Access, Audit, and Control Are the Product. Harness Bias, GLM 5.2, and the Local Hardware Boom.

[00:00] Dave Pengelley: Oh yeah, Richard doesn't know. [00:03] Richard Webbe: I'm excited. [00:04] Dave Pengelley: Yeah, you should be. You should definitely be very excited because we are here. We are doing shows. Oh, I forgot to download a new snippet you haven't seen, Matt. So let's just um while I actively not am not looking up websites and downloading assets, I will say um you said late 90s. So do you like late 90s stuff, Matt? [00:25] Matt: Yeah, yeah, look, that's when I was born. Uh I was actually done early nineties, but you know that. era. I was aware of myself by the late 90s at least. [00:34] Dave Pengelley: Yes. [00:35] Matt: Yeah, it was a it was a good time. [00:41] Richard Webbe: I was born [00:41] Matt: Yeah. [00:41] Dave Pengelley: No [00:41] Richard Webbe: in the [00:42] Dave Pengelley: doubt. [00:42] Richard Webbe: sixties. [00:43] Matt: Yeah, my uh my my upbringing was somewhat uh sheltered in that in that sense Dave with regards to ratings and whatnot. My parents meant well. I had I had my father uh force me to return Super Smash Brothers on that 'Cause it was too violent. Um yeah, it was it was an era. [01:05] Richard Webbe: You [01:05] Dave Pengelley: Wow. [01:05] Richard Webbe: had a caring family, well [01:07] Dave Pengelley: That's [01:07] Richard Webbe: done. [01:07] Dave Pengelley: that would that would yeah, that was definitely not the 80s parenting that I experienced where there was there was nowhere near enough supervision and censorship on the content we watched. Um which you [01:20] Matt: But [01:21] Dave Pengelley: know my my son's getting a very different uh experience to what I had. [01:25] Matt: making up for lost time. No, I used to love um going to friends' house. And they had all of like the the violent games. So like we played Goldeneye and Mortal Kombat and like all the cool things. [01:37] Dave Pengelley: Yeah, yeah, I'm not sure my son's seen Ghostbusters yet, even though he's 15, because I was always wary of the bit where the ghost unzips Dan or Aykroyd's pants. [01:50] Matt: Oh, [01:50] Dave Pengelley: Um, like violence and stuff we're okay with, but you know, that's that saucy kind of adult relationship stuff is the kind of arena that we've been uh paid more attention [01:59] Matt: yeah. [01:59] Dave Pengelley: to. when it comes to shielding our children. [02:01] Matt: Fair [02:01] Richard Webbe: Okay. [02:01] Matt: enough. How are we going to fix the population decline though? [02:06] Dave Pengelley: Not by getting fifteen year olds pregnant. [02:11] Matt: Good good answer. Good [02:13] Dave Pengelley: Yeah, [02:13] Matt: answer. [02:13] Dave Pengelley: on that, let's let's let's uh let's break it let's um do a show start [02:18] Matt: Wow. [02:19] Dave Pengelley: playing [02:34] Dave Pengelley: The wrong one. I realized I was like, no, that's the wrong one. That's the end of the show one. [02:38] Matt: Oh well it'll [02:38] Dave Pengelley: Spoilers. [02:38] Matt: worked. [02:39] Dave Pengelley: Spoilers. [02:40] Matt: The little little Gemini cameo there, you know, Gemini coming in strong with the with the new footage. [02:47] Dave Pengelley: No, Grok imagine. [02:49] Matt: What was the diamond? There was the little Gemini diamond. [02:51] Dave Pengelley: No, that that's that's a grok [02:54] Richard Webbe: AI [02:54] Dave Pengelley: thing. [02:55] Richard Webbe: operated logo. [02:57] Dave Pengelley: How hang on. What what are you saying? [03:02] Richard Webbe: Let's [03:02] Dave Pengelley: Hang [03:02] Richard Webbe: get [03:02] Dave Pengelley: on. [03:02] Richard Webbe: professional, fellas. Well you've distracted the the [03:04] Dave Pengelley: Let's [03:05] Richard Webbe: MC, [03:05] Dave Pengelley: watch this again. Why is Gemini taking [03:17] Dave Pengelley: credit for Groc's work? [03:20] Richard Webbe: what's going on? [03:21] Matt: It it's just like when you get new Chinese models and they say that they're clawed. [03:27] Dave Pengelley: Like that is a hundred percent I'm sure like I was oh that one might have been grok actually. Oh I'll I um Gemini. I apologize. I was doing experimenting in two different platforms trying to get the best result. That one might have been a Gemini special. The other one for the end of the show is a Grok. [03:47] Richard Webbe: Now while we're talking about different LLMs, I noticed a uh comment on LinkedIn, and this gets to the edge of politics, but it's not real politics. And um a friend of mine sent it to me who was doing some research. Research and they sent it to me and said, Hey, look at this. And uh, if this lady's listening, and she may well be because she was an AI ethicist, and I love this birth of ethical AI people appearing everywhere without really knowing what the friggin' hell AI is. But anyway, and so this person, and I respect them for their comment, wrote an article about all the answers they were getting from AI, Groc, from OpenAI, from ChatGPT, from you know, Claude, you name it. And they They didn't actually say it, but they're hedging towards political bias in the responses. And I got so grumpy about this comment because we know that AI is just the sumnet average of everyone's opinion or the available data to that LLM to make a decision on what the reality might be, right? Now, those that are experienced have a mathematical mind understand what that means. This person was trying to drum up ethical AI business for themselves, and I respect that, but the you know the The implication of a conspiracy in the background of LLMs giving certain answers for political bias or whatever, it just floated my boat and I was triggered. [05:10] Dave Pengelley: There there is it's not even completely unfounded though, Richard. I mean to be fair, like it is they are trained on a certain set of data, but [05:17] Richard Webbe: That's [05:17] Dave Pengelley: you know, [05:18] Richard Webbe: right. [05:18] Dave Pengelley: the companies can [05:20] Richard Webbe: They can do, and [05:21] Dave Pengelley: yeah. [05:21] Richard Webbe: that's right. And these and then and there's there's severe um investigation and understanding of that, but the way this person framed it was well, it's pretty obvious. You know, because it's Groc, Elon Musk is influing us all, and because it's OpenAI, this person. And I was just really triggered by that. I mean, at the end of the day, the growth in AI is now driving a lot of hardware sales and it's driving a lot of infrastructure sales because people, as you articulated very well at our conference we spoke at the other week, that no longer are people having their, you know, their algorithms, their ideas and devices here and all going back to the LLM. They're all going back in their own spot and only going to the LLM when they require a complex calculation. Right. And so it's all self-contained. And that self-contained rag, if we want to call it that or something like that, actually runs on the basis that it's using its own data lake, its own source of data. So you're only going to corrupt answers if you corrupt yourself. And at the end of the it you challenge me, at the end of the day, the AI answers depend on the data you have available. And as someone said, if you get a thousand people to say the sky is purple, I'm sure an AI is going to come back and say, sorry, the sky's purple. [06:31] Matt: I have I love this. I this is such an amazing way to start the podcast. This is going to be a great episode. Uh so everyone tuning in, strap in, because there's a few things that we say on repeat. You know, there's a few things that I bring up on repeat. You could go back to the very, very early episodes, and I spoke up a few concepts around this exact topic, you [06:52] Richard Webbe: Yeah. [06:52] Matt: know, with regards to how to protect yourself from it. Not whether it does or does not exist. Can or cannot happen, but that it will. And how do you protect yourself from it? How do you remain aware of it? You know, just because the road is empty does not mean you can close your eyes and cross the road. [07:15] Richard Webbe: Very true. And it's confirm you've got what you're saying, Matt, and I totally support it, is remove your own confirmation bias and remove your lazy, what I call lazy bias, because you think someone else who knows more has told you said that must be the truth. So you've got Confirmation bias and lazy [07:31] Dave Pengelley: Well, [07:31] Richard Webbe: bias, [07:32] Dave Pengelley: it's [07:32] Richard Webbe: right? [07:32] Dave Pengelley: the appealing to experts. That's [07:34] Matt: Yes. [07:34] Dave Pengelley: the argument that people appeal to the expert. And more the expert said that we saw a lot of that a few years ago around um a certain sickness that happened around the world. [07:46] Richard Webbe: You try so you try so well to stay away from politics [07:50] Dave Pengelley: No, [07:51] Richard Webbe: that [07:51] Dave Pengelley: I on [07:51] Richard Webbe: they [07:51] Dave Pengelley: the [07:51] Richard Webbe: can [07:52] Dave Pengelley: on [07:52] Richard Webbe: do [07:52] Dave Pengelley: the buy [07:52] Richard Webbe: it. [07:52] Dave Pengelley: stuff. On the bias up, we're like they've done studies, I've done research, they've done tests and said if you run a question through chatgpt.com, the Versus directly to the OpenAI API, you will get different responses because the [08:04] Richard Webbe: If [08:04] Dave Pengelley: chat [08:05] Richard Webbe: we try [08:05] Dave Pengelley: GPT, the harness, we talk about the harness all the time. [08:07] Richard Webbe: do it, to [08:07] Dave Pengelley: That set of layered guide rails and rules and instruction sets they put in on top of anything you ask it influence the answers you [08:15] Richard Webbe: a bit can't get little more. you [08:15] Dave Pengelley: are you can and cannot get. And at the API level, there's probably still some extra guardrails, which we'll get onto in the news around some of the newer, more advanced models that they're they're putting in place. But I mean, one of the common examples around you know the ethical, moral, political bias. These things that comes up is when they say, uh, give me a critical analysis of the negative aspect of religion A versus religion B. And one religion goes, Oh, well, I don't touch a religion, I can't talk about that. And then we go, Well, on this one, these people say this and they do that and the other thing. And it's like, that's that's interesting that you would refuse to give negative critical uh analysis of one religion, a certain religion, but the other one is sort of open slattered and they go nuts on [08:55] Richard Webbe: I [08:56] Dave Pengelley: us. [08:56] Richard Webbe: I I look, I I watch two guys, I forget who they are. Now, but they had, I don't know if they had a podcast. Yeah, the one had a it was only audio, but they were years years ago, they were going to Google and and other search engines and they were asking a specific question. And one of them asked, Who is Jesus? And it came back and said, A mythical creature. Well, by all data, uh, Jesus is not mythical, he was a real person. Whether you believe he was a prophet of God or a head of a religion, that's that's the other bit that's up for debate. And then they went and asked about other, you know, religion. Rig religious leaders, Jehovah, Moses, and it said they were all real, but said Jesus was a mythical creature, to which the Christian uh lobby got very upset about because that's actually not true. [09:41] Dave Pengelley: He's one of the most [09:42] Richard Webbe: But [09:42] Dave Pengelley: well documented historical figures every [09:44] Richard Webbe: exactly. But it it's interesting in that in that someone deep in the bowels of of of that library [09:51] Dave Pengelley: year. [09:51] Richard Webbe: setting up that answer was confirmation bias, and they didn't hadn't studied theology on what it really was. And so they said that they thought. Jesus was a mythical creature. [10:01] Dave Pengelley: So, to give your ethical AI um friend a break, there is some argument that some models or some harnesses wrapped around certain models will steer answers more one way or the other. And the Matt's point, if you're not aware of that, and this is why we need to make sure this next generation, even the current generation, they have this critical thinking literacy to challenge and question answers and to verify like Pixar didn't happen kind of thing. [10:30] Richard Webbe: of a rant you know when occasionally it's late at night you're on Facebook or LinkedIn and you decide to express an opinion without those filters [10:36] Dave Pengelley: No, I don't I don't do that. [10:40] Richard Webbe: I'm looking down sadly you know that carton that had a picture of the guy madly tapping on the on the on the keyboard at 2 a.m at night and his wife says to him what are you doing he said someone's wrong on the internet I must fix it now that was me yes anyway so um [11:00] Richard Webbe: Oh. [11:00] Dave Pengelley: It's not gonna change anything anyway, and we just discard draft, [11:02] Richard Webbe: Don't [11:02] Dave Pengelley: it's like [11:03] Richard Webbe: send that email. Yes. Anyway, so um I did have a little rant and I think I was quite articulate, but the opening might have been using words like silly and what were you thinking? And you don't know much about AI if that's your answer. Um, I subsequently have deleted my comment if that person's watching now, so they'll probably be happy about that. Um and their response. But what I did do is is I said, if you're going to ask AI. [11:29] Dave Pengelley: a lot of same. the [11:30] Richard Webbe: To give you a moral, political decision or answer. You might as well might as well ask a two-year-old how to fly to Mars, right? It's a dumb thing to do. And as you're saying appropriately, Dave, forewarned is forearmed, is prepared. So AI is not perfect. It's not the answer to everything. But can I tell you now, there are so many people out there now that are falling into that habit. And they think exactly it is the sumnet of the smartest people in the world. And it's not, it's the lower average. It's the Lowest common denominator. And that's how data mining actually works. That's how we get statistical analysis to try and come up with the answer to things. It works pretty well when you're talking about medicine and lots of data. And will I operate this or will I do that? Or will do that. And I saw a very interesting video on the weekend about the rise of China, uh, BYD and electric cars and AI. And it was rather amazing. And the same thing, they're talking about how the perception of what's happening. And it's happening is completely different. [12:32] Matt: Link [12:32] Richard Webbe: I'll shut [12:32] Matt: it [12:32] Richard Webbe: up. [12:33] Matt: link it in the chat. We should watch it some other time. [12:36] Richard Webbe: I will. I'll link in the chat. I'll put in the chat. You're [12:39] Matt: I [12:39] Richard Webbe: right. [12:39] Matt: have [12:39] Richard Webbe: Good [12:39] Matt: a um [12:39] Richard Webbe: point. [12:39] Matt: I've got so many comments. I've got so many thoughts. Do you guys want to keep going down this path? [12:43] Dave Pengelley: Let's let's let's let's do this for another few minutes and then we'll cut to news and talk about all the uh the [12:47] Richard Webbe: Go, [12:47] Dave Pengelley: model new [12:48] Richard Webbe: Matt. [12:48] Dave Pengelley: things that have been happening. [12:49] Richard Webbe: Matt, [12:49] Matt: All [12:49] Richard Webbe: you're [12:49] Matt: right. [12:50] Richard Webbe: the best with expert comments. Go. [12:57] Matt: And uh I think what you should Do is is write down that author's name. Let me let me just reach behind into my library. [13:06] Matt: This is how you spell it. Let me do this one. This one's probably the easiest to read. Robert Cialdini. [13:12] Richard Webbe: Oh, feel Janey. [13:14] Matt: With uh with C I A. And um, so [13:18] Richard Webbe: Yeah. [13:19] Matt: his [13:19] Dave Pengelley: You gotta [13:19] Matt: first [13:19] Dave Pengelley: you [13:19] Matt: book, [13:19] Dave Pengelley: gotta do [13:20] Matt: Influence. [13:20] Dave Pengelley: this when you say his name, you gotta like [13:22] Matt: Yeah, Gialdini. [13:24] Richard Webbe: I'll hold the book higher [13:24] Matt: Um [13:25] Richard Webbe: up. Hold the book higher [13:26] Matt: so [13:26] Richard Webbe: up. [13:26] Matt: influence, I think you would Benefit greatly by going to your AI of choice and saying exactly those arguments that you've just brought up with a little bit of context and saying what would Robert Cialdini say about this with regards to his book Influence and also Prasuasion. [13:47] Richard Webbe: Oh, persuasion. [13:48] Matt: So presuasion is adding bias. [13:54] Richard Webbe: Yeah, confirmation bias and turning it off and on, and yeah, I understand. [13:58] Matt: So the idea of going Going into a conversation where you know you want to go down a particular path and have the person ask a particular question or provide a particular answer. So you start talking about elements that are somewhat adjacent before you mention the main core topic. [14:13] Richard Webbe: Always, always. It's manipulated at the base. [14:17] Matt: Yeah, but this this is sales. You know, I another thing that I say a lot, and I haven't actually brought it up on this podcast ever, but I, you know, you guys hopefully agree with me. You've got a few more years on on yourselves than me. And I believe that the entire world, the entire Like our existence as an interpersonal relationship race, everything is sales. [14:35] Dave Pengelley: New [14:36] Richard Webbe: Oh [14:36] Dave Pengelley: evil [14:36] Richard Webbe: yes [14:36] Dave Pengelley: capitalist [14:36] Richard Webbe: it is. [14:37] Dave Pengelley: view. [14:37] Richard Webbe: I've absolutely and you'll get my book soon as soon as the publisher who's sitting to the left of you finishes getting the typesetting right for me, [14:45] Dave Pengelley: He's [14:45] Richard Webbe: but [14:46] Dave Pengelley: working on it. Um it's [14:49] Richard Webbe: I [14:49] Dave Pengelley: interesting [14:49] Richard Webbe: think [14:49] Dave Pengelley: you [14:49] Richard Webbe: you're [14:49] Dave Pengelley: uh you [14:49] Richard Webbe: right. [14:50] Dave Pengelley: talk about bias and political bias in models and stuff, but I mean there's the broader uh media bias around all this sort of stuff that everyone forgets to talk about as well. Well, and just accept, oh, it's in the news. And what is it? It's the um oh, I can't remember the name of the effect where you watch something on the news, which you is a topic you know well. Like see you see and see an article, you know, all about sales, and you go, That's bullocks, that's not how that works at all. But then it's like you turn the page and it's about some international affair, and you just go, Well, this must be true. Uh, and you just you just you just separate the two and go, even though the same media source just told complete lies and and non-truth. The next article you read from them, you go, Well, that one must be 100% true. Um, [15:34] Richard Webbe: It's [15:34] Dave Pengelley: that's [15:34] Richard Webbe: like the Women's Weekly. They used to put the new cake recipes at the front and the diet recipes at the back. They're both counterintuitive to the end [15:42] Dave Pengelley: but [15:42] Richard Webbe: goal. [15:42] Dave Pengelley: no one talks about that. That's um, I can't remember the name of the effect. That's um anyway, Michael Crown or someone his friend was named after him. Uh oh, well, it won't come back to me. But anyway, on that, look at this. Buckle up, the bad guys now have. A model as powerful as mythos. So let's just label China as the bad guys. That's [16:04] Matt: Well [16:04] Dave Pengelley: that's helpful. Um, [16:06] Matt: Yeah, that that's that's not an that's not an okay statement. [16:10] Dave Pengelley: right, like that's ridiculous. Like, come on, like, so yeah, this is um in reference to the GLM 5.2 model, but you know, just the bias is everywhere, and we all need to be aware of where whichever source we're getting it from, whether it's from an AI or a media news source or a podcast. Um if you disagree with anything we say, drop it in the comments. Even just now we had Richard talking and and we're getting counterpoints. Like that's what we do. That's that's the point, right? Conversation debate breeds better outcomes. It is one of the strongest arguments for free speech is that we will get better outcomes if we just have the chats. [16:51] Matt: Absolutely. [16:51] Richard Webbe: Yes. [16:52] Matt: I'd love to see more live comments come in and like trip us up. You [16:55] Richard Webbe: I [16:56] Matt: know, [16:56] Dave Pengelley: Yeah. [16:56] Matt: like [16:56] Richard Webbe: just I know this is a technical question while we're live, but How do I put stuff in the chat? It won't let me write anything. [17:02] Dave Pengelley: That uh you probably can't post in the chat uh back to other people, maybe. [17:07] Richard Webbe: I'll [17:07] Dave Pengelley: I [17:07] Matt: No, [17:07] Dave Pengelley: don't [17:07] Richard Webbe: forward [17:08] Matt: we're [17:08] Dave Pengelley: know. [17:08] Richard Webbe: to you what to put in the chat. I'll do it now for you. [17:11] Dave Pengelley: If you just want to send stuff to us, like a link you want me to put up on the screen, there's a [17:15] Richard Webbe: Yeah. [17:15] Dave Pengelley: private chat at the bottom of our [17:17] Richard Webbe: Oh, [17:17] Dave Pengelley: environment. [17:18] Richard Webbe: the public chat. I want [17:18] Dave Pengelley: On [17:18] Richard Webbe: to put [17:19] Dave Pengelley: the [17:19] Richard Webbe: one in [17:19] Dave Pengelley: public [17:19] Richard Webbe: some [17:19] Dave Pengelley: chat, there's [17:19] Richard Webbe: bit. [17:20] Dave Pengelley: a can you put chats on Matt on the right [17:22] Matt: able be do going not that. to to [17:22] Dave Pengelley: edge? [17:22] Matt: The [17:23] Dave Pengelley: Or is [17:23] Matt: only [17:23] Dave Pengelley: that just [17:23] Matt: way that we can do it as a [17:24] Dave Pengelley: no? [17:25] Matt: guest. [17:25] Dave Pengelley: You just gotta log into the actual YouTube platform, whatever, and [17:28] Matt: Yeah. [17:28] Dave Pengelley: comment out [17:29] Richard Webbe: I'll [17:29] Dave Pengelley: there. [17:29] Richard Webbe: send it. And you can add [17:30] Dave Pengelley: I'm [17:31] Richard Webbe: it in, [17:31] Dave Pengelley: like [17:31] Richard Webbe: Dave. [17:31] Dave Pengelley: He-Man, only I have the [17:33] Richard Webbe: It's [17:33] Dave Pengelley: power. [17:33] Richard Webbe: all magic in the background, team, when you're watching. Don't worry, [17:36] Dave Pengelley: I [17:36] Richard Webbe: it's [17:36] Dave Pengelley: mean, [17:36] Richard Webbe: not really happening [17:36] Dave Pengelley: not [17:37] Richard Webbe: well. [17:37] Dave Pengelley: AI related, but I saw the He-Man movie on Thursday and I loved it. I thought it was fantastic. [17:41] Richard Webbe: You know, I missed the whole, I'm too old. I missed, I saw [17:44] Dave Pengelley: Yeah, [17:44] Richard Webbe: He-Man [17:44] Dave Pengelley: yeah. [17:44] Richard Webbe: on the Kelly when I was older, [17:46] Dave Pengelley: You're [17:46] Richard Webbe: but that was for [17:46] Dave Pengelley: busy [17:46] Richard Webbe: younger [17:47] Dave Pengelley: working [17:47] Richard Webbe: kids. [17:47] Dave Pengelley: a job when kids like me were [17:49] Richard Webbe: You're [17:49] Dave Pengelley: watching. [17:49] Richard Webbe: super excited about [17:51] Dave Pengelley: Oh, [17:51] Richard Webbe: it. [17:51] Dave Pengelley: it was so good. It was just nostalgic and it was well done. And it was a I I I feel and there's different takes out there, but I feel like a respect. The source material [18:02] Richard Webbe: Well done. Sorry, some guy named Matt just put in the chat. I disagree with that red-headed guy. Should put on there who needs a haircut. [18:11] Richard Webbe: The nod. [18:13] Dave Pengelley: here. Um, [18:16] Richard Webbe: Now, Matt, oh, you're going to say more about you told me the books. We're going to look them up. We'll get the boy to put them in the chat for other people. But what else were you saying about this subject? Because it's [18:26] Matt: One [18:26] Richard Webbe: the most interesting, [18:26] Matt: more. [18:26] Richard Webbe: I think, of all of them. Go. [18:29] Dave Pengelley: I'm not sure what I'm [18:30] Matt: I believe it is a direct analogous parallel to this. You know, it's to do with uh psychology, influence, persuasion, like those Chieldini books. And this is magic, you know, like performance magic, magicians, um, with regards to the idea of specifically mentalism. [18:49] Richard Webbe: Matilism, yep, go on. [18:50] Matt: So think about what mentalism is and what they do. It's usually some sort of routine where at the end of it they say a thing that you've had in your mind, you know, or you say a thing and they they know. Knew it because you know their card has the thing that you said. I'm like, whoa, how did they write that down? [19:04] Dave Pengelley: I was I was reading an early draft of this book called How I Screwed Up Selling, and it was talking about how uh Sales guy stuffed up a relationship because he did this secret magic trick that uh really annoyed someone. [19:14] Richard Webbe: Yes. [19:15] Matt: Yeah, for so that that whole thing, you know, mentalism in general is is seeding thought. [19:21] Richard Webbe: Powerful. [19:21] Matt: It's it's adding bias, it's pre-suading, [19:24] Richard Webbe: Very [19:24] Matt: it's [19:24] Richard Webbe: powerful. [19:24] Matt: influencing. [19:24] Dave Pengelley: Richard, you do this, right? You've done this, you've got your card trick. [19:27] Richard Webbe: I do a yes, for those that haven't seen it. Some have, although I did go on a date once with a girl and try and do the card trick, and she already knew it, it really ruined my date night. But um, so the card trick I do is about using a series of questions to control the person listening, which is what magicians do, either image-wise or physics-wise. And we talk about selling as um herding cats or getting people to focus on what the value really is. So, salespeople, as Dave knows, we don't try and sell things that people don't want, but we do try and highlight what we think. Is valuable for them. And as we all do in anything, like with you taking your partner out to dinner and you want to go to a certain restaurant, you'll tell them how great a certain meal is that they like. That's at the restaurant. So it's all about changing priorities and changing value and changing perception of value. And that's where that what you just said earlier: the the what was it? Pre [20:22] Richard Webbe: suasion. That's how it works. And it's all the same stuff. And we've got to arm ourselves with, I think, perspective. And I was reading a book on Confirmation bias. And there were two ladies, um, I think Iranian, who were talking about AI and confirmation bias, and how when certain people were presented with absolute facts would still not change their mind. Because you've hung your whole persona, everything that's you about you, on this fact, and then someone tells you it's not true. Like I lived for years thinking that Simon LeBon was um uh uh who's Mr. Stick Together um the famous. Singer uh was his son. Uh I I've oh and now I've stuffed it up. Roxy Music. Who's the lead singer of Roxy Music? [21:11] Richard Webbe: Ah. [21:11] Dave Pengelley: Yeah, [21:12] Matt: Okay, [21:12] Dave Pengelley: yeah. [21:12] Matt: okay. [21:13] Richard Webbe: Anyway, he looks a lot like Simon LeBron, who was the lead singer of Duran Duran. And for years someone told me when I was living in London, oh that's Simon Leah Simon LeBon's his son. And for years, thank you. And for years I trade on the fact that absolute fact that Simon LeBon Brian Ferry's son, and it's not Not true at all. But anyway, that's life. I also thought Thomas Dobby was related to the guy who invented Dobby noise systems. And that's not true either. But it's very saddening when you get [21:39] Dave Pengelley: She [21:39] Richard Webbe: that information. [21:40] Dave Pengelley: blinded me with science. [21:41] Richard Webbe: That's it, thank you. [21:46] Richard Webbe: And then [21:46] Dave Pengelley: Matt's [21:46] Richard Webbe: we throw [21:46] Dave Pengelley: like, [21:47] Richard Webbe: away [21:47] Dave Pengelley: what [21:47] Richard Webbe: lines. [21:47] Dave Pengelley: am what [21:47] Richard Webbe: So [21:47] Dave Pengelley: am [21:47] Richard Webbe: Matt, [21:48] Dave Pengelley: what [21:48] Richard Webbe: where [21:48] Dave Pengelley: am I [21:48] Richard Webbe: is [21:48] Dave Pengelley: doing here? Matt says Matt. Matt's like, What is what is happening? [21:50] Richard Webbe: Matt? Sorry, we're yeah, you were still not even a twinkle in someone's eye when we were listening to this stuff. Um, you're going on to say some more stuff about confirmation bias and [22:00] Matt: If I'm if I'm quiet, it means I'm in just enjoying it. I'll s I'll I'll step over you voc vocally if I need to add something in. [22:08] Richard Webbe: Might be good for us three people listening might not be instead of [22:11] Dave Pengelley: My [22:11] Richard Webbe: using. [22:11] Dave Pengelley: my kids know they are blinded blind. She blinded me with science. I'm like, no, not that again. [22:18] Richard Webbe: I have to meet your kids. Having been to your house last week for our conference, which I thought went very well and I appreciate being involved. Um yes, it's a sign of a very happy, productive and interesting family. [22:31] Dave Pengelley: Yeah, I was impressed yesterday. My daughter, my 11-year-old daughter out of nowhere, just asked um the smart speaker to play um uh Stone Sour, like um through glass. And I was like, what? Like she loves Shine Down, she loves all these things. I was like, it's great. I'm I'm doing something right. [22:49] Richard Webbe: Well, [22:50] Dave Pengelley: Likes [22:50] Richard Webbe: yes. [22:50] Dave Pengelley: rock. Um, [22:51] Richard Webbe: So [22:51] Dave Pengelley: although [22:51] Richard Webbe: then [22:52] Dave Pengelley: then a nickelback song came on, and I was like, just so you know, like it was um this is how you remind me. Like the radio hits are great. And I said, just so you know, don't just ask. [22:59] Richard Webbe: do going have that. to to we're [23:00] Dave Pengelley: From Nickelback songs. So she's like, What? Who are they? I'm like, they're the ones playing this song now. So they've got some spicy lyrics like saucy, [23:05] Richard Webbe: Yeah, [23:06] Dave Pengelley: but [23:06] Richard Webbe: there [23:06] Dave Pengelley: no [23:06] Richard Webbe: you [23:06] Dave Pengelley: no nickelback [23:06] Richard Webbe: go. [23:07] Dave Pengelley: for you. She's like, Oh okay. [23:08] Richard Webbe: So, Matt, in this subject of confirmation bias or political bias or control bias in AI, particularly at the start of LLMs, we can probably remove it or introduce it when we're building our own rags and our own isolated edge LLM or edge AI systems. What do you think is the answer? [23:30] Matt: what how to remove [23:31] Dave Pengelley: 42. [23:31] Matt: it. You you can't. You the only [23:35] Richard Webbe: Great. [23:35] Matt: thing you can do the only thing in you can do is add lenses. [23:39] Dave Pengelley: Yeah. And and I mean more more and more of these systems and tools, and we discussed it last week or the week before how OpenRouter have their fusion model, whatever, which will like link two or three different models together. Um, Hermes is now starting to push the mixture of agents as a tool you can use. More and more of these harnesses are now actually going. Going, you know what, if you want the best answers, maybe you don't just go to a fable or a mythos type thing, just get three of the really good ones and get them to fight. [24:10] Matt: Yeah, [24:10] Dave Pengelley: So [24:10] Matt: and all of this comes back to the loops and validation and back pressure. So everything is a cycle here. Everything is the same thing, just in different flavors. [24:19] Dave Pengelley: let's let's speaking of some different flavors and some new new toppings of the local LLM and Larry's, let's do this. [24:28] Richard Webbe: You think yourself? [24:35] Dave Pengelley: No [24:35] Richard Webbe: Sorry. [24:36] Dave Pengelley: more in the news. [24:38] Richard Webbe: In [24:38] Dave Pengelley: I've [24:38] Richard Webbe: any [24:38] Dave Pengelley: run [24:38] Richard Webbe: case, anyone [24:38] Dave Pengelley: with a [24:39] Richard Webbe: can hear the [24:39] Dave Pengelley: much [24:39] Richard Webbe: noise [24:39] Dave Pengelley: more [24:40] Richard Webbe: school [24:40] Dave Pengelley: professional. [24:40] Richard Webbe: holidays is on in Melbourne and I've got I've got children climbing up the stairs. Sorry, go on. So [24:44] Dave Pengelley: That's [24:45] Richard Webbe: the can [24:45] Dave Pengelley: right. [24:45] Richard Webbe: I did [24:46] Dave Pengelley: Did you like [24:46] Richard Webbe: you [24:46] Dave Pengelley: our [24:46] Richard Webbe: like [24:46] Dave Pengelley: new [24:46] Richard Webbe: any [24:46] Dave Pengelley: network [24:47] Richard Webbe: demo [24:47] Dave Pengelley: news to [24:47] Richard Webbe: news [24:47] Dave Pengelley: you? [24:47] Richard Webbe: together? I did, it looked great. It was great. I was busy screaming at the children, but no, it was great. [24:54] Richard Webbe: I have news. I have news. You guys have always got news. I've got [24:58] Dave Pengelley: Oh [24:58] Richard Webbe: news. [24:58] Dave Pengelley: please, please do. [24:59] Richard Webbe: So So the increase in hardware sales, based on the architecture that you both have described before, and how agentic architecture is changing the landscape of computers and how we process systems. People like Lenovo and Dell and HP are reporting market increase in hardware sales in a world that has become so SaaS in love with SaaS and offline on-site hardware purchases to grow your AI system footprint and come. Capability and keep the costs from tokenomics down is really taking off. [25:36] Matt: If I had the cash now, I would already have it at home. And I've priced everything up regardless of availability. To get near state of the art. So let's say you are running GLM 5.2 or Kimi 2.7, those kinds of models, which are quite large, you need to have somewhere in the realms of 50 to 100,000. [26:01] Matt: And then the [26:02] Richard Webbe: Yes. [26:02] Matt: power to supply that. [26:05] Richard Webbe: And [26:05] Matt: So [26:05] Richard Webbe: then you go and buy a big a big GPU and off you [26:08] Matt: and [26:08] Richard Webbe: go. [26:09] Matt: that now, you know, when somebody uh when you're in the shower and somebody flushes and then the shower kind of like dwindles. The same thing is true with this, you know, setup, with this hardware, this inference that you run yourself. That is only good enough for say like one, maybe two users. So if you're someone who usually does parallel sessions, all of a sudden everything's gonna slow right down. And you know, let alone multiple users itself. So you know that's then probably another another fifty or a hundred thousand. So it's achievable, it's just very expensive still. [26:38] Richard Webbe: Seventy-eight percent of uh businesses across the board have reported injecting a new lens of ROI into their LLM AI usage and their system process usage. And while no AI, the system usage is way too expensive, they start injecting AI into it and they're learning how to streamline and automate and save time. time and cost and money. And now they're looking at the tokenomics and the expense of using AI. And as David David's mentioned, for Ed yourself, it's blown out of the water. And now they're going back to buying more hardware. And we're talking millions of dollars of hardware across the board or billions. And that's lowering the cost of the AI while still lowering the ROI on the total system. So we're getting efficiencies and performance benefits, but like with client server in the old days, the the spend is shifting back to a centralized processing capability. capability. [27:31] Matt: Yeah, I agree. And I I I'm really excited to see where it goes. I I really like this whole shakeup. I think it's very, very important. And uh, there's a there's a an acronym that is used a lot on X. Uh it's probably on available on other social medias as well. And uh it is N GMI. Have you seen that before, Richard? You know what that [27:56] Richard Webbe: I [27:56] Matt: means? [27:56] Richard Webbe: have, but I do know what it was one of the questions I had for you guys. I saw it in I think I saw in G M I on the weekend. What's it mean? [28:03] Matt: Dave? [28:05] Dave Pengelley: I know NGL, I'm trying to look at what MGMI is. [28:08] Matt: So it's it's a it's a phrase that basically sums up what you've just said, Richard, and the people that are resistant to learning and discovering and experimenting and being curious about this and what it will bring, they [28:24] Matt: NGMI, which it literally stands for not gonna make it. [28:30] Dave Pengelley: Different from NGO, which is not gonna lie. [28:33] Matt: Yeah. [28:33] Dave Pengelley: Yeah. [28:34] Matt: Yeah, not gonna make it. So it's a phrase, it's a comment, it's a claim that comes out a lot with when people get a bit stuck and they're spinning their wheels, or they're resistant to certain things, um, where the writing is clearly on the wall. And you know, that kind of stubbornness, that head in the ground ostrich, you know, and GMI. [28:55] Dave Pengelley: Well, yeah, the the the the the the Luddite kind of theory, right? And people are like it's really interesting. Have you guys seen on Amazon the show called Upload? I don't know if I mentioned it before. [29:03] Richard Webbe: I saw the first series. I haven't seen the second [29:06] Dave Pengelley: It's [29:06] Richard Webbe: one. [29:06] Dave Pengelley: it's really good. Like it actually takes a bit of a critical lens to the future of virtual reality and and AI and things and where it can go if the bad people do bad things with it. And it's just and it's funny, it's like a sitcommy kind of it's more comedy than drama, but it does have that sort of slight drumedy kind of edge to it, like with a this sort of underlying plot line through the whole thing. Um, but it you know, it looks at these people that before they die, they upload themselves into a virtual world as sort of like their own live forever and then their relatives can still FaceTime them because now they're living in this virtual server world, but it's the haves and the have nots. And then you've got the the the sort of Luddite people that are refusing any technology and there's sort of harvesting things and living offline. And yeah, interesting interesting perspective on a fut on a future world [29:52] Richard Webbe: I thought [29:52] Dave Pengelley: where [29:52] Richard Webbe: it was a great [29:53] Dave Pengelley: these [29:53] Richard Webbe: show. [29:53] Dave Pengelley: things [29:53] Richard Webbe: I [29:53] Dave Pengelley: are [29:53] Richard Webbe: I enjoyed [29:54] Dave Pengelley: going. [29:54] Richard Webbe: it. Yeah, I agree. [29:55] Dave Pengelley: Yeah. Anyway, [29:56] Matt: Would [29:56] Dave Pengelley: funny, [29:56] Matt: you say [29:57] Dave Pengelley: funny [29:57] Matt: it was? [29:57] Dave Pengelley: funny stuff. [29:58] Matt: On on Amazon or something. [29:59] Dave Pengelley: Uh Yeah, think it's a prime show. [30:01] Matt: All right, I might still have Amazon because we're watching the boys, and that's on that. [30:05] Dave Pengelley: Yeah, [30:05] Matt: So [30:05] Dave Pengelley: yeah, upload. It's [30:06] Matt: all [30:07] Dave Pengelley: um [30:07] Matt: right. [30:07] Dave Pengelley: got oh the guy who is the brother of the guy who is arrow. [30:15] Dave Pengelley: Robbie someone. Anyway, good show. I [30:18] Richard Webbe: We'll [30:18] Dave Pengelley: enjoyed [30:18] Richard Webbe: get Richard [30:18] Dave Pengelley: it. [30:19] Richard Webbe: Wilkins on next week to answer all these questions for her. [30:22] Dave Pengelley: From MTV. Um showing my age. All right. [30:30] Dave Pengelley: Okay, okay, let's let's let's I've got I've got some news links links to that. Okay, so in the last week, since we last spoke, um new models have been announced. OpenAI came out today. We're introducing a limited preview of the GPT 5.6 model family, Sol, Terra, and Luna. Uh and [30:47] Richard Webbe: Wow. [30:47] Dave Pengelley: so they're saying Sol is their new, super expensive flagship model, which is like the best ever, ever. Uh Terra is at least as good at 5.5. [31:00] Matt: At [31:00] Dave Pengelley: About [31:00] Matt: half [31:00] Dave Pengelley: half [31:00] Matt: the [31:00] Dave Pengelley: the [31:00] Matt: price. [31:01] Dave Pengelley: cost. And Lunar is almost as good as five point five and much cheaper. [31:05] Matt: I I have a sneaking suspicion that Luna is in early preview at the moment in ChatGPT as their 5.5 instant [31:13] Dave Pengelley: Oh [31:13] Matt: model. [31:13] Dave Pengelley: really. [31:14] Matt: Yeah. Because 5.5 instant had an update and it's actually really good. Like I don't use the ChatGPT app much, but every now and then you just want to have like a quick, you know, chat based interaction question [31:25] Dave Pengelley: Yeah, yeah. [31:26] Matt: or something. And um yeah, it it's meaningfully. Different. [31:30] Dave Pengelley: Yeah, although instant, I don't know, I found sometimes instant was a bit lacking because I was using a little bit for like some of the new media assets and just back and forthing a few images and stuff, and I was like, eh, I think I'll turn thinking back on. Um, but [31:44] Richard Webbe: I I do find it interesting how such a burging industry has managed to accelerate its segment marketing capability. [31:53] Matt: Restrict it so much [31:54] Dave Pengelley: yeah, [31:55] Matt: at [31:55] Dave Pengelley: well, [31:55] Matt: the same [31:55] Dave Pengelley: well, this [31:55] Matt: time. [31:55] Dave Pengelley: is the thing. So um it's on uh a limited preview, and so you read. The article and they go after the whole um anthropic issue with them releasing Fable and then having to be pulled 72 hours later. They've gone, oh, we're now doing this limited preview with some government agencies and some uh some big tech firms just to make sure everyone's happy with it before we put it out in the wild. Um, because it's you know seen as that same sort of fable class model. If we go to this benchmark post, breaking open I just took the crown back, 5.6 solars live. Soul Ultra 91.9. Base Soul 88.8. Mythos 5 88. Fable 84.3. So it's smashing [32:37] Matt: So [32:37] Dave Pengelley: it. [32:38] Matt: base [32:38] Dave Pengelley: Soul. [32:38] Matt: base soul is the same as what Mythos was, which, you know, Mythos was obviously just a little bit higher than Fable being [32:45] Dave Pengelley: Yeah. [32:45] Matt: not so bottomized. [32:47] Dave Pengelley: Um, and Opus 4.8 all the way down here in the 70s. Oh. Even Luna. Luna is beating Opus 4.8 apparently on this particular one benchmark metric. And uh, I don't know exactly what that's met. Measuring, but anyway, they're just saying that these new models are pretty good. And that has caused them to go, Well, we are conscious of US regulators. We want to make sure we don't get this pulled and go through all of that drama. Um, and it's got some people, and it's funny, you talk about the hardware thing, um, Richard. I saw this guy, he's one of the the hype merchants on YouTube. Good for him, that's his business model, Alex Finn. And he wrote, It's never been more important to learn about local AI, and he was fully blackpilled about. I can't believe this is like a democratic freedoms issue that we can't get access to these models straight away. And I'm like, I think you misunderstand, like in how markets work and that private companies don't have to give you this entitlement syndrome that there's a new model. I have to be able to access it. This is a restriction of my freedoms if a company won't allow me to access this thing that I've built. I'm like, I don't quite agree with that point of view. And he's fully getting to this whole personal sovereignty, and he's got multiple Mac Studios. Goes and he's built a an RTX 9000 NVIDIA thing that's costing him tens of thousands of dollars. And doing what you're talking about, Matt, where he's building up this whole thing. He's like, okay, on my DGX Spark, I'm running a full GLM 5.2 model, but it's slow. And then I've got these other ones which have like the NVIDIA doesn't have as much VRAM. And so we can run this on that. And on the studio, I think he's he's got different models on different hardware because just the way the RAM restrictions work and the the data pipelines, there's sort of some are better for. Others than the others based on bandwidth and size. So anyway, he's building out his full local lab, and [34:35] Richard Webbe: Yeah. [34:35] Dave Pengelley: he's going on about how important it is you've got to be own your own AIs and this and that and the other. And this was a few days ago when he's like fully, fully like gone off about this whole I can't believe they're not sharing these models. Uh what a travesty. [34:50] Richard Webbe: So [34:51] Dave Pengelley: And [34:51] Richard Webbe: he's [34:51] Dave Pengelley: then [34:51] Richard Webbe: changed his confirmation bias. [34:54] Dave Pengelley: sorry. [34:55] Richard Webbe: He's changed his confirmation [34:56] Dave Pengelley: Well, well, [34:57] Richard Webbe: bias. [34:57] Dave Pengelley: no, but but then then okay. So so next next next new. Article that's popped out today is Claude Sonic 5 has come out, our most agentic Sonic yet. It makes plans, uses tools like browser terminals, runs autonomously, blah blah blah. So Claude's brought out their new cheap model. They're trying to get in and steal some of the lunar uh terror thunder. And uh then you go back to Alex's page and you see actually incredible, Claude Sonic just dropped. I'm changing how I use AI. So it's like three days ago, he's all like, You can't rely on these frontier models. I can't believe that. You've got to go all local, all do it yourself, all run your own home labs. And then he's like, well, Sonic 5 is incredible and it changes the way I think about AI for the future. And I'm like, oh, come on, [35:39] Matt: That's [35:39] Dave Pengelley: come [35:39] Matt: why [35:39] Dave Pengelley: on. [35:39] Matt: we exist, man. That's why this podcast exists. Yeah, it's not as dopamine rich [35:44] Richard Webbe: Oh [35:44] Matt: as [35:44] Richard Webbe: it's [35:45] Matt: watching [35:45] Richard Webbe: no [35:46] Matt: Al [35:46] Richard Webbe: it's [35:46] Matt: Alex. [35:46] Richard Webbe: changing quickly. [35:47] Matt: Yeah. [35:47] Richard Webbe: I had [35:48] Matt: Yeah. [35:48] Richard Webbe: uh I had uh uh quick coffee with one of the leaders of one of the world's largest optical manufacturers and executive from Corning, and um they were telling me how their sales are just through the Roof for everything from long distance to the short distance loop cables on the back of the GPU for optical fiber because of the increase in the amount of data we're going to be exchanging. So neural networks are just, you know, they're accelerating at a great rate. And of course, uh, you know, the quantum computing uh can you can we physically buy quantum computers off the off the shelf yet? Or are they all we're not we're not quite there yet? Okay. [36:30] Dave Pengelley: There's a lot of development in Australia um on on it, but you can't just go down to Harvey Norman and buy yourself a quantum laptop yet. [36:38] Matt: It's in all these more bunnies. [36:41] Dave Pengelley: M Wave, M Wave have them apparently, I think. AliExpress. Get it from China, the enemy, the enemy, apparently the bad guys. The bad guys have them. [36:51] Richard Webbe: But who would have thought a few years ago we could get the power of a GPU in a little laptop today? [36:58] Dave Pengelley: Well, [36:58] Richard Webbe: Someone [36:58] Dave Pengelley: I mean [36:58] Richard Webbe: would have [36:58] Dave Pengelley: that's [36:59] Richard Webbe: said. [36:59] Dave Pengelley: that's what we saw. And then Jensen Huang, we spoke about this the other week. Uh, he the NVIDIA's got their new system on a chip. It's kind of their version of Apple Silicon for Windows users, really, where they've got all the processor and the RAM and everything in a single entity. So from those data pipelines and the bandwidth restrictions, they're not there because it's traveling micro microns rather than you know millimeters for it to get data from one place to another and [37:24] Richard Webbe: Yeah. [37:24] Dave Pengelley: it's all embedded and fully integrated. Um so yeah, I mean Moore's law, I think, is still holding up, but quantum breaks that um [37:36] Richard Webbe: It does. [37:36] Dave Pengelley: and step [37:37] Richard Webbe: Yeah. [37:37] Dave Pengelley: change. [37:37] Richard Webbe: Yeah, it it does, it does [37:39] Dave Pengelley: Um, [37:39] Richard Webbe: indeed. [37:39] Dave Pengelley: which everyone's worried about, how it's gonna change encryption standards and everything. [37:45] Dave Pengelley: Um, but uh on the on the sonnet thing, I mean, Matt, you were saying you hadn't seen a lot of positive feedback around Sonnet so far, and obviously Alex is very positive about it. But I did see this other post that I thought you'd appreciate as well. Um Okay, I can't believe I'm gonna say this, but Sonic 5 Max is too high effort. Like, damn, they ain't playing around if it's like Sonic 5 Max effort. It's like giving a box of squirrels a bunch of cocaine and saying, go with God, and just seeing what comes out the other side. [38:12] Richard Webbe: Oh, that is funny. [38:16] Matt: Yeah, I should have I should have prepped some links for you, Dave. Like I've got so much. Um so artificial analysis, the guys that do the probably the most relevant series of bench aggregation that exists, they they Clocked Sonnet 5 costing a lot more than Fable did on their actual badge um, their actual intelligence index, they call it. So, yeah, and that's because it does that, what Shapiro just said, you know, it just spins up so many loops of uh, you know, need to do this. And then the the the amount of reasoning that it does is similar to Fable, but the amount of actual loops and checks, you know, like the squirrel logic. The funniest part, and I genuinely say this uh in seriousness, but in [39:07] Richard Webbe: Oh, [39:07] Matt: jest, you know, the bad guys. Um [39:09] Richard Webbe: yeah. [39:10] Matt: GLM 5.2 has been out for a while now. It is very good and it is still better than Sonnet 5. And it's cheaper. So those in the Claude ecosystem, sure, you know great, Sonnet 5 is a meaningful upgrade from Sonnet 4.6, as long as you don't need to worry about squirrels. But if you were choosing, if you had zero like loyalty to anything, there's no reason why you would choose son of over GLM still. [39:38] Dave Pengelley: Well, I mean, here's here's a very real example with it's just from my own noose portal, uh, where I was getting um to I use a bit of Opus 4.8 the other day, like paying API pricing, and that's the red line there. And this number of tokens, and then I also use some GLM 5.2, and so you can see I use more GLM 5.2, and I've used some more today. That's the total tokens. If I look at the space Spend comparison [40:02] Richard Webbe: Wow. [40:03] Dave Pengelley: Opus twelve dollars versus like two bucks. [40:07] Richard Webbe: Yeah. [40:08] Dave Pengelley: Like for the side for like GLM was more tokens and it's like a fraction of the price. [40:14] Richard Webbe: Well, I had uh I had Claude analyze the ROI for me across, you know, LLM usage with ma maximum agentic leverage and rag. And of course, all it is is the line between what you're using it for, whether it's personal productivity, system productivity. competitive advantage where you're pulling your data from and how you have your data organized and it's it's really interesting but the top end uh I had it tell me up to 300% ROI if you shift the line between where you're doing local AI and LLM calls [40:48] Dave Pengelley: Yeah, I mean and when you say local AI, do you mean local models or just deterministic like back to programmatic stuff like we talked about on the um the [40:56] Richard Webbe: well [40:56] Dave Pengelley: like short? [40:57] Richard Webbe: yeah well it depends on your needs, right? Whether it's personal productivity, enterprise knowledge assistant or system uh competitive advantage and where you're mining your data from. So it's that whole triangle is it, you know, where you're doing your processing, where you're doing your calculation, and where's your data and [41:13] Dave Pengelley: Yeah. [41:13] Richard Webbe: how well organized is your data to answer those questions or provide that success. [41:18] Dave Pengelley: I mean, as as we've been discussing, I I was heavy into co work and had all my daily briefings and had all these sort of routines and systems running, but it was phenomenally expensive to run because the agents were doing all the work and That exact scenario we talk about not doing where you just have the agent doing calls back and forth, back and forth, reprocessing stuff. And the you know, I fit inside my subscription usage. So I was like, who cares? But once I ditched that and went to having to pay for actual things or using other models, um rebuilt a lot of those workflows to be scripts. So and Matt and [41:49] Richard Webbe: Okay, [41:49] Dave Pengelley: Matt, you'll love this. So then now now the cron runs a Python script which goes and fetches a bunch of data into some static files, and then [41:57] Richard Webbe: I [41:57] Dave Pengelley: once we've got all the data fetched with no agents involved, [41:59] Richard Webbe: a good idea. that's think [41:59] Dave Pengelley: no AR whatsoever, then we run a specific call to an agent passing it those inputs and asking it for a certain output, which then it spits out and then scripts format that into HTML, not an agent, [42:11] Richard Webbe: Yep. [42:11] Dave Pengelley: uh, which then sends that that daily summary email to me programmatically with checks and verifications, not relying on agents to hit errors and then debug and loop and fix it and then try again and waste tokens. [42:25] Richard Webbe: And [42:26] Dave Pengelley: So [42:26] Richard Webbe: that goes back to first principles. It's all about knowing the question you want to. Ask before you actually ask it. [42:32] Matt: This sounds a lot like what I teach in my school. [42:35] Richard Webbe: It is. It is. [42:38] Dave Pengelley: Right. It's [42:38] Richard Webbe: Sorry, [42:38] Dave Pengelley: and [42:39] Richard Webbe: sorry. [42:39] Dave Pengelley: it is. And as we talk about the token rug pull coming, etc., all these things, you've got to think about where you spend your tokens, how you spend them. I experimented with Minimax M3, and I've actually dropped that sub after a month because it just, you know, that reasoning thing just kept circling and double-checking itself and over-verifying. And I just wasn't enjoying the experience. Um of long-term usage with Minimax M3. I've got my Grok subscription. I'm using Composer 2.5 fast. That seems pretty good as a daily driver model at the moment. But I'm excited for the new OpenAI models like Terra and Luna to give them a crack as my Hermes daily driver agent and see how quickly they burn through my codex usage and if that's sustainable. But this mix of mix of models, future, I think, and you know, as we are going to have less subs and have to pay more. Per token, we're gonna have to be really thoughtful uh around. And when they go, you know, SaaS is dead, software's dead, people can code anything, vibe code it all. The party's gonna be over pretty soon where you just it's not cost effective to do that, and you go back to hiring a developer who is augmented by these tools versus replaced. [43:51] Matt: There's a lot here that is very, very similar to manufacturing. Now now [43:57] Richard Webbe: Oh, [43:57] Matt: I've never [43:58] Richard Webbe: extraordinary. [43:58] Matt: been a in manufacturing. [44:01] Matt: But I've had a few good chats with people and um I love speaking to people like that. I love speaking to actual, legitimate real engineers, you know, people that are like business engineers. You know, I I would love nothing more to just sit in rooms with those people all day. [44:17] Richard Webbe: I [44:17] Matt: Um [44:17] Richard Webbe: had a friend Matt, I had a friend who was uh I think an Australian leader in what they call just in time manufacturing for reducing supply chain. And uh I also was involved in installing the BHP IT. Steel supply chain system, which was one of the best in the world. And it's exactly what you're saying. It's building what you need at the time you need to build it without wasting your time doing other things, similar to what Dave was just saying, making all these spurious calls that you don't need to do. And you're right, I think it is exactly like manufacturing. Where's the most efficient way to process that information before I get to the final answer? I think that's a good description. [44:55] Matt: Hmm. I have a little like statement here that my my agent helped prep for me for this episode. Because it's all the, you know, this everything that I do is an amalgamation of my my voice as well through my systems. And this this concept is basically literally the one liner that it gave me, um, which verbatim, it says the interesting thing apart, the interesting part about the models getting smarter is that access, audit, and control are becoming the product. You know, there's where there's two sides, there's the chat layer and then there's the ops layer. So, [45:28] Richard Webbe: Yeah. [45:29] Matt: you know. The whole manufacturing side of things is that ops layer, which is what Dave was just talking about, you [45:35] Richard Webbe: Yeah. [45:35] Matt: know, actually creating the correct pipelines, the correct deterministic workflows. [45:40] Richard Webbe: Yeah, when should I do this? When should I do that? Absolutely. [45:42] Dave Pengelley: Yeah. [45:42] Matt: And only using AI where you need [45:45] Dave Pengelley: Right, [45:45] Matt: to. [45:45] Richard Webbe: Yeah. [45:45] Dave Pengelley: even even for this show last week, and this is one of the things that drove me away from M3. When I finish this show, I I download and tag the transcript with who said what and pass that through, and it writes the new title for YouTube that gets rid of the generic first of July AI operator. That you see right now, and put gives it you know a episode title, puts a description, puts all the chapter markers, and I had skills that go through and they know process this summary, create that, do that, and then push to the API endpoints, update YouTube to update the website because we put the episodes on the website too. And I was just going around in circles and kept complaining and saying, No, you need to verify this. No, I can't do that. Oh, oops, I didn't escape that character, and and so the call failed. And so now I've I haven't tested it. Yet, because I haven't done a new episode since. But I've got it so that we'll output the summaries and then hand that to a script which will check for those escape issues, which will check for those things, and then programmatically, deterministically push the update up to the APIs versus relying on the agent to do that. The agent can do all the summary, it can do all the prep, it can actually generate the actual data, but it's no longer responsible for pushing it up to the API. [46:54] Richard Webbe: So it's an algorithm [46:55] Dave Pengelley: I can't trust [46:55] Richard Webbe: versus an [46:56] Dave Pengelley: it. [46:56] Richard Webbe: LL. [46:58] Matt: What was that statement, sorry? [47:00] Richard Webbe: An algorithm, an automated algorithm versus using LLM. It's not deterministic, it's just a known process. [47:09] Dave Pengelley: Well yeah, because because Python Python can run run a script across a string and go, is there a character like is there a quote character in here that doesn't have a backslash next to it to make sure that we just read it as text and not an actual quote character, which means something. And so it can deterministically run through and check for those sorts of little data cleansing errors that will stuff the API. Whereas the agent goes, oh no, I got this. Oh no, maybe it was this. Oh, I'm gonna try this. Oh, let's do this. Oh, that version didn't work. Let me try again. And then it goes, oh, I've run out of turns and I've spent a million tokens and I didn't get anywhere. Versus, you know, just generate the summary data I need, however it generates it with quotes and columns and who knows who what, who cares, dump that into a text file, which then my script will pick up and go, okay, I know how to clean this up and make sure it's ready. It's like if you've ever cleaned up a Word document where people use multiple carriage returns. And or and or spaces [48:01] Richard Webbe: Yes. [48:02] Dave Pengelley: instead of tabs and things like that, and you just find and replace, you [48:04] Richard Webbe: Yes. [48:05] Dave Pengelley: programmatically know okay, if I need to clean this document up, I need to look for every instance of two carat Ps next to each other and replace with one carat P. And then I'm going to put some paragraph spacing in. Maybe yeah. [48:16] Richard Webbe: Syntaxing, [48:17] Dave Pengelley: So [48:17] Richard Webbe: yep. [48:17] Dave Pengelley: you just tell the scripts can do that without needing an [48:20] Richard Webbe: Yep. [48:20] Dave Pengelley: agent to work it out dynamically every time. [48:23] Richard Webbe: Yeah, exactly. Yeah, so that's a a saving, absolutely. [48:27] Dave Pengelley: So [48:27] Richard Webbe: Um, go on. [48:28] Dave Pengelley: yeah, no, so so. I mean, and this is this is what we we're talking about. Is as the costs of AI come home to roost, people need to work out what is an agent. When do you need that interactive experience? When do you need that sort of thinking with tools capability? When do you need it to just be a link in the chain that does some work and passes it back so then the chain can continue? Um, and when are you just using it to build another tool that has no AI? Like my app, chartreporter.com, is not an AI tool, but I used AI to build a tool uh that is then Of this is a program that I used AI to build a tool to fix a problem. [49:04] Matt: I love that too. I think that we've said that in a previous episode that you don't necessarily need AI to do the AI. You use the AI system to build the tools that you'll actually be running. [49:14] Richard Webbe: Yeah yeah so we're we're talking cost offload and what I call about my concentric ru ring rule where is it most expensive to do what you need to do and where is it most cost effective and and I think that's going to be a constantly changing landscape because I've been As soon as the uh AI LLM providers see that their calls are dropping and not growing as much as they want, they now realize they're in competition with local hardware processing and they'll change their economic model to suit. Or it'll be more like uh all-in-one black box theory, people will start shipping boxes. [49:49] Matt: Yeah, well, like the hardware thing is actually really, really important for people like me, like running a lot of stuff. You know, I'm not running my own AI locally at this point, but I'm still doing a lot of process. Processes on my machine, you know, where I'll have, you know, maybe like six or so different projects that I'm actually working on for various personal or professional reasons. And each of those may have multiple AI sessions all chipping away at it. You know, all of a sudden now my system is CPU pinned, RAM pinned, and I'm on the verge of capping out my what's called the swap, you know, with regards to storage space that can be accessed. And oh, my machine. Machine is at its limit. And [50:32] Richard Webbe: You can cook an egg on it, and then the next consideration that comes in is power. [50:37] Matt: yeah, exactly. So, like this whole like hardware shortage thing is actually really affecting me and on flow affecting my clients because I can't deliver as fast as I should be able to. You know, I went to look at actually upgrading my machine the other day, and directly from the Apple website, like if I go to order a new Mac, um, six to ten weeks, and and the prices are like 50% more than what they were like maybe three, four months ago. [51:04] Dave Pengelley: Wow. [51:04] Matt: It's crazy. Like I've seen a few people who like they picked up an M5 hundred and twenty-eight gigabyte MacBook um Pro when it came out, you know, like the the newest top of [51:13] Dave Pengelley: Yeah. [51:13] Matt: the range fully kitted out thing. They spent probably like, I don't know, six to seven US on that, and now they're about 10. Like [51:21] Dave Pengelley: Wow. [51:21] Matt: it's just crazy. [51:24] Dave Pengelley: Yeah. It's I mean I move from my Mac minis and my daily driver to my ACS ROG because it had a NVIDIA. A chip in a and just generally was a beefier machine than my little Mac Mini, my old M1 Mac Mini, uh, which only had eight gig of RAM. So I moved to my ACES because I was hitting limits. Like and back then I was heavy than things like anti-gravity, which was a massive memory hog anyway, on on either platform. Um, but you know, like yeah, that's that's real when you're running multiple things and you start setting up uh automated processes, so they just start randomly in the background and running their own little processes and checking things and you're like, oh man, don't launch that thing now. Like why? [52:00] Richard Webbe: Hey, just changing tax slightly, but on that point, sitting as a non-technical sort of solution architect person that I am trying to solve problems for my customers now as well. Can I get some of those programs and just reverse engineer them in vibe coding and the AI models that we've got today? [52:20] Matt: No, I'm gonna say I'm gonna say no because of your understanding of the whole process. Um [52:28] Richard Webbe: Right. [52:28] Matt: you can But can you right now with your current knowledge set? No. [52:32] Richard Webbe: No, that's okay. So assuming that I'm not an idiot, [52:36] Matt: No, [52:36] Richard Webbe: which [52:36] Matt: that's [52:37] Richard Webbe: I [52:37] Matt: not what [52:37] Richard Webbe: am, [52:37] Matt: I'm saying. That's [52:37] Richard Webbe: I think [52:38] Matt: not what I'm saying. [52:38] Richard Webbe: I know what you're saying. So, but if if I had the perspective and context of knowing the variables I want within that program, and I was experiencing it using one of these, I can just go in and [52:51] Matt: 100%. So if you can speak the lingo that it needs to do the job properly, 100%, this then becomes a really long project because You can't just do it instantly with one shot. [53:02] Richard Webbe: No. [53:02] Matt: So therefore, you need a really good um state engine for this. You need a good job queue, you need a good repeatable system and like [53:11] Richard Webbe: Sandbox [53:12] Matt: algorithm. [53:13] Richard Webbe: testing, all of that stuff. So [53:15] Matt: Yeah. [53:15] Richard Webbe: I hunt. [53:18] Richard Webbe: Thank you. [53:19] Dave Pengelley: I [53:19] Richard Webbe: Well, [53:19] Dave Pengelley: I [53:19] Richard Webbe: we're [53:19] Dave Pengelley: refer [53:19] Richard Webbe: going to [53:20] Dave Pengelley: I refer you to my previous efforts to try and rebuild such a thing and uh understanding how complex these tools like Claude Code, Hermes, etc. [53:29] Richard Webbe: do have that. to [53:29] Dave Pengelley: are. As far as their what they do deterministically. And that's what a lot of people I was posting. People saying someone else on LinkedIn was going, oh, the model's not the moat anymore. I was like, yeah, I know, we've been saying that. And then I talked about harnesses. He goes, Oh, it's not just the interface, but I'm like, no, it's not just the interface. Don't don't don't misquote me. I'm talking about the whole harness, the underlying instruction set and the prompt layers that it puts on top of anything you give it, and the way the deterministic programmatic handling, when you give it commands and things really [53:57] Richard Webbe: So I'm I'm I'm I'm even extending beyond Our AI tools that we like to use that we can I'm talking about in the monolithic software space [54:05] Matt: Yeah, [54:05] Richard Webbe: like [54:05] Matt: it [54:05] Richard Webbe: FAP [54:05] Dave Pengelley: Yeah. [54:06] Matt: was talking [54:06] Dave Pengelley: And [54:06] Matt: about. [54:06] Dave Pengelley: you yeah, [54:07] Richard Webbe: or [54:07] Dave Pengelley: you you you can build anything, but but you know it comes back to you know I was we we were so so excited thinking about whether we could do something. Nobody stopped and asked, should [54:18] Richard Webbe: when you [54:18] Matt: So [54:18] Richard Webbe: do something, [54:19] Matt: this [54:19] Richard Webbe: yes, [54:19] Matt: like [54:19] Dave Pengelley: should [54:19] Richard Webbe: I'll [54:20] Matt: Richard, [54:20] Richard Webbe: do [54:20] Dave Pengelley: should [54:20] Matt: this [54:20] Dave Pengelley: we create dinosaurs? [54:22] Matt: Richard, this is actually what I do. This is that that's what my professional business is. It's creating custom solutions and those [54:29] Richard Webbe: it. [54:29] Matt: custom solutions. Solutions are quite often a case of somebody coming to me and saying, Hey, we currently pay for this tool, we only use five percent of it. Can you rip replicate that five percent and create our own stack or add it to this tool for us? And [54:43] Richard Webbe: You are [54:43] Matt: I [54:44] Richard Webbe: Ben [54:44] Matt: say [54:44] Richard Webbe: Affleck, aren't you? [54:45] Matt: no. [54:46] Richard Webbe: From that [54:46] Matt: Um [54:46] Richard Webbe: movie. What was that movie where he [54:48] Dave Pengelley: You [54:48] Richard Webbe: reverse [54:49] Dave Pengelley: want [54:49] Richard Webbe: engineered [54:49] Dave Pengelley: the new [54:49] Richard Webbe: your [54:49] Dave Pengelley: benefit? [54:50] Richard Webbe: product? That you are Ben Affleck with the beard. So okay, so then of course that introduces all the other issues, doesn't it? That once I do that. I then now have scale, support, API, [55:04] Dave Pengelley: My upgrade [55:05] Richard Webbe: multiple issues [55:06] Dave Pengelley: security. [55:06] Richard Webbe: that I'm introducing. And it depends, like you said, if they're only using one or two percent, maybe it's worthwhile to create a bespoke version for your business, uh, or otherwise you just buy the big one. [55:17] Matt: So [55:17] Dave Pengelley: Yeah. [55:18] Matt: it really it just literally depends on what the company needs. And sometimes what the company needs, what they think they need, is not actually what they need. So [55:24] Richard Webbe: No. [55:24] Matt: they, you know, for example, I'll give you a literal example. I can't name names obviously, but I can tell. Tell you the color of it. [55:33] Matt: I had a quite a large company come to me and say, hey, we currently use this particular software tool. It costs us $1,000 a month and all we use is the dashboards. You know, so can we yeah, we need a dashboard tool, but that means that we need to we need to aggregate and we need to ingest data sources from they have 17 different other tools, 17, that I need to create a a source mirror. So my own database layer that is an exact source mirror of all of those other sources, then use correct database architecture and engineering practices to be able to serve that properly without just skyrocketing hosting fees, you [56:16] Richard Webbe: Yeah. [56:16] Matt: know, because of uh you know incorrect indexing on a DB and just collect all, you know, like any dev listening to that would just, you know, it's it's a it's an N plus one, like an unbounded collect, you know, that whole situation. So that's one layer. Yeah. And then you've [56:30] Richard Webbe: Yeah. [56:31] Matt: got their their feedback, you've got to be able to add that in, you got to be able to iterate and continuously improve. Um, [56:36] Richard Webbe: Getting back to the guys comment, we can, but should we? [56:39] Dave Pengelley: Well, even even when we we started working together, Richard, and we started talking about CRM, I've been building my own custom CRM in Airtable because at the time that seemed like a good option. You know, their free tier was pretty generous, and I can build a custom data model. But then for my own personal use, I've had to like rebuild entire things like the whole [56:57] Matt: yeah. [56:57] Dave Pengelley: email BCC so I could actually store and track. Emails against contacts, which is a fairly common out-of-the-box CRM thing to do. But that was a huge amount of like building workflows and things to look up things and reconnect them and save it and store it. So when we started working together, I was like, I'm not doing that maintenance. Let's just [57:13] Richard Webbe: No. [57:13] Dave Pengelley: find something off the shelf to use as a CRM for us because [57:16] Richard Webbe: You [57:16] Dave Pengelley: that's [57:16] Richard Webbe: need a support. [57:17] Dave Pengelley: too that that's that's not the core business I want to be focusing on. I'm not here to build a CRM. Um, I need a CRM, but I don't want to build and maintain [57:25] Richard Webbe: And [57:25] Dave Pengelley: one. [57:26] Richard Webbe: this is going to describe this basically describes how people who are previously doing certain roles will be doing different roles uh inside the uh administration of our world, right? In the IT processing and and data data capture and data processing because it's going to have to uh we have to have other support functions because producing the software is not used to be producing the software was the magic. That's not hard anymore. It's now managing the interpretations, the usage, the connectivity, and where the data is located. [58:00] Dave Pengelley: I mean creating software is still hard to do it well. I mean it's the done well, done fast, done cheap. People are doing it fast and cheap or on lovables and things like that, but not necessarily well. [58:11] Richard Webbe: And that was the theory of rapid development, wasn't it? The the term rapid where it's better to get something out there that doesn't work very well than nothing at all. [58:21] Dave Pengelley: But but even in that, Matt, and then chime in, sprints are built around stories and epics and stuff, and you actually are doing small bits of iteration quickly. in order to you know test verify move forward but not to output junk [58:34] Richard Webbe: Yeah, yeah, well. Yeah, I found some of the new project management strategies have been outputting junk. I think you're right. [58:43] Matt: I have so many words, but it's the end of the show, so this needs to come next time. [58:47] Richard Webbe: It does. I know I actually have this, I have another meeting to go to. [58:50] Dave Pengelley: very good very good well uh that is the hour thank you all for watching us if you joined us live uh appreciate you if you're watching the replay which is probably most of our audience at the moment uh Many watch the replay versus the live, but uh it's great to have you and your input. Thank you, gentlemen. Really appreciate your time here today. Uh good chats, as always, discussing all the ins and outs and not the hype. So uh we are yeah, as you might have noticed, we have our new logo in the top uh corner. We had a new logo on our opening stinger, and we're gonna finish with our new logo, which was inspired by our mascot, who did not make an appearance this episode. But Ruby is around working. [59:30] Richard Webbe: Love you know. [59:31] Dave Pengelley: um [59:31] Richard Webbe: Good good job, guys. [59:33] Dave Pengelley: so we've got the we've uh we will end on our new outro bumper stinger i'm very very excited all this new stuff um so it's definitely more polished than the old one but uh thank you gentlemen we'll see you next week