different ways of, like, liquid salt cooling and all that kind of underground nuclear things, you know, next gen safer nuclear. [02:18] Dave Pengelley: Um, but he couldn't get the funding to do it properly in America, and he was gonna run off and do it in China, and then, you know, Bill was very, very upset with Mr. Trump for not allowing him to go and proliferate nuclear, uh, energy generation in China. Yeah. Now, given what I've heard, um, during the election campaign around Trump's views on generalized nuclear proliferation and his fears of more nuclear weapons, maybe that played into it. [02:46] Dave Pengelley: Plus also the China threat and him worrying about them getting cheap, free power and all kinds of bits and pieces, like, there's a lot of stuff going on there I'm sure that we will never know about. But- [02:54] Richard Webbe: Sure. [02:55] Dave Pengelley: Yep ... um, it doesn't surprise me that Trump would take a issue with next generation advanced [00:03:00] technologies leaving US borders to go live in Chinese data centers [03:04] Richard Webbe: Hmm That's, uh, and that's the thing that, um, Reagan was talking about years ago, wasn't it? [03:09] Richard Webbe: I, I heard about that this morning. Um, and that, uh, you know, Japan did it with America and it backfired on them. Uh, but, uh, the US and China may be doing that together. I'm looking at an article that says AI regulations that were gonna be federally run and not allowed to be run state by state for consistency moving forward. [03:33] Richard Webbe: A national policy framework- Gotcha ... executive order 14365. [03:39] Dave Pengelley: Okay. [03:40] Richard Webbe: I wonder, the Uniform National Framework, of course this would stop, um, I think, you know, politically aligned states in the US going against the federal administration for political reasons rather than a united front when they go to negotiate with China and Russia and other people and that- Hmm [03:56] Richard Webbe: on what AI uniformity is. [03:59] Dave Pengelley: Yeah. I, [00:04:00] I don't know enough to, to comment deeply on the specific, specifics of that. Oh, by the way, but th- but thank- thanks for the spoilers, Richard. We finally finished Person of Interest, and for the last two and a half seasons I've been thinking your, your comments about how it ends, and that's just been ringing in my head. [04:14] Dave Pengelley: So I've like, the whole like last 30 episodes plus I haven't been able to properly enjoy 'cause I'm like, "Where is this going? Oh, he told me it's not... Oh, man." So anyway, I'm- I can't [04:23] Richard Webbe: remember how it ends ... [04:24] Dave Pengelley: remember you said not well. That was enough. [04:30] Richard Webbe: I say that about all- Just- ... all the series that I liked and I don't just- ... [04:34] Dave Pengelley: poisoned apple'd the whole, whole, uh, whole end of it. But, um- [04:37] Richard Webbe: Well ... [04:38] Dave Pengelley: yeah, but good show. Great show. Again, like the way it was talking about AI and its ability to do things and track and monitoring movements and pattern recognition, like the fact they spawned that show 15 years ago, um- Yeah [04:52] Dave Pengelley: and they were talking about some of that advanced intelligence stuff, pretty crazy. [04:57] Richard Webbe: Yeah, very true. Very true. [00:05:00] Well, Hal, close the bay door please, Hal. [05:03] Dave Pengelley: I can't do that, Dave. [05:06] Matt Slager: You know what? I've actually never seen Space Odyssey. [05:09] Richard Webbe: Yeah. You should see it. You'd like it. It'd be a bit frustrating for you though, Matt, 'cause your brain would be down the end of the movie. [05:15] Dave Pengelley: Let, let, let me know if you can hear this. [05:18] HAL 9000: I'm sorry, Dave. Yes. [05:19] Richard Webbe: I'm [05:19] HAL 9000: afraid I can't do that. [05:21] Dave Pengelley: Yeah, I got it, got it on my soundboard. There we go. Oh, really? That's cool. Yep. [05:27] Richard Webbe: The end of that film, there's a couple of scenes where he goes from room to room to room seeing an older version of himself doing different things in each room, and I read the article the other day that explained what that meant. [05:37] Richard Webbe: I won't read it out. You can look it up, but I've forgotten. Mm. But yes, it was very- That was very- ... quite advanced. [05:43] Matt Slager: Mm. [05:43] Dave Pengelley: Yeah. Yep. Um- Speaking of advanced. [05:49] Matt Slager: Sp- did you say speaking of advanced? Yeah. I, I had put together this super advanced, uh, workflow. So, like, I had my workstation, you know, my cockpit [00:06:00] ready, and I was going hard. [06:03] Matt Slager: I literally, like, after Codex did a, a quota reset, you know, I, I think that was about Friday or maybe earlier in the week, I can't remember now. It's been a blur. Anyway, they did a full reset, and I'm on the, the, the max, max Codex plan right now, and I obliterated that, that weekly reset. Uh, 50% of it was just gone in a day. [06:24] Matt Slager: Uh- [06:24] Dave Pengelley: How is that possible with the five-hour chunks? Like, that, that seems insane that with the five-hour limits you can still get through so much weekly quota so quickly. [06:32] Matt Slager: Zero clue. Yeah. I'm assuming the five-hour is, like, more than 50% of the week or something if you do that. Yeah, it, it, it just went heavy, and then all of a sudden, um, the, the, the... [06:46] Matt Slager: Using, like, Codex through OpenAI as a provider for these agent harnesses- [06:50] Dave Pengelley: Yeah, yeah ... [06:51] Matt Slager: to use the quota, it all of a sudden just stopped working. Like, you, you were talking about some issues with Hermes before, and- Yeah ... I was curious, like, if that was [00:07:00] the same error. [07:00] Dave Pengelley: Yeah. It's- Uh, I... Weird. Like, like, trying to... [07:06] Dave Pengelley: So I, uh, for the show, for the listeners, we'll, we'll get onto that as part of Wins of the Week, I suppose, but, um, yeah, I think the way the APIs respond and when you hit usage limits and what these third-party things do when they, when they're rate limited, it gets really funky in, in how they manage their fallback providers and, and all this kind of stuff I've been on struggle street a little bit with. [07:27] Dave Pengelley: Um- Yeah ... just, yeah. I think, yeah, it's one of those things where, uh... And so it g- gets you to thinking if, if people are running these in business critical environments, it becomes really essential to have that multi-model fallback plan. Um, but, uh, sound- sounds like we're talking about AI, so let's talk about AI. [07:56] Dave Pengelley: I like to feel fancy like we're on a real television show. [07:58] Matt Slager: Yeah. [00:08:00] So, like, the thing that I was experiencing, like, you know when you open a website or something and it just sits there spinning, you know, like a- Yeah ... like an infinite loading? That was literally happening in my agent harnesses. So using Pie, using Homies, using Droid, using OpenCode, any of those- It wasn't even telling you [08:15] Dave Pengelley: that it was r- rate limited, [08:16] Matt Slager: it just was hanging [08:17] Matt Slager: zero error. It was just sitting there [08:19] Richard Webbe: spinning. It wasn't that fault you had last, wasn't that fault you had last week with something down. [08:23] Matt Slager: Nah. No, this is different. And made zero sense, and I completely table flipped because, like, I was working the hardest and fastest that I'd ever been because I had these huge deadlines coming up. [08:37] Dave Pengelley: Yeah. [08:37] Matt Slager: And then all of a sudden everything stopped. So I literally had to pivot and move my entire workflow into a different series of harnesses and apps because this sort of seems like tinfoil hat, like, a, a little bit of a conspiracy theorist moment, but I had to use Codex. I had to use the Codex CLI and the Codex desktop app. [09:00] Matt Slager: Nothing else was [09:00] Matt Slager: using that subscription inference properly. [09:02] Dave Pengelley: Wow. [09:03] Matt Slager: That's weird. And, yeah, what a journey that was, and now I actually really like the Codex desktop app, so. [09:08] Dave Pengelley: It's pretty good, isn't it? Yeah. Not [09:09] Matt Slager: bad. It works great. And I, I... Theo had a really good video that I watched this morning as well, where he was talking about Claude Code's easily the one for just people that don't really know what they're doing to get into. [09:22] Matt Slager: It's the one that's the most gamified, the most- Yeah ... the most instantly satisfying It, it just, it's fun to use, you know? And you can still get good work done in it. Yeah. And then Codex is the one that is for people that wanna actually do work. You know, it's, it's the one that's not fancy, it's not shiny, it doesn't have gamification. [09:42] Matt Slager: It's just, "Oh, you want that fixed? Okay, there you go. It's fixed." [09:44] Dave Pengelley: And, and it's got the multimodal stuff. It can do image gen built in. [09:47] Matt Slager: Yeah. [09:48] Dave Pengelley: It's just like it's complete. [09:50] Matt Slager: Um- As far as, um, end-to-end closed loop, uh, UI testing as well, so like, just spin up... You know, I, I always have my dev servers running, and the Codex can just [00:10:00] open its own web browser now- Yeah [10:01] Matt Slager: and click around. And like I've been using Playwright for a really long time, but this just works better. And- [10:07] Dave Pengelley: Yeah, it's just 'cause it's integrated into the, into the development environment harness and- Yep ... um, yeah, I was building... I was playing around with a game. I've, I've been pausing on it for a few weeks 'cause I'm using tokens elsewhere, but I was just experimenting building a game using Godot. [10:20] Dave Pengelley: Um, and that was awesome 'cause it would actually then, you know, using the computer use, just fire up and start play testing there, and I found that... And, and it could generate, um, visual assets. So I started in Claude Code with it, but then I was like, "I, I need some visual assets in this, this game." And so I moved to Codex, and next thing you know, it's generating assets and pumping them straight into the Godot engine and while it's play testing, and I was like, "This is awesome." [10:41] Dave Pengelley: Like- [10:41] Matt Slager: That's cool ... [10:42] Dave Pengelley: um, so yeah, the, the, the, the out of the box harnesses are really good for doing some of those jobs. But for the general agentic life thing, some of the new harnesses like Hermes are also pretty fun and exciting, and that's where I've been diving into, [00:11:00] when they work with the APIs and so on. [11:03] Dave Pengelley: So poor Richard- Yeah ... sitting there going, "What are, what are these two nerds talking about?" [11:08] Matt Slager: Have you got your Hermes agent running, Richard? [11:11] Richard Webbe: No. I thought I had Hermes, but I went and saw the doctor, he said it's okay. Oh, very funny. [11:17] Dave Pengelley: They g- gave you a cream for it. Um, the, uh, I mean, you, you had... Did you get your cowork working after all that, with all the administrator permission issues we were having on your laptop? [11:27] Richard Webbe: Yeah, sort of. I've gotta go back to it, but I've still got other problems with my laptop like yours. So I'll report on that next week. I was having a bit of a play, but I've had a bit more success. I've had a little bit of a win there, so yes. [11:38] Dave Pengelley: Yeah, because, because, uh, Ma- Richard had got multiple user logins on his, on his single user laptop for reasons that are beyond the scope of this show. [12:00] Dave Pengelley: Um, but cowork would work in one of them but not the other, and I think it's because one of them had admin rights and the other one did not. Yeah. And so when we enabled your, your main profile that you actually [12:00] Dave Pengelley: like as admin, then it worked and stopped giving you that update error? [12:03] Richard Webbe: Yeah. Yeah. [12:04] Dave Pengelley: Brilliant. [12:05] Dave Pengelley: Brilliant. So we were going around in circles, going and said, like, "Run the update to use cowork," and you go, "You're fully updated. You're working in the other profile. What are you doing?" "No, no, close and re-update it." Like- And I think it's 'cause Cowork wants to do stuff with the operating system and wants to move files and wants to have access to your desktop and do computer use or whatever, so, um, it won't run if you're not in admin profile, which I thought was interesting on Windows. [12:27] Dave Pengelley: Yeah. [12:28] Richard Webbe: Yep. [12:28] Matt Slager: That makes sense. It's sort of like in the land of, um, Linux operating systems, it's kind of just like the super user do, you know, the old pseudo. Um, so yeah, you kind of need to have admin rights to do some stuff depending on- Yeah ... like how you do it. It's [12:45] Dave Pengelley: not obvious and it doesn't tell you that. [12:47] Dave Pengelley: I was like old school debugging just trying to work out why is this working in one of Richard's profiles but not the other. [12:52] Richard Webbe: Yeah. [12:53] Dave Pengelley: Um- [12:54] Richard Webbe: Very frustrating. [12:55] Dave Pengelley: Yeah. Yeah. But, uh, I mean, Co- Cowork does some really good [00:13:00] stuff and like, uh, like as you guys know, I've been running like hard in Cowork Hana for a long time now, like a couple of months. [13:06] Dave Pengelley: That's a long time in this space, um, to doing it. [13:09] Matt Slager: It's a very long time. [13:11] Dave Pengelley: It is, right? Like, it's crazy. When I started trying to do my Valid Agenda business thing back in October, I was... Last October, I was all gung-ho for just teaching people how to use ChatGPT and do ChatGPT coaching. [13:24] HAL 9000: Hmm. [13:25] Dave Pengelley: That, that was where I started, and I hadn't touched an IDE at that point yet. [14:00] Dave Pengelley: I was still generating some, some little programs for doing some trading stuff, and I'd been copying and pasting that out of ChatGPT browser sessions like, um, to where we are now, where, you know, built and deployed full SaaS app with APIs and MCP servers and everything else. It's just wild that in like eight, six to eight months how quickly and how fast this landscape has changed to the point that I've gone from that to Antigravity to manual workflows in Antigravity, to [14:00] Dave Pengelley: automated workflows in Cowork, to now moving to Hermes, um, with automated project boards and stuff. [14:08] Richard Webbe: I had something very weird happen 'cause, you know, listening to you guys and I got into the Cowork, and then I suddenly looked like that. Again. I'm not used to looking like that. [14:19] Dave Pengelley: I don't, I don't, I don't think Cowork can generate that picture, can it? That, that looks like a ChatGPT. No. [14:26] Richard Webbe: I looked in the mirror after playing all the technology you told me to do, and I came out a hippie. [14:30] Dave Pengelley: Yeah, fair enough. Um, free- [14:32] Richard Webbe: It [14:32] Matt Slager: happens. Happens to the best of us So with, with your Hermis, I, I'm super curious, I'm sure other people are too, like how, how are you running it? Like, are you using their, their live sort of web server portal? Or are you just doing it through Discord or Telegram or terminal? [14:50] Matt Slager: Like how are you managing and executing it? Yes. [14:52] Dave Pengelley: The short answer is yes. [14:54] Matt Slager: All of it. [14:55] Dave Pengelley: All of the things. Um, as I, I go to the terminal sometimes for some [00:15:00] debugging stuff like today where I've been trying to debug issues and things and, um, I do a little bit in the terminal 'cause I'm, I'm running it on my Windows PC just in WSL, so it's just- Yeah [15:10] Dave Pengelley: just running locally. I'm not running a hosted VPS or anything, so it's just on my Windows machine. Yep. Um, so I've got that and I'm primarily trying to use it through Discord. So I've, uh- Yeah ... set up a whole hierarchy of Discord channels, and so I'm trying to have conversations in the right Discord channel. [15:27] Dave Pengelley: I've arguably got too many channels, like it, it proliferated and built me like 500 channels and I probably need five- Right. [15:35] Matt Slager: Mm ... [15:35] Dave Pengelley: given the way it works with threads and stuff. So anyway, I'm doing that. I've got Telegram, uh, and I can use Telegram if I want. I get my gateway restart notifications on Telegram. [15:46] Dave Pengelley: Um, but primarily Discord and the Kanban itself, just trying to like drop stuff in the triage queue of the Kanban. [15:54] Matt Slager: Yeah. [15:55] Dave Pengelley: And then let it, it read through that and allocate that. Um, to the point that, you know, 'cause I've got Telscale [00:16:00] set up, I got it to set the dashboard, web dashboard up So that way I can access it from my mobile anywhere. [16:09] Dave Pengelley: So I've got, I've got saved a Hermes icon to my iPhone home screen. Yeah. And I tap that, and it brings up the Kanban board for me, so I can actually access my Hermes Kanban anywhere, even though it's running on my home laptop, desktop. [16:20] Matt Slager: Yeah. Nice. That's awesome. Yeah. I, um, I was gonna ask that, if you'd exposed that so you could run it from your phone and, um, yeah, that's cool. [17:00] Matt Slager: Uh, I haven't really done too much with it. I had a quick play, and I think it would work really, really well with some of my workflows. Um, I'm definitely using Hermes as a- Mm ... just another layer in between. Um, I've experimented with some of my individual projects just using Hermes and pumping 5.5 through it, and it's very interesting 'cause it's a, um, it's kind of like a do-everything agent, and its sort of awareness of, uh, skills and, you know, being able to curate and create its own skills is [17:00] Matt Slager: interesting, but when it determines one being stale that actually is relevant and you just haven't used it for a while, that is potentially, like, a workflow problem. [17:11] Matt Slager: Um, and I experienced that 'cause all of a sudden it started doing weird stuff. I'm like, "Come on, man. You should know this." [17:16] Dave Pengelley: But it's... Uh, uh, I know I've got that concern too with its curated memory. I mean, it specifically only keeps, like, a couple thousand characters of memory- Mm ... so that way it stays tight and efficient. [17:26] Dave Pengelley: But I'm running it across, you know, four or five different business entities and projects and things, and I'm like, how can that small, tiny memory know all the things it needs to know about all the different things I'm doing? Yeah. Um, and after watching a NetworkChuck video, I've linked it up to Honcho, so I think it's getting some backup support from Honcho for, for memory- [17:45] Matt Slager: Yep [17:45] Dave Pengelley: management. [17:46] Matt Slager: You should be able to see the Honcho traces and- [17:48] Dave Pengelley: Yeah ... see, [17:49] Matt Slager: like, uh, an example input payload and, and how it's, um- [17:52] Dave Pengelley: I can go to the Honcho website and it shows me, you know, 2,000 memories logged and stuff and, and how it's inferring. [17:58] Matt Slager: Yeah. [17:58] Dave Pengelley: Cool. Yeah. But the re- the reason, the [00:18:00] reason I- I'm, I'm loving it is, as you guys might remember, I was trying to build my own harness, um, which, you know, has now... [18:10] Dave Pengelley: Ah, do this one. Do this one? is winding down. I've killed it, um, in favor, and I'm sending people to Hermes instead. Um, I fully, on LinkedIn, posted a video of me setting a piece of paper on fire to indicate this, and I still could barely get any impressions and views on LinkedIn. Like, I'm freaking doing a video setting something on fire, and I still can't get likes. [18:34] Dave Pengelley: What's, what's going on with this world? [18:35] Matt Slager: Probably get banned for, like, arsoning or something. [18:37] Dave Pengelley: Oh. I mean, I think the problem is because I said I'm winding down sergeant.com, and it goes, "Oh, you put a URL in there." I'm like, "I'm not trying to divert people there," but as soon as you put a URL in, LinkedIn goes, "Ngh, get down." [18:50] Matt Slager: Yep. " [18:50] Dave Pengelley: Keep people off my feed." [18:51] Matt Slager: I've got that, I think, as well. I've got some auto moderation in one of my client apps where there's actually a community feature. He wanted to merge... He had a circle community, [00:19:00] so he wanted to merge that functionality into this other app as well, where his clients do things. [19:04] Matt Slager: Anyway, um, I have auto moderation in there for the community feed, and when the business owner posts a message, it gets flagged for self-promotion. [19:14] Dave Pengelley: Hmm. [19:15] Matt Slager: Which is kind of funny. Like- Yep ... it's his own thing. Stop it. [19:19] Dave Pengelley: It's an edge case you gotta, gotta allow for. But, uh, yeah, my, my, my platform that I was building was this concept of having, um, the business in a box and having different worker profiles that were actually hierarchical talking to each other, so you could have compartmentalization. [19:35] Dave Pengelley: So when you talk about, you know, token burn, rather than have mega, mega prompts for everything, it would break the work down into little chunks and then feed that out to sub-workers who were very compartmentalized and specialized. They could keep their own memories so they could try and iterate on their business unit to make it better. [19:50] Dave Pengelley: It was kind of trying to take some of the OpenClaw heartbeat aspects and that self-iteration stuff, but build it into this sort of structured business in a box thing. [19:58] Matt Slager: Yep. [19:58] Dave Pengelley: Um, [00:20:00] building a harness is really, really hard, and I think I didn't appreciate how much of the harnesses is actually deterministic code- [20:07] Matt Slager: Yeah [20:08] Dave Pengelley: where it's not actually the LLMs making decisions at all. It actually really is just a program- Yeah, and then like- ... making decisions old school-wise, just from semantic text- Layers upon layers as well ... analysis and stuff. Um, and when you look at the way Hermes does it, the first thing Hermes does on half the stuff you do is it references a skill about the Hermes agent to learn about itself, and all the- Yeah [20:26] Dave Pengelley: all these little things you don't realize when you're just a dude going, "I'm gonna make my own AI harness." Like... [20:32] Matt Slager: Yeah. [20:32] Dave Pengelley: Whoops. But, uh, yeah, the, the Hermes have, they have this thing called the Kanban board, which, I mean, most people are familiar with what a Kanban board is nowadays. Um, but, you know, and, uh, this isn't really anything super sensitive from my point of view. [21:00] Dave Pengelley: This is me trying to build trading robots. Um, but I've got this sort of Kanban board now where I can drop tasks into triage, and then when it, when it goes through and does a review, it'll actually break them down and create them into [21:00] Dave Pengelley: to-do lists. Um, and then, you know, you got your scheduled, what's ready, what's in progress, blocked things. [21:06] Dave Pengelley: So I can come in here and go, "Okay, read some notes on this one. Failed to fetch. Ah, it's broken." My Kanban keeps getting corrupted. There's, there's some, there's some bugs to fix here. It, it's actually... The bug I think is with my Windows file system, and it's actually doing this read/write lock thing, and- [21:23] Richard Webbe: Mm. [21:23] Dave Pengelley: I'm trying to fix it. It's not perfect. Um, which, you know... But that's why guys like us have to experiment with it before we put it into businesses. That's right. That's where we come [21:31] Richard Webbe: in. [21:32] Dave Pengelley: Um, but yeah, so, so these, you-- I can click on these like a linear task. I can add comments, I can unblock it, and then it'll go back into in progress and stuff. [22:00] Dave Pengelley: But this ability to have these things, and I've got... You can see, like, I've got FX, OS, Ray, Harry, Perry, Felicity. So I've created different agent profiles, which is all my tr- my, my, you know, different personas that have different specialty skills. So it assigns the tasks to the right persona, um, that can all be working on these things, and it'll break tasks up and assign [22:00] Dave Pengelley: worker to other workers. [22:01] Dave Pengelley: And so it doesn't have that, that full hierarchical thing like I was building in Sergeant, but it's, you know, got 80% of the same functionality, plus being built by a proper research lab, sophisticated harness. [22:13] Matt Slager: Mm. [22:13] Dave Pengelley: So that's why I'm like, "I'm not gonna burn more cycles trying to build my own thing." Um, this is really good, and it's gonna get better. [22:21] Matt Slager: You could, uh... I don't wanna, like, get you to pull the ashes out of the bin, but you could literally wire Hermes into that- [22:30] Dave Pengelley: Nah. Nah, not doing it. Nah. I, um, I, I put a pro- on LinkedIn and, and a, and a friend sort of said, "If only you could fork it, fork Hermes and do something." I was like, "No." I'm like... He's like, uh, he's like, "But you would, you could get agents to do it." [23:00] Dave Pengelley: I'm like, "No, I don't have the tokens. I'm not, not focusing on that." It's, it's, it's an interesting side project and it's a diversion, and it's not- It's not gonna make me any money, and so I can't afford to, to spend money on building it at this point in time when, you know, we've got research-backed funded teams [23:00] Dave Pengelley: that can-- that are building this sort of stuff, and they're doing it really, really well. [23:02] Dave Pengelley: But- [23:03] Richard Webbe: Yeah ... [23:03] Dave Pengelley: um, despite some issues with my own personal setup as far as, you know, hard drive write issues and, and locking my database and corrupting it if I'm not careful, and I've just put through a fix to try and remove some of the journaling functionality, which should solve that. Um, but you know, I'm doing a live demo, so of course it was gonna break. [23:23] Matt Slager: Well, it's like back in the day when everyone was frothing on OpenClaw and I, I built my cloud version of a railway template that was fully secure. You know, like it's- Yeah ... it was super, super locked down, and then I had various layers of bubble wrap that you could remove. I remember that. I could, like, ascend that security tier, and even that, that was so interesting because by nature of me locking it down, it literally was deactivating and breaking certain features. [23:52] Matt Slager: So you could use it. In the, the lowest security setting, it was basically just like a personal ChatGPT, you know? [23:57] Dave Pengelley: Yeah. [23:58] Matt Slager: It was just like a, a thing that had its [00:24:00] own little file system that it could save files and read and write on that cloud container, but it couldn't really do anything. There was no tooling. [24:08] Matt Slager: There was, you know, like- [24:09] Dave Pengelley: Yeah ... [24:09] Matt Slager: no web-related exposure, um, other than your own tunnel, and then that was it. And then when you ascended those tiers, all of a sudden now you could go, "Okay, now we have the ability to put API-based tooling in it," so, you know, custom binaries. Like, I could connect it to my core service, for example. [24:27] Matt Slager: And then web searching, which obviously without web search it's kind of a bit dumb because you're just relating, relying on internal and trained knowledge. [24:36] Dave Pengelley: Mm-hmm. [24:36] Matt Slager: Um, but yeah, you're right. Like, the whole building a harness thing and, and making it work and all of the variables, it's a lot. It's a hobby. [24:46] Matt Slager: It's kinda like when people wanna get into coffee, like, you know, if you guys... I, I can't remember if I've told you, like, I used to work at a specialty coffee roastery. [24:54] Dave Pengelley: Yep. [24:54] Matt Slager: Which meant that I had to learn everything. You know, like people talk about getting barista trained. [00:25:00] Yeah. I was training baristas. Wow. [25:02] Matt Slager: I was, I was tuning their espresso machines in the morning, um, because they just couldn't pour good shots. Anyway, like coffee, you know, proper espresso is a hobby. You, you can't just dabble. You can buy a Barista Express or some Breville machine and make okay coffee, but it's not the same. [25:17] Richard Webbe: Matt, I hear it's an art. [25:19] Matt Slager: Uh, define an art. [25:22] Richard Webbe: Good point. I can't do that. Sorry, you're right. [25:25] Matt Slager: It's a hobby. It's, it's like, it's a profession. [25:28] Dave Pengelley: Yeah. There, there, there's so much to it. A- a- and it's not the, like... And, and what I was doing with Sergeant was, was getting there really well, and in some ways I was over-indexing on trying to do the user interface and make it really user-friendly compared to any of these other things where it's all terminals and entering this, and I was trying to make it a windowed experience and have really clear settings menus, and just trying to build the UX of it was actually a huge challenge to make it usable and functional and readable. [26:00] Dave Pengelley: And when I look at what the Kanban in Hermes does, where it shows me little task IDs with the parents and [26:00] Dave Pengelley: children, and I'm like, "Oh, I had all of that with my work orders." Um- Yeah ... except when I clicked on mine, I could navigate through to the parent or the child, whereas the Kanban on Hermes, it's just a s- a, a pill with no linkage. [26:11] Dave Pengelley: I'm like, "Why wouldn't you let me link and just open up the thing?" So in some ways I was, I was ahead of the game, but in other ways, trying to get the tools using and getting it to read and write on the file system and do web search and, and build that tools meta layer, really, really tricky. [26:28] Matt Slager: Yeah. Yeah, it's cool. [26:30] Matt Slager: There's lots of stuff involved. [26:31] Dave Pengelley: So yeah. So anyway, I'm, I mean, I, I hadn't done any development on Sergeant for a few weeks 'cause I've been trying to, you know, I've been focusing on Chart Reporter and AI operators, IRL events, and other bits and pieces. So came to a point where I'm like, you know what? Now that I'm seeing what Hermes is doing, I'm not gonna spend the time trying to build my own thing that does the job. [26:50] Dave Pengelley: Like it's... It was an interesting thought leadership piece and, a- and, and stuff that I was trying to put out there, and I think I can still claim a bit of that. And- [26:57] HAL 9000: Hmm ... [26:57] Dave Pengelley: I've open sourced, um, the, the Git [00:27:00] repo so people can see what I was doing with, um, with my Rust-based backend system and so on. If anyone did wanna iterate and experiment with it, that's there. [27:10] Dave Pengelley: Um, but [27:12] Matt Slager: Very [27:12] Dave Pengelley: cool ... no more. Um, [27:14] Matt Slager: yeah One, one last like note on, on Hermes in general as a, as a project. Um, I don't know if this will happen, but I was seeing some, some chatter on X about, um, SpaceX I, XAI. Mm-hmm That's such a weird word to say now. Um, the, the, the team behind Groq- [27:30] Dave Pengelley: SpaceX? [27:31] Matt Slager: Yeah. Yeah. [27:32] Dave Pengelley: That sounds like something that astronauts do in their private time. [27:35] Matt Slager: True. No, the... Yeah, so the, the team behind Groq, they were... They... You know, you guys know that they're, they're partnering with, um, Cursor at the moment. [27:43] Dave Pengelley: Yeah. [27:44] Matt Slager: So there's all of that Cursor training data that's going into the next release of Groq, which is gonna be out in probably like two or three weeks. So the current Groq model is a like half... [28:00] Matt Slager: It's like 500 billion, um, [28:00] Matt Slager: parameters. [28:00] Dave Pengelley: Mm-hmm. [28:00] Matt Slager: And it's okay. It doesn't perform, you know, that well for the upper echelon of tasks. But this new one, like their, their like pre-trained one, is already miles ahead, and it hasn't even had the Cursor data put into it yet. So I think this new one, which is gonna be a 1.5 trillion parameter model- Wow [28:20] Matt Slager: as long as it's got the right architecture to handle that, is gonna be genuinely competitive, hopefully comparative to what GPT 5.5 is now, or at least 5.4. Um, because that, you know, that will then become... You know, in Cursor they have, uh, Composer. Yep. So there's Composer 2.5 at the moment. I'm assuming Composer 3 is gonna be Groq based instead of Kimi. [28:43] Dave Pengelley: Yeah. [28:44] Matt Slager: And then... Anyway, what made me start talking about this is they are actually talking about purchasing Hermes. So imagine if the whole Groq system had Hermes actually running it, um, with the Cursor data. Yeah. [28:56] Dave Pengelley: There, there's a lot of talk around like how the Super Groq and the XO auth is working really well in [00:29:00] Hermes apparently. [29:01] Dave Pengelley: Um, I just bought $5 of, uh, like XAI API credits to try and circumvent because I'm having issues with my Codex auth at the moment, and that's, you know, that's Groq 4.3. Um- Yeah ... there's no fast model anymore. Everything costs the same amount. There's no cheap and fast sort of flash, spark fast model at the moment. [29:20] Dave Pengelley: Maybe that's because of what you're talking about, and that'll be iterated and, you know, when Groq 5 or 4.4 or whatever comes out, the current model will be deprecated to be the fast model or something. I don't know. [29:30] Matt Slager: Yeah. They should still have the, um, their previous ones. Uh, I don't think their endpoints are disconnected. [29:36] Dave Pengelley: 4, 4.0 fast, uh, apparently EOLed about two weeks ago, and it's not available. It doesn't pop up as a model- Oh ... to select anymore. [29:42] Matt Slager: How interesting. [29:43] Dave Pengelley: Mm. [29:44] Matt Slager: Yeah. I... So many different things to keep track of, like all the Chinese labs as well. Like- Yeah ... I'm a huge fan of Minimax, and Minimax 2.7 has been my just general fallback go-to for everything. [30:00] Matt Slager: And- Their M3, [30:00] Matt Slager: so 2.7, like the next one's 3, um, it's going to be in the next couple of weeks and it's gonna... They, like, they've never been state-of-the-art, like actual top. 3 is gonna be at the top, um, for sure because, oh, uh, the... I can't really talk about it 'cause I don't really know too much about- Yeah, that's fine [30:19] Matt Slager: the, the really deep machine learning stuff. But trying to understand their papers, it gives me a headache. But um, they're very confident in it and, um, it could just be marketing hype, but I'm pretty keen. [30:32] Dave Pengelley: I think, I think what's really interesting about all this sort of stuff, and, and Richard, to, to tap back into something that probably, uh, is easier for you to jump in on, um, is All this model stuff, I can see a point where, you know, whether it's Opus or, or Codex or any of these, it will switch from a purely get it from us and pay tokens to a license and run it on your own hardware at some point. [31:00] Dave Pengelley: It has to. Mm-hmm. Definitely. Because all these open source models exist, [31:00] Dave Pengelley: um, and you can, you know, put it on LM Studios and Ollamas and, and run a lot of these mo- models on your own hardware s- supposing you have powerful enough hardware. Um, there's gonna be a point where... And, and MIT just re- released a report, I'll see if we can find it, um, that sort of says that that sovereign capability to have all your own data on your own metal and run your own agents and models on your own metal is gonna be critical with the amount of explosion of agentic AI in the enterprise. [31:28] Dave Pengelley: Like, A, people don't wanna be sending it to who knows which data center with all their data in it, but B, just the cost and, and containing the cost is gonna be critical in running your own models. [31:38] Richard Webbe: Yeah. Serend- serendipitously, and I've been watching a few of the, uh, a few of the, uh, narratives and comments around, you know, uh, true general AI and when's that coming, or AGI as they call it, artificial general intelligence, and the whole idea that agents start to build up like pathways in our brain. [32:00] Richard Webbe: Like, you know, here's the language center, here's the mathematical center, here's the [32:00] Richard Webbe: control the... And that's all accelerating quite fast, and it's driving, I think, well, the, um, certainly the share price of people manufacturing the, the processes to do it on site and offload to certain algorithms or intelligence you don't need. [32:14] Richard Webbe: But when you're mining your own data and you wanna create a RAG and an, a sovereign AI environment that's protected, there's gonna be big spending in this area. As a user, I always like to come back to as a user, I've been using a few government services and health services and large enterprise services now, and I ran for a while with my friend Mike 24/7 AI, which was a BPO, you know, chatbot AI business a number of years ago, and it was so rudimentary then back then. [33:00] Richard Webbe: Now, I was on a site recently and the AI conversation that I was having was much, much better. So much more intelligent, so much more intuitive. Uh, I got through what I wanted to get done quicker. I thought I'm gonna have to hit that button and say, "Please let me talk to someone," but it [33:00] Richard Webbe: questioned me Took the answers, went to the right location, went to the right narrative and said, "That's okay, Richard, we're doing this," and on we went. [33:08] Richard Webbe: So I think growth in CPUs, obviously data centers. I think networking is starting to explode again because as you said, people are gonna bring it back in onto their own metal. That means they're gonna have to get the answers out quicker, so bandwidths on the inside, on the ingress and the egress are gonna expand. [33:27] Richard Webbe: Mm. And of course the big one, uh, well, we know in the background there's a problem with energy generation, but it seems to be as we need more energy, it goes up like this. Um, the way we make things that uses less energy is also coming down a bit, so we're hoping they'll cross out quite neatly. And the big thing, of course, is enterprise AI integration, and it's the same story it always was. [34:00] Richard Webbe: People with big software are writing nice big answers, but the customization of that for competitive advantage at the business end of a business is still gonna be left to people like us to come in and plug those little gaps, put those [34:00] Richard Webbe: little bit of mortar filler in between so it runs for the advantage of the organization. [34:05] Richard Webbe: That probably wasn't the answer you were asking for, but I just thought I'd shoot my mouth off anyway. [34:09] Dave Pengelley: No, no, no, it's good. Um, I, I wanted you to, to jump in on, on this sort of stuff. So here's, here's the report, uh, the report about the report- About the report ... that, uh, my- That my AI sent ... my AI sent me this morning- Sent me this morning [34:19] Dave Pengelley: as part of my, my news review. This came out a couple weeks ago. Um, "Sovereignty is the new operating system for agentic AI, new MIT Technology Review Insights report finds." This is a report about the report. You can click through and, you know, put your stuff in their lead magnet website if you want to get the full- Actual [34:36] Dave Pengelley: actual report. Uh, but you know, 95% of organizations plan to establish their own AI and data platforms within the next three years. [34:43] Richard Webbe: They have to. [34:44] Dave Pengelley: Um, that's, that's huge. So at that which point, like When are they gonna wanna run their own models internally, right? Um, there's an interesting quote down here, um, "Sovereignty defines which agents can touch in which region under which policies," blah, blah, blah, blah, blah.[00:35:00] [35:00] Dave Pengelley: Uh, the big concern is if you're deploying an AI infused application with a cloud-based large language model, are you losing your IP? Are you losing your competitive position? [35:07] Richard Webbe: Yes. [35:08] Richard Webbe: Yes. [35:08] Dave Pengelley: Um, yeah. Uh, sovereignty is not, uh... Where, where was it? Uh, where was the line? Wha- There was a line here specifically I wanted to read. [35:19] Dave Pengelley: So many bits and pieces in here. Um, but basically... Oh, here we go. Thales, who are, you know, a defense contractor, argues that, "Sovereignty is what unlocks the highest value AI u- use cases." So when you can trust the data is safe, when you can trust the models are safe, then you'll actually really give it things that matter from an enterprise point of view. [35:39] Dave Pengelley: If you're worried about, uh, "Can I do this? Can I say that? Will that be all right?" And you're pussyfooting around, then you're never gonna really move the needle effectively. They're saying, you know, if you can unblock these things by having those sovereignty controls in place, then you're really gonna see the biggest changes. [35:56] Richard Webbe: Yeah. Look, technology advancement and, and the sovereign data and [00:36:00] the competitive advantage is what's gonna happen, and it's always gonna leapfrog it. I always think of, uh, that old marketing curve we used to talk about in, in all products, but especially in technology, and it's a marketing S-curve, right? [36:10] Richard Webbe: So people get the bit of the technology and they like, "Oh, it's very productive, I'll use that." And it drops into what I call the trough of disillusionment, where people have got all this technology, the agents and all that, but like we're saying here, there's a couple of problems, it's not working here. But eventually it'll hit an angle where it's creating so much competitive advantage, so much benefit, it goes along. [36:30] Richard Webbe: And how long that growth on the S-curve depends on how well you implement the technology, sovereign, not sovereign, the right data, and it could last you 20 or 30 years, or it could last you three weeks. And things are moving so fast, you need to have a handle on where your data is, what data it is, is it up to date, are you secure? [37:00] Richard Webbe: 'Cause sovereign says three things. It says I'm secure, it says I can trust the data, and it says no one else... And, and it says I'm getting competitive advantage 'cause I'm doing it my way with my [37:00] Richard Webbe: experience and my expertise. So it's, it's gonna be a really interesting advantage. And these competitive advantages that people are gonna try and take from agentic AI and getting customization their way inside their organization is creating new consulting industries as well, like Syllogism. [37:17] Richard Webbe: And to which we should point out that, uh- June 23? [37:24] Dave Pengelley: Yes ... [37:25] Richard Webbe: in the evening at the Australian [37:28] Dave Pengelley: Brewery, [37:28] Richard Webbe: I know the pub ... Brewery Club in just- Rouse. Rouse ... West Sydney, Rouse Hill. We're going to be there from what time? [37:36] Dave Pengelley: 6:00. Here we go. [37:38] Richard Webbe: We'll be there from- [37:39] Dave Pengelley: Let me bring the thing up ... [37:40] Richard Webbe: oh, beaut- There we go ... come and have a beer, and we're gonna have a chat about this. [37:44] Richard Webbe: We're gonna get some people to talk about it and offer some real data behind it, and we encourage the audience to contribute and talk to us about it, aren't we? [37:53] Dave Pengelley: Yeah. Yeah, so, um, it's, it's gonna be a great night. Our first IRL event for, from the AI Operators, [00:38:00] uh, banner. So yeah, 6:00 o'clock doors open. You can't see the 6:00 o'clock on the side there. [38:05] Dave Pengelley: The barrels is there. Why isn't that showing? There we go. [38:08] Richard Webbe: And we're not charging for this one, are we, Dave? [38:09] Dave Pengelley: Uh, this one, this one's, I mean, you gotta buy your own beers. We're not, we're not, we're not paying for the drinks. [38:13] Richard Webbe: No. [38:15] Dave Pengelley: But yeah, people come. They, they can, uh, grab a beer, um, have a chat. We'll, we'll do some talks. [38:20] Dave Pengelley: Uh, we've got some local business people coming in to, to do some talks as well. And, and we're still announcing the speaker lineup. Gonna be doing that over the next few weeks. Uh, and then a bit of a Q&A panel with more time for people just to, you know, shoot the breeze like this. We've always said AI Operators is like a pub chat that we're letting people listen in on. [38:37] Dave Pengelley: Why not come along and have your own chat at the pub? Um, and so yeah, um, registrations via Luma. Uh, we've only been advertising for a week or so. We've got, uh, 14 going. There's only about 40 seats in the room, so it's filling [38:52] Richard Webbe: up. And then look who we are. Yep. Sean Barkley from Optus is coming. I know Sean and a few others. [38:57] Richard Webbe: It's good. Some heavy hitters turning up. [38:59] Dave Pengelley: Very, [00:39:00] very good. So, um, yeah, looking, looking forward to that. That's, that's gonna be fun to get out there and, and have chats in the community with people and- And the door prize is a, a free- ... not sit in an hour. Sorry? [39:12] Richard Webbe: First prize, first prize door prize is a dinner with you, Dave [39:16] Dave Pengelley: Sure [39:17] Richard Webbe: And last prize is three dinners with you, Dave. [39:20] Richard Webbe: Just kidding. Just kidding. [39:22] Dave Pengelley: Uh, yes. Um, yeah, so we're... That, that, that's gonna be good. We're look- looking, looking forward to that. But yeah, there, there's so much happening with this space, and yeah, I think on the models thing and running your own models, uh, like I can't see OpenAI and that not offering some kind of self-hosted model in the future at some point, right? [39:43] Dave Pengelley: Like, they're already doing with it through Azure and AWS, and maybe they're going, "That's good enough," because big enterprise organizations already trust Azure and AWS, and so maybe that's enough. But then I think you're still paying per token for those. You're not just paying for your hardware there. [39:58] Matt Slager: Yeah. It's token. And [00:40:00] then you also can deploy your own custom version of that on your own hardware. Like, there's a company called Base10 that I... That's an American one- Yeah ... that I have experimented with in the past, and it worked really good. One thing that I think a lot of people don't realize, though, is what, what actually is required to run some of the big models. [40:20] Matt Slager: So, like, Opus 4.7, uh, called Mythos, you know, um, as well as GPT 5.5, um, most likely this new Grok model that's coming. Like, all these high-end big models, when they say big, they, they genuinely mean it. Like, you remember the, the computers of the past, you know, the... Take up a whole room. Yeah, you'll- Yep ... that would be, that would be GPT 5.5. [40:45] Matt Slager: It would take up a whole room, if not [40:47] Dave Pengelley: more. The old, the old Cray. The Cray supercomputer. Yeah. [40:49] Richard Webbe: Oh, Dr. Cray. I met Dr. Cray many years ago. [40:52] Dave Pengelley: Really? With, with... Q- I, I love that they put couches in the computer. That was a awesome. [40:55] Richard Webbe: Yeah. The design of the Cray had the, uh, the seating around it. A little bench seat. [41:00] Richard Webbe: And you [41:00] Richard Webbe: could... To get to the cabling, you lifted up the little seat. Uh, I worked for Control Data, of course. Their, uh, original, um, uh, CTO was Dr. Cray, and then he left Control Data to form Cray Computers. So I was in the supercomputer game, and it was rather amazing, uh, some of the things they did there. I like what you're saying there, Matt, in- Yeah, they'd get it up there. [41:21] Richard Webbe: Used to be one of those sitting up in the Bureau of Met in Melbourne. Oh, that's ANU, isn't it? That one there- Oh, yeah ... is the Australian National University. I think I was there a few months ago talking to the guy running- Mm-hmm ... the supercomputer there. The, the one in the square box, not the other one. [41:35] Dave Pengelley: Yeah. Yeah. Yeah. Yeah, no, I, I love that they built a little couch on the outside of the computer. [41:39] Richard Webbe: Yeah. Well, it was, as an engineer, we used to find it fun 'cause we would sit down. [41:45] Matt Slager: Um- 'Cause Richard, you said before, um, with networking and, and people needing bandwidth, the, um, the world of LLM inference, the, the bandwidth metrics there and the actual, you know, [00:42:00] I, I don't even know. Like I've started to look into local hosting just for myself, but that's one user. You know? Yeah. As soon as you have a company that has 100 users, 1,000 users- Mm-hmm [42:09] Matt Slager: and they're trying to all pull from their own sovereign infrastructure- [42:12] Richard Webbe: Mm. Yeah, yeah ... [42:13] Matt Slager: you, you have to start scaling. [42:16] Dave Pengelley: Well, the client- When, when I was, you know, had big dreams about Sergeant and I started building and thinking about the potential for it to not just be a desktop personal agent, but roll it out to teams and roll it out to enterprises and ha- have a cloud version, but then a self-hosted version of the Sergeant harness where you could have deployed to all of your enterprise and have every worker has their own little business in a box with their own virtual team. [42:38] Dave Pengelley: I started doing the maths on, on the tokens, and I can't remember off my, top of my head, but it was gonna be, you know, hundreds of thousands of dollars of tokens very quickly, um, if you wanna roll this out across a workforce of, you know, 500 or 1,000 people, whatever. To the point where I started getting into sort of forecast out the maths for paying for the metal for that kind of stuff. [43:00] Dave Pengelley: And yeah, it [43:00] Dave Pengelley: became pretty relevant, pretty, pretty obvious pretty quickly that if people really wanted to deploy mass agentic systems out, there comes a point where you stop paying for tokens and you start paying for server cycles instead. [43:11] Richard Webbe: Yeah. Yeah. ' [43:11] Dave Pengelley: Cause that's gonna become a lot cheaper. [43:13] Richard Webbe: It's the client server versus on-site versus off-site efficiency cycle model, and it's influenced by the build shock model, which we know quite well. [43:22] Richard Webbe: So wherever the most cost-effective place is to do the processing, and wherever the most cost and easiest way to get the information to the right people, the right applications at the right time will drive it, and that will continually move. We've seen the, the client server model move, and we've seen the cloud model move. [43:38] Richard Webbe: A lot of people bringing things from cloud back in-house, 'cause it's a lot cheaper to do some of their processing in-house rather than with the build shock they're getting from the- Yeah ... hyperscalers. And that's not gonna be different to AI, and it's gonna create a very positive competitive environment for cost-effective use of AI in the future, I think. [43:58] Dave Pengelley: Well, I think NVIDIA [00:44:00] sell, or, or they, they like... There's various versions of it, but it's based on the NVIDIA DGX or something, Spark, whatever. Yeah. But they're about eight grand, I think. So for eight grand you can have a reasonably Half decent local model server- Mm-hmm ... running on your LAN. [44:17] Richard Webbe: Mm-hmm. [44:17] Dave Pengelley: Mm. Um, that's, you know, I think, I think for your tech nerds and, and people that are really intense into this or your working from home people, that's gonna become more and more likely. [44:27] Dave Pengelley: The same way people go... Used to be, like, the joke was, "Oh, you wouldn't run your own power station, would you? You centralize all that." But now everyone's put solar panels and batteries on their house. Like- [44:38] Richard Webbe: It's all about cost ... um- It's cost and availability- It's- ... and the cost of transmission. [44:42] Dave Pengelley: So there's gonna be a point where, you know, that's part of a standard house setup. [45:00] Dave Pengelley: Not only do you have a, a data rack in your brand new build, but you cost in, there's a Tesla battery for 10 grand and there's my, my model host for another 10 grand of hardware as just part of the build of the house because I want [45:00] Dave Pengelley: to have the smart house. I want all the smarts and have- Do we want [45:02] Richard Webbe: every house to be a data center? [45:04] Dave Pengelley: Well, yeah, every house... Well, not every house, I'm sure. You know, that's an overgeneralization, but, uh, but it's gonna be a lot more prevalent. The same way that people are getting into Bitcoin, buy their Bitcoin nodes and they run their own Bitcoin node in house. It's all part of that sovereignty conversation. [45:17] Richard Webbe: Yeah. [45:17] Dave Pengelley: Yeah. Different kind of sovereignty with the Bitcoiners, but, um, [45:20] Richard Webbe: related. I've seen some presentations recently of various architectures for AI. Um, one particular chap put up his Mac Mini as his central processing point with links to all the LLMs and links to his data through, you know, G Drive and- [45:36] Dave Pengelley: Yeah [45:37] Richard Webbe: iCloud and all of that, and he's brought all of that together. But as you're saying, Dave, will that scale? [45:45] Dave Pengelley: Yeah, he's gonna hit hardware limits of, of the Mac and then, then, then the new, the new Mac Minis and the Macs are, are very good and they can do a lot but, you know, there's a reason why NVIDIA stock has gone up a lot over the last few years because the GPUs are really good for running these models and [00:46:00] things on. [46:01] Dave Pengelley: Um, and GPUs, yeah, I mean, we- it's just one of these things. We're gonna see more of it. Like, there's gonna be more of that. And at the personal level, at the small business level, and at the enterprise level. Back when I started doing IT 20 plus years ago, it was really common that small businesses had a Microsoft Small Business Server 2003 in the corner. [46:22] Dave Pengelley: That ran their user logins, it was their file server, it was their print server, ran their Exchange for their email. Like, they ran that. There was no cloud hosting for this sort of stuff. You ran your own SMTP and POP server or IMAP server off an Exchange box that lived in your office in the corner. I, I worked for a smash repairs at one point. [46:40] Dave Pengelley: They had a little server rack in the corner running an SBS 2003, um, edition. Like, this was the thing. Like... And I feel they never even think... Got it outsourced into, um, Office 365 and, and Azure Cloud, but some of this sort of stuff is gonna head back out to the edges. [46:58] Richard Webbe: It is. No, you're absolutely right. [00:47:00] It, it'll be a cost processing power, availability, accessibility model and it will continue to change. [47:08] Matt Slager: You're absolutely right. [47:10] Dave Pengelley: Mm. [47:10] Matt Slager: The, um, the, the idea of the economics of token-based pricing, you know, there's already companies, like big enterprise companies, that have just turned off their, their company's access to, to Claude and stuff like that because, you know, their budget for the year is blown out in a month. [47:27] Matt Slager: And, you know, all that, like companies like Uber, and I'm pretty sure even hilariously there was a presentation of Microsoft using Claude agents. Um, so yeah, it, it's just gonna keep getting more and more interesting. [47:41] Dave Pengelley: Yeah. Yeah. I've, I- I've canceled my Claude. I, um, um, was on the Max 5X and that's, you know, Australian 150 something a month. [47:50] Matt Slager: Wow. [47:50] Dave Pengelley: And I have been maxing it, like, uh, last week, uh, well the last two weeks I was like 90, 100% by the time I hit my [00:48:00] refresh. Yeah. Uh, last week... And then, then I got scared off and we've had a lot of stuff going on with family the last week, and so last week I only used like 60% of it, um, if that. Um, and then this week I was back to 90% before reset. [48:14] Dave Pengelley: So I'm, it's not like I'm not using it, but I'm also gonna offload a lot of that stuff from Cowork into Hermes. Yeah. I'm also getting more efficient with the way I'm using it. I was like, "5X, I've got tokens to spare. It doesn't matter if my processes are not super efficient, they're running anyway. I've got enough tokens." [48:28] Dave Pengelley: Whereas now as I move to Hermes and I'm trying to bring the cost down and, and optimize, A, for my own, you know, hip pocket, but also because these are lessons I want to be able to help other businesses and customers with, right? That, that's, that's half the point of, of doing this kind of thing. [48:43] Richard Webbe: Yes. [48:43] Dave Pengelley: Um, and going, well, you know, I was just bringing all the data in through some MCPs and making Claude do all the disaggregation and filtering and sorting. [49:00] Dave Pengelley: Hey Hermes, let's build a few like end-to-end workflows that pull the data. Let's keep track of the delta of what the old [49:00] Dave Pengelley: data was, what the new data, so when you do a pull for the morning review, just pull the newest version of the data. You don't need all the old data to sift through as well, and really start optimizing how much data I feed through and how many tokens I'm gonna burn on my daily workflows, right? [49:16] Dave Pengelley: Um, and so there's a lot of that work you can do. Can [49:18] Richard Webbe: I ask a specific question that people who watch this will probably be asking? I'm a medium-sized business. Let's say I've got between 50 and 100 employees. I've got calls coming in, meetings organized, product shipping out. Uh, I've got to schedule services, things like that, and I'm looking at going into AI to help me clean up, maybe, uh, redeploy a number of those people that just do paperwork and systems that could be automated. [49:47] Richard Webbe: If I'm looking at that, where do I start? What do I do? [49:52] Matt Slager: It's a, it's a great question. I think the... 'Cause I've had a few of these chats recently. The biggest thing that always comes out [00:50:00] is if you're the business owner in this case, CEO, the operations, whatever, you actually don't know your system as well as you think you know. [50:09] Richard Webbe: Yep. [50:09] Matt Slager: That, that's where it comes down to first. It's, it's understanding system decomposition, and, uh, just purely that. Just systems decomposition. You know, y- you could look of it, look at it as, like, you draw on your whiteboard the most zoomed out version possible of your business, and then you pick one of those blobs and then you zoom into that one and draw that one again until you understand who's responsible for what, what those processes are, what the actual sovereignty and provenance of those processes are, and where everything goes, where's the time spent, and which tools are actually involved here by which person. [50:45] Matt Slager: And until you can define that, you won't know it, 'cause that's what someone like me would say to you, is, "We need to get this mapped out before we do anything." Yep. Um, so you either do it now or you do it when you start consulting. [50:57] Richard Webbe: Absolutely. What do you think, Dave? [00:51:00] [51:00] Dave Pengelley: Yeah, I mean, with that 50 to 100 people thing, it's where is the value gonna be driven? [51:04] Dave Pengelley: And it comes back to, like Matt said, what are, what are your core processes and where are the bottlenecks? Where are the costs? And, and are you augmenting everyone or are you just augmenting and automating specific parts of a process that, uh, multiple people touch? Like, that's an area where it's, I, I... [51:21] Dave Pengelley: there's probably not an ROI on giving all 50 to 100 people their own personal agent and their own personal token subscription. Yeah. Uh, that's where you are looking at, you know, putting in a centralized system that has some agentic knowledge or summarization or coordination skills involved, but it's more that sort of systems factory line point of view than virtual employee sitting at everyone's desk. [51:45] Matt Slager: Yeah. The way that I would sort of define what you just said is it's not about replacing or automating people. Like you used the term augment. Uh, augment is the best way to describe it. Yeah, I'm [51:56] Richard Webbe: with you, I'm with you on the augment. [51:58] Matt Slager: Yeah, 'cause your, your people already exist. [00:52:00] Like, if you wanna just cut people out of your business, fine, that's your decision. [52:03] Matt Slager: But, you know, they already exist, they're already there working. You, you know that they're probably wasting time on certain things. So it's not about their position, it's not about their role- Mm-hmm ... it's about their individual workflows, their actual- Yeah ... tasks and the granular steps of each piece. [52:19] Richard Webbe: Yeah. [52:19] Matt Slager: And that's where you need to look at augmenting or automating those little individual pieces. [52:24] Richard Webbe: Yeah. [52:24] Dave Pengelley: Yeah. Yeah. And, and how do you do that most effectively? And so, and I think Sean even raised this a few weeks ago when he was on the show talking about, you know, that, that data scatteredness, and the first thing most businesses need to do is actually order where all their data sources are and potentially sort of rationalize some of their, their data infrastructure in order to make it accessible, um, for agentic systems and automations. [53:00] Dave Pengelley: Like, a lot of the things and improvements are just automations and wiring systems together that don't necessarily need a lot of agents. Like I said, my... even my morning reviews where I'm pulling all my emails and doing this and that and the other, half of that is just in the workflow step. I was [53:00] Dave Pengelley: making the AI do all of that with cowork because I could. [53:02] Dave Pengelley: I said, "Yep, go pull the, the emails and then sort through what's new today versus yesterday." And now I'm going, "No, I'm gonna have a workflow that knows we already looked at yesterday's emails. Just pull t- today's ones- Yeah ... uh, and, and pre-filter those, and then just give me the meaningful bits, because that's what actually matters." [53:19] Dave Pengelley: And so it's easy to be lazy, but as the costs go up, um, the value is not there to be lazy, and we need to be really targeted and specific. And you're gonna see a lot more of this where people are running a lot of agent experiments. We'll be shutting them down and turning them off, I think, because they... [53:34] Dave Pengelley: there's no value there. It was just, "Oh, well that was fun while it lasted, but it didn't really matter. Let's get back to work." [53:40] Richard Webbe: Mm. Back to the person getting the, uh, the Windows program to run their house. It doesn't really add value. So I suppose what you're both saying, and, and I agree with, is know your process, know what your business does, know the touch points, know where your data is. [54:00] Richard Webbe: And then the other thing is- I, I think it's know what your unique value is. What part of that process [54:00] Richard Webbe: has made you successful above everyone else or kept you in the game? And I think there's gonna be some real exciting and interesting things for businesses to discover in that area. They may find it's the receptionist who treats everyone on the phone or comes in really well, and they come back because of that person. [54:19] Richard Webbe: And I think one of you mentioned it last week, uh, one of the heads of the AI thing said, "Get your mental health, get your soft skills, get your emotional intelligence intact because that's what's gonna differentiate you in the future, not process." 'Cause process will be able to touch, feel, automate, move, but it's what the humans do with that that's more important. [55:00] Richard Webbe: In IBM, when I worked at IBM, there was something like 40,000 daily reports were produced on the massive data processing around the globe for IBM, right? Now, I'm sure these reports are everything from, you know, leave and who broke their leg last week and who got a promotion and who's gonna get one and all of [55:00] Richard Webbe: that. [55:00] Richard Webbe: But they found out between 1 and 5% of those reports were of any value to anyone. Mm-hmm. And they found that .01% of them got read. So I imagine a world in the future where we're not focused on the reports at all. We're focused on the process and the agentic automation of that process. Then we don't need reports to tell it's, it's speeding up or slowing down. [55:26] Richard Webbe: Do you know what I'm saying? It's like- [55:28] Matt Slager: Yeah ... [55:28] Richard Webbe: why would I look for improvements? I know where it is. I don't need... It's 2% better than it was last week. I just need to know that if I accelerate that, my business value goes up. Does that make sense? [55:39] Matt Slager: It does. It's so strange. It's like when people say that I, I always would prefer human service over, you know, work- uh, doing the self-scan at, at Coles. [56:00] Matt Slager: And my, my, um- Uh, dystopian kind of take on this is that it's a [56:00] Matt Slager: weakness that we have as a human being to, to need to have this connection. You know? If you didn't need that, all of a sudden it's, everything f- is fine and, like, all the issues of the world kind of reduce. Um, but- [56:15] Richard Webbe: COVID taught, COVID taught us that. [56:16] Matt Slager: Y- yeah, right. And- So the, the, the concept, it remains, and it will persist, and if, in perpetuity, we need human, human interaction, and that will define a lot of our own values and a lot of our own decisions. So it's exactly what you just said, Richard. Like, if, if there's a company that's like, "Man, we've got this receptionist that's doing nothing," like, we can just replace that with a voice AI, surely. [56:37] Matt Slager: But all of a sudden, their regular people that alw- only came in to speak to that receptionist, you know, it, it's a, it's a consideration that they have to try and make. Like, um, when you talk about, like, your individual sort of method, like the reason why you're successful, like your unique spice compared to your competitors, it's usually a human thing. [57:00] Matt Slager: Uh, it, it's very rarely, like, [57:00] Matt Slager: some sort of secret process like KFC spices, you know? Um, it, it's usually the, something to do with, like, "Oh, I just like that company better because they're, they... I don't know. I just like them," you know? Like it's some sort of weird human thing. [57:12] Richard Webbe: And I use this example I think the first time you guys invited me on to be a part of the team, and it was the customs officers in the UK. [57:20] Richard Webbe: Because sometimes, as you said, you don't know why you're so successful. You keep doing things well and improving. You tend to throw a lot of good stuff at a lot of different moving parts. And of course, they were trying to spot people that were doing something criminal that were coming through Heathrow Airport, and they tried to capture it on software. [57:38] Richard Webbe: This is like, I'm talking 35 years ago, and they weren't sure, and they had... I know anyone who's been through Heathrow, there's all those mirrored windows you walk past, and you know that on the other side of there, there's probably a dozen agents taking notes, probably some facial recognition software, all that stuff [57:53] Matt Slager: going. [57:54] Richard Webbe: Human agents. Human agents, yes. And, and they were, like, so, and they're all taking notes, looking. And so [00:58:00] they couldn't get, like, they still only got 50/50 on who was a crook, but the customs officers, they were 95% accurate So something was wrong between getting the AI or the, the algorithm to spot a crook and not. [58:15] Richard Webbe: And so they went to the officers and said, "You know, w- what is it?" And they say, "Oh, it's in the eyes. We can always see it in the eyes." Mm-hmm. So what they did is they turned their analysis not onto the people, but onto the customs agent on what they did to spot ... And it wasn't the eyes. After they did all this exhaustive, and I won't go into all the details, they found out, and a lot of criminals might be listening to us today, it's the stiffness of your neck, right? [58:40] Richard Webbe: So the customs officers would see the eyes and go ... And if they were going like this, they'd get a easy lock in their eyes, because they're not moving their neck much because they're uncomfortable and they're stressed, and they're probably doing something illegal. It was just the most amazing case study of [59:00] Richard Webbe: Spent, they spent millions trying to find out, how do we catch someone for being a crook? Uh, and [59:00] Richard Webbe: there's other ways they do it. In the gait, in the way they walk, in all those things. But you need to be open to identifying what it is before you can capture it and reuse it and accelerate it. It's gonna be a big change, an interesting change [59:13] Matt Slager: Hmm. [59:14] Matt Slager: It makes me think of big data and, like, people getting tracked and their behaviors online getting identified and, um, hypothesized and potentially extrapolated. Like, you know, what's this person doing, and what are they thinking, and what are they going to think in the next few weeks? Um, yeah, I love all that stuff. [59:33] Matt Slager: It's very interesting. [59:34] Dave Pengelley: Watch, watch Person of Interest. [59:35] Matt Slager: Yeah. I need to try and find some time. [59:39] Dave Pengelley: But I mean, the, the, the end, end, end point of that is the whole pre-crime issue, which, um, you know, Minority Report highlighted- [59:47] Richard Webbe: What a [59:47] Dave Pengelley: great show ... which we've gotta be careful of, right? Um, pattern behavior and, and past recognition isn't always a, the, the definite predictor. [59:55] Richard Webbe: Oh, you can get it wrong. You can get it wrong. My mom was one of those receptionists at the local real estate agent, [01:00:00] and they had an extremely high amount of business. Um, and I remember someone walked in and they were talking to my mom. I was sitting in the reception waiting for her, and someone else popped in and said, "Oh, why are you using these guys? [01:00:11] Richard Webbe: They charge way too much money." And the guy just quickly turned around and said, "Yes, but I don't get to talk to Pat, and if I couldn't talk to Pat every day, I wouldn't care." So that was my mom. So you're right, human connection is the thing you want to accelerate the most. [01:00:25] Dave Pengelley: Al- although I go, I go to the checkout and I go to the human, not because I want to have a chat with the person specifically. [01:00:31] Dave Pengelley: I just think it's good that we still employ humans, and they're not paying me to, to, to check out my own groceries. If they gave me a 5% discount for checking out my own groceries, then maybe I'd do it, but they're not. So they can pay someone to pack my bags for me. [01:00:46] Richard Webbe: I agree. I find it a problem that I have to pack my own bags. [01:00:49] Richard Webbe: I think that's insulting. [01:00:51] Dave Pengelley: Although my wife prefers packing her own bags because half the, the teenagers that do it are useless, and she's like, "Ah, just why, why do we have to go through a human?" Like, "Dugh." She just [01:00:59] Richard Webbe: wants to pack her own bags. You would, you would be [01:01:00] amazed. I accidentally talked my way into, and I won't say who, um, uh, one of our largest supermarket's, uh, secret labs and how they were using AI to track what we buy Grab it, charge when you get there, which I think happens in a number of places already, and we'll start to... [01:01:17] Richard Webbe: And the amount of technology they're using to track what we do and how we do it, it's, it's... I don't think we're gonna have checkout people in the world in very much longer. [01:01:26] Dave Pengelley: Well, that, that was the Amazon. Amazon had that store where you'd walk in and there's some kind of RFID tag or facial recognition that picks up who you are. [01:01:33] Dave Pengelley: You just walk in, grab stuff off the shelf and walk out, and it would just come off your Amazon bill. [01:01:38] Richard Webbe: So I'm not sure if this is sure you're exactly right, but sometimes when I'm trying to use the checkout and it's not working, it knows what I've put on before I put... There are a number of camera doing this stuff. [01:01:48] Richard Webbe: So I don't think they actually need the checkout scales and that, the barcode scanners anymore. I think they're getting close to not needing that at all. [01:01:58] Dave Pengelley: And I do like that they're using the picture [01:02:00] recognition so that way you don't need to tell what kind of fruit or whatever, or it narrows it down very quickly- Yeah. [01:02:03] Dave Pengelley: Okay. Yeah ... to the green fruits or the yellow fruits or whatever. Like- [01:02:06] Richard Webbe: Yeah ... [01:02:06] Dave Pengelley: some of that sort of stuff's pretty- [01:02:07] Richard Webbe: It [01:02:08] Dave Pengelley: is good ... handy. But yeah, I, um, as a rule, unless I'm in a real hurry or would like to just get one or two things, or sometimes they just don't have the humans open, I will generally stand in queue and wait for a human, to the point where they've actually had to open up extra checkouts because there's a few of us waiting. [01:02:25] Dave Pengelley: I'm like, "Excellent. That's what I want." [01:02:29] Richard Webbe: I love it. Very true [01:02:31] Dave Pengelley: So, uh, yeah, so my journey's gonna continue this week. It's trying to find out, uh, what other models and subscriptions and, and things I can do to replace my giant Claude one. I've got a couple weeks left of my Claude sub before that expires, and so I'm still trying to max out and balance between Claude and, um, other things. [01:03:00] Dave Pengelley: But, uh, yeah, I'm, uh, if anyone in the audience wants to reach out to me and tell me you've had great experiences, you know, with, with Minimax or Kimmy or GLM [01:03:00] Dave Pengelley: or Groks or whatever those things are, I am looking for options because it's one of those things, like I can do the report, I can, I can... And I've had Claude and ChatGPT and Gemini all go and do deep research reports going, "You could use this plan and this plan and this plan and this one, but pay for that one by the token and the API." [01:03:17] Dave Pengelley: And I'm like, "There's so many different options, and what if they all suck? I don't even know." So I just want to back to the connection. I would like some humans to say, "Hey, I found really good experience with this, and the $20 model gave me heaps of usage," or, "That one was, uh, better to be on the API usage, and I smashed it and only paid 10 buck-" Like whatever those stories are. [01:03:36] Dave Pengelley: That's-- But no one's really talking about that, and you can't find that easily online, and the hype bros are just hyping, and they don't even care how much money they're spending on tokens. I think they're just making YouTube videos out of it all, so. [01:03:49] Matt Slager: I probably have a lot of that information for you [01:03:51] Dave Pengelley: already, Matt. [01:04:00] Dave Pengelley: You probably do, Matt. But, uh, sometimes I feel I, I, I'm like, you've got so much information sometimes and you say, "Just do this, this, this, this, this, and this." And I'm like, there's [01:04:00] Dave Pengelley: too many options there. Like, like in, uh, in that group chat where I was saying, "What are the options other than these models?" And I would like, A, I talk to you all the time, and I like your input, but I was hoping to get the broader consensus for other people. [01:04:12] Dave Pengelley: But you just put it like, like a million models, and then I think everyone's like, "I'm too scared to answer now." [01:04:16] Matt Slager: Do just OpenCode Go. Just get an OpenCode Go subscription. It's 10 bucks US, and that will get you everything that is just not Frontier. [01:04:28] Dave Pengelley: Yeah. Yeah. Okay. OpenCode Go, I'll look in that. Apparently, the NASA Research one comes with some great, like included fire crawl type stuff and, and that, so. [01:04:36] Matt Slager: Yeah. Yeah, that one's worth the 10 at the minimum. [01:04:38] Dave Pengelley: And I'm looking at said Super Grok as that evolves because that's got pretty good image and video generation. [01:04:43] Matt Slager: Yeah, and voice, so if you're wiring in voice stuff. Um, but yeah, as far as your own personal subs, stick with, um, stick with OpenAI for the work you're doing, and if you need to add Claude into it, then I would do it through Cursor, [01:05:00] um, at the [01:05:02] Dave Pengelley: 20. [01:05:02] Dave Pengelley: I, uh, I definitely need additional models in Hermes because my ChatGPT five hour one, I'm just maxing all the time at the moment [01:05:09] Matt Slager: with- Dude, if you just need it for Hermes and it's just for agentic tasks, then do the either OpenCode Go or Minimax, and the Minimax $20 one. [01:05:20] Dave Pengelley: All right. Thank you. I appreciate that. [01:05:22] Matt Slager: No worries. All good. [01:05:25] Dave Pengelley: All right, guys. We are, we are over the hour. Um, we didn't do all our little segments today. We've, um, nerded out on Hermes chats and things, so, uh, with some Sovereign discussion sprinkled in there too, so... [01:05:40] Matt Slager: It was good. [01:05:41] Dave Pengelley: Yeah. Yeah. I think, uh, the, the, the, the pace is moving fast. Uh, if you joined us on the chat, thank you. [01:06:00] Dave Pengelley: You didn't join us... No one chatted with us today, but we've had a ha- a, a handful of people, um, watching us live. I can see that in the numbers. So thank you for joining us. Hope you enjoyed the, the nerdy brainstorm chat there as well. Have you experimented with Hermes? Drop a [01:06:00] Dave Pengelley: comment in the video. Comments help the algorithm show our chats to everyone. [01:06:04] Dave Pengelley: Make sure you like, hit the subscribe button, and the little bell icon, so you get notified each week when we go live at midday Wednesdays, uh, East Coast time. So do that. Um, and if you're interested in the live event, go to aioperatorspod.com, and there's a link to the IRL event there where you can register to, uh, join us for a beer and some discussion out in Rouse Hill. [01:06:27] Dave Pengelley: So thank you everyone. Appreciate it, guys. We'll see you all same bat time, same bat channel next week.