The AI Productivity Paradox: Why Doing More Feels Like Burnout: EP99.31

And I think. Yeah, sorry, you just paused up. I just. My brain just paused up.

0:00

All right, so, Chris, before we start, today's episode two in one week. What an achievement. But before we start, we said we were going to have a form that you could fill in if you wanted to join us on the still relevant Australia tour. This is the Australia tour, but there might be sort of like a world tour coming up as well. So if you're interested and you live overseas, like you live in the States or Europe or anywhere in the world, you can also fill in this form. You'll just have to say, no, I'm not in Australia, but notify me of future tours and that will allow us to sell your data to NORDVPN to get money. I mean, sorry, sorry, sorry.

0:11

No, it's the only VPN I trust.

0:50

To figure out, you know, where we might go when we take the still relevant tour internationally. But if you are in Australia, Australia, you can put in your details and register for your city. And we're not taking questions though, in Tasmania. Just keep that in mind. So fill in the form. There'll be a link in the description below. No matter where you consume the podcast, there should be a link in the description somewhere. I don't really know how to get to it, but this day and AI is still relevant to. And as a bit of an Easter egg, on the website there is a little figure that looks strikingly similar to Geoffrey Hinton, but it's not him for the sort of legal reasons. And yeah, anyway, have fun with that. So that is the still relevant Australia tour coming soon. I also should announce, Chris, a huge milestone has been achieved our LinkedIn group, our professional LinkedIn group has exceeded 200 members. 200 members. So if you want to join and get in on the world's best AI average influencer LinkedIn thing, also link in the description. I don't know why I'm plugging this because I hate LinkedIn and I hate logging into it. But anyway, it'll be good for advertisers so they can target you. So if you're wondering why we both have bags under our eyes, that's because we've been doing a lot of late night vibing. A lot of.

0:53

Actually, I actually decided to wear my glasses today to try to hide how bad my eyes look, but I think it's just making them look worse.

2:14

But, but I, I did want to touch on this topic. Not the late night vibing necessarily, but this idea that once you, like, once you have the capability to multitask across, say, tabs or across these agent interactions, you get this weird sense of exhaustion. And I've seen a few other people talking about this online. It's this, your mind starts to feel overloaded. And I almost want to dub it like AI exhaustion as this new feeling. It's like your brain at the end of a week feels fried. Time seems to have, like significant time seems to have passed in your mind, but very little at the same time. It's like you sort of completely lose track of time. It's like an AI psychosis sort of thing. And I find that the more things I work on at once. So if I open like six tabs and I'm trying to do six things at once, I do have this multitask cognitive load problem now where I just like, it feels more productive. But I question, is it more productive?

2:23

Yeah, I definitely know the time I'm finished working each day when I'm just sitting there staring at the screen and not doing anything, even though the next steps are sitting there telling me exactly, precisely what to do. Or in some cases the AI has already done it and I just need to test it. It's amazing how, you know, like you always think that, oh, well, if I'm tired and someone just tells me what to do, I can do it. But I'm saying that I think sometimes even then you can't do it because it's just too much mentally to cross that chasm.

3:26

I think this is one of those things too, where if you. We were joking about it last week where people like I spun up 70,000 sub agents to, you know, in order to, you know, be more productive or do do more work. And of course you, you can do all this stuff, like you can get it off going and producing all sorts of things at will really. But I think the thing I keep coming back to is the end output still for most of these projects, I would assume the vast majority is for humans. And so I find that if you're building a product for a human or a presentation for a human or a document for a human, you do need to validate it after that work has been done. And so once you start opening up all of these threads, I find you, you lose. You sort of have this cognitive load problem of like, oh, where was I? Like this context switching starts to become a real challenge. And it, I find personally now I've gone back to single tasking where I'm like, I'm just going to work through waterfall style my tasks for the day with the AI Because I can't handle the cognitive load of, even if it's going off in the background doing stuff, I struggle to then sort of fill in the pieces on it.

3:55

Yeah. And I think because ultimately, I mean, at least in our case, with the stuff we're working on, we are the stakeholder. So I really need to check that the thing I asked it to do is what I wanted, that it looks the way I want, that it behaves the way I want, and test it. So my role in the loop, even when it's actually making the updates itself, without me manually doing anything except asking for what I want, I still need to go off and try it and check that it fits with what I wanted. Did that meet my expectations? And in a lot of cases, when you're doing the sort of Agile style iterative development, for example, you might realize, oh, my initial assumptions were wrong in the first place. I actually, now that I see it, I need to do this other thing. So I agree. You often have to drill down on one thing to get that right. The second thing that I find with multiple things going on, if it's on the same project, for example, is you don't want them interfering with each other. Like one may end up working and one might not. And then you say, well, okay, I want, I want to keep these changes, but I want to discard these other changes. So if you've done multiple at the same time, it really muddies the water on that front.

5:09

Yeah, you can, you can start to go down paths where one thread has completely lost context on the overall project. And I think that, I mean this, in theory, this could be solved with like sub agents looking back at the latest context constantly and doing all of this fancy stuff. But I just question at the end of the day, in my direct experience, does this make you more productive or does it just give you like exhaustion, cognitive overload, and ultimately like, you feel like you're more productive. But the reality is there's that like, human block blockage where like, no matter how good the tooling is, the human is going to just get in the way for the foreseeable future. And it kind of needs to, because you like, who are we building these things for?

6:16

Well, and I think when you look at a lot of the examples on X that people give of these epic projects where they've had 20 sub agents building a project and then they finish with a piece of SAS software, in a lot of the cases they were doing it for the sake of doing it. Like as in to say they did it or to make something new that they, they wanted. But in a lot of cases they don't really mind about how it turns out because the goal was just to make something. And so therefore they don't need to care about the, the individual details of that or complete, completely understand it or anything like that. It's like I built this amazing thing, now look at my software I've made. It's not like an existing project where you've got constraints and you've got specific ideas of what you want to do. And I think the second you get into something that's real and meaningful, you start to actually care about the details more. So, for example, I was doing something during the week where I needed to contact a series of suppliers and I had the, I had SIM theory and simlink go off using browser use to research them, extract their email addresses, compose an email for each of them and then propose to send it. Now that's a lot more serious of a scenario than me just vibe coding an app or a game or something because I'm going to contact real people in my name. Like, I can't have it just spewing out random garbage. Like it needs to be accurate and get me what I want, really. And so I think in those scenarios you really do need to be mentally involved in the process and watch what's going on. And like, yeah, you might not be taking the individual actions, but you need to be supervising them to make sure that they're in line with what, like the way you want to work.

7:02

Yeah, and I think there's this, this also this division between what people think of an agentic loop or an AI agent versus like sort of how they're being implemented today. And then also just there's sort of, as we call it, like the tool called loop and continuation of a conversation or a collaboration between an assistant and yourself working throughout the day. And I think that the lines are a little bit blurred when it comes to white collar work because I think similar to your example, in the week I had a number of documents and meeting notes and requirements and emails and I needed to build a presentation. And so like the cognitive load there is, I've got to kind of bring all those documents and meeting notes and things into my own mental context and then build a presentation with how do I want to tell or craft this story. And in the, in the recent past, the way I would do that is I would go to the AI and say, you know, hey, I'd cut and paste a bunch of stuff from all over the place and be like, can you give me an outline for the presentation? Then I'd sort of go off and build it, maybe screen share with it and check in and see what it thought. But I think for me that's evolved into, okay, well, should I just give the agent that task and say, like, put together a first draft of this presentation and then go back and forth with it. And that's what I was able to do for the first time very successfully through the week. And it did open my eyes to this, like, unlock of White Collar work now that we're seeing where like I can now actually do this. And you might ask like, what's actually changed here? And the answer is nothing. It's just the tooling around the modeling and releases like Nano Banana have become like so good that, you know, I can now do it. But so the steps I took was with my new email integration in Sims 3, and this isn't a plug, it's just showing that I'm more likely to use it.

8:45

It's just teasing the audience really, at this point.

10:50

Yeah, yeah, it's a little mean. So I, yeah, basically what I did is I went in, I added all the emails into the context, I added the relevant documents from the file system using SIM link. I put meeting notes in there, which SIM theory can now, like record meetings, which is really helpful and transcribe them. So I was able to put them in. And then there's this new, like, I built a slide, like a presentation MCP that can build styled presentations. I gave it an example of a previous presentation and the theme of it. So like what it looked like and it absolutely blew my mind. Like it not only built the presentation, it built a talk track for me and it kept everything on brand. And then I went back and forth with it a little bit and this is where that sort of collaboration piece comes in, where it just can't be agentic because the human is the consumer of the content ultimately. And so it was in that scenario where I, I had to say like, oh, you know, on this slide, like slide six, it's ugly, it's too much text. Fix it, go back and forth, finally export it as a PowerPoint. And I sat back and then I, I did a run through of the presentation like I normally would, where I go through the slides, I act like I'm doing the meeting, like I rehearse the meeting and I was keying off its key points in the meeting notes and I'm like, you know, it's still my, my work to Some extent in the sense that I've gone back and forth with it and tuned it how I want. It's still my context that formed the slide presentation ultimately. But then I was able to get to a final output for the first time where I was like, I, I, like, I can't really explain how good this was without showing. And I can't really show because it's like fairly like personal, proprietary kind of information. So I, unfortunately I can't directly share it. But along the way I was actually tweaking how SIM theory worked because I was like, oh, I still want the email of the requirements sitting up right in my workspace so I can just like constantly refer back to it and decide if, you know, if it's meeting them. So I even allowed emails and documents for the first time to like pop out and just like sit over the session so I could see them and refer back to them while I'm kind of working this workflow. Anyway, it got me thinking about really, as we've been saying for quite some time, the untapped potential here is still the software layer and the tooling around this stuff and just sort of like discovering what's possible and making the AI even more context aware. And one of those things that I came to the conclusion of was all the files, all the things that you have open in the workspace, you know, they should just instantly be considered as part of that context. So anyway, like this is a pretty long ramp, but I do want to unpack it because I think there's some interesting advantages here. Like if you think about a desktop operating system right now, like Windows or Mac, it's like, well, this stuff should be like, like. I guess the conclusion I came to is like, why is it Copilot and Microsoft doing all this stuff?

10:53

And I think to come back to the point that started this about the cognitive overload, what we're seeing here is the AI stack gradually being able to fill up more of that context for you in convenient ways. So it can access your file system, it can access your OneDrive and get things in there even in, in shared environments, in a role based setting. It can use skills to keep things on your brand guidelines, which is something you previously would have had to direct it to or maybe it couldn't have even done. It can refer to previous conversations and things like that. So lots and lots of the things where before, like if you go back, say 50 episodes I was really big on, take the time, take 10 to 15 minutes to build a full context to get the Most out of the model to get your tasks done. Now, that same process that took 10 or 15 minutes might take 20 seconds because you can just reference all of these things or like you say, they're already up and it has them available. So the actual mental fatigue of building that context and doing those steps is now being taken on more and more by the AI software, like the AI system around it. So you're spending more time on the actual task itself, more time evaluating the output and whether it's actually doing what the original goal was. So you're getting from that first point to the later point more easily. And what's fascinating about this is this is cross model. So I've been doing similar things to what you've been doing. So just to give you an example of my own, we have a Simlink browser plugin which allows full browser control now, which will come out as part of Simlink. And I. You always hate my designs. I had just Vibe coded some design that looked like crap. And then what I wanted to do was update it to be more in line with our style guide. So I'm like, okay, well how would I do this normally? And so in the past I would probably copy paste the code and then describe what I want it to be like and then copy paste what it said back into the system, check it, see how it looks. But I thought, no, let's just try using it using the new context and tools. So I put in a screenshot of our application, I put in a copy of the logo, and then I said in my Chrome browser plugin, I would like the styles to match this plus and it went off, found all the relevant files that it needed to work with, inferenced across those images, and updated it in one shot. And then because of the way Chrome plugins work, it auto reloaded. So I literally just clicked on the icon and it was done and updated in one shot. So a task that would have taken me say 15 to 20 minutes before, by the time I went through the process, I had done in a single command because of that. And here's the crazy thing, I did it with GPT5. So a model that's what, like at least a generation old now and most people wouldn't even bother with anymore, worked just fine. And I've been doing similar tasks with like Gemini, Flash, Haiku Queen quite a lot. And my point is that we've been saying this for a while, that the models don't actually even have to get any better in order to magnify the effects of what we're able to do them because we have better ways of working now. And I really do believe that this gathering of context, and as you said to me this morning, not too much context, so it isn't just stuffing it all in and having millions of tokens. It's. It is having the right tokens, the right package that's put together to work with the, the, the LLM to actually take the steps necessary to get things going. And we're getting closer and closer to that effect. And I think just to do a rant of my own here, I think the reason that you and I are starting to hit this cognitive overload of doing tasks is because we, we are not used to getting so many big steps done in a single day like that. Like, it's just, it's a lot to take on that I'm able to do so many things towards my goal and, and, and keep up with that.

14:06

It affects my memory too. So like, I can't remember what I did earlier in the day anymore. I used to be able to clearly think back and be like, oh, you know, I worked on this one thing or two things in a day. And now it's like, yeah, you work like say 17 or 20 things. And I think back to the cognitive overload, though with that presentation, I come back to that task probably previously would have taken me several hours. That took about 20, 20 minutes. The longest part was rehearsing the presentation and then tweaking it with the AI assistant in order to, you know, just get it a bit more in my style and, and tell the narrative that I wanted to tell. But I think going back to what I said earlier, it is somewhat multitasking, but through a single thread. So the AI is going off and getting all this context and taking on some of the cognitive load in that single task for me, synthesizing it. And then the output is, from the human point of view, a single output, which is, I'm trying to work on this presentation, but it's able to take all of that other load off me. So in that case, I think it's a reduction of cognitive overload. I think when it becomes a cognitive load is when you try and work on that presentation, write some code, reply to emails, and you just. I just fundamentally think humans are bad at multitasking if you want to get really high quality output and not have to like go back and correct like a million mistakes. So, so I think, yeah, I mean, in software engineering maybe it's different in the agent, agent point of view, right? Because if you work in a big project, there's a bunch of like JIRA tickets and you want the AI to go and have a first pass on all of those. I can totally see how it could go and like solve like 20 bugs, right? And then you just sit there going through, reviewing them. But it's still, again, the end output is still the human in the loop. Validating that for humans. I don't care how many automated tests you have. The reality is a human still always, in my opinion in software needs to validate something. And, and I don't think that's very different to any form of like white collar work either. But I think the thing I struggle with mentally is like what has changed? Like what in reality, what is what has changed that now I want to attempt these projects and I would say one of those changes is definitely Nano Banana because it's able to produce slides that quite frankly, like you don't need to edit the slide anymore. You once, once you have it in this slide MCP sort of framework and the right prompting around it, it just nailed like, it just nails it, it just, it just works. I think it could be better in many ways, but, but it will get better this year. Like I know that now and so and I think, and you talk about using GPT5, right, for this stuff. So it's like, well, so that means the models already kind of work for this a while back. But why are we all just discovering feels like at the same time. And I think again, it's just the software, the tooling, the looping, all these techniques may be coming together. Does that your view?

18:11

Yeah, I think, I think it's a sort of shared discovery kind of thing. Like the new technology is made like for example, the invention of electricity, right? Electricity came out. You didn't immediately have like music players and you know, I mean, look, I don't know the history that well, but you know what I mean. A new technology, a profound technology comes out. It takes a while for everyone to discover all the possibilities for I bring you back to my song Endless Possibilities that I wrote about sim theory using AI, right? When the music models came out, that song was so profound that probably one day it'll be in a museum or something like that. But anyway, my point is that there's a lot of shared discovery going on and someone will have an idea. Like remember there was like the think step by step thing and then suddenly everyone was like, oh, this is, this has made the models better. But the models hadn't changed. It was Just better prompting. And I think this is all just evolution of that kind of thing. And I think there's a simultaneous evolution going on, which is where the people who are working with AI day to day in their jobs, and I don't just mean developers, I mean everybody who's working with it in their businesses or life have discovered better ways to work with it themselves. And I don't think you could necessarily go from like never having used AI to these advanced agentic workflows where you've got like masses of context and all this sort of stuff going on and then producing these incredible presentations. I think you yourself have to actually go through a bit of a process of working with it to understand what's possible and therefore know what to ask for. And I think there's a real, going to be a real education gap there where people need to be taught the different ways in which you can make the most of all of the stuff that's available to you. Because sometimes I even look at what you do with it with it, and we've got access to all the same stuff, even the more it, like advanced stuff that we haven't released yet. And you do so much more with it than I can. And I'm like, but I have all the same stuff. So I think that there is definitely a sort of technique and skill to it and almost like an art form of how you use it that needs to be learned by people. So I agree.

21:20

I mean, it's just a new way of thinking and working. But ultimately I think that most of these sort of AI influences, call them at the moment, are doing huge disservices to people by acting like all this stuff's magical and it just works out of the box. And you know, if you don't understand this, it's a skills issue that kills me. Like I think it is a skill issue, partly, partially, but it feels like almost gatekeeping on this stuff, saying, oh, it's just a skills issue and making people feel small, like they, you know, they haven't figured it out yet. To me, like it is just playing around with the tooling and having access to all the tooling. That's a big thing here, right? Like it's almost an accessibility thing.

23:24

Induced apathy. They're trying to say, well, this is all being solved, so don't you bother. Because the future workers, I, I've seen a lot of exposts this week that are sort of like, there's a big divide between the people who are going to hyperscale and become wealthy and the people who are going to be crushed by AI. And I'm like, life has never been like that. When new technology comes out, yeah, some people do really well out of it and some people don't. But a society in general becomes more capable. It's not like some, like the, the society just gets completely destroyed by some new technology. And I don't think that's what's going to happen here. But I do think it's accessible to anyone, this stuff. And I think the best way for people to actually learn it is to work in their own domain with data, that data and documents and stuff that they understand. Because it's only when people start working with their own things in their job or their life that they understand that they understand the impact of what this stuff is able to do. And I think that's the really important thing. It isn't just going in and vibe coding SaaS software that you've never done before. It's going in and doing like a work process like you've done here, with a detailed presentation with all real source documents that you're working with, and then producing something that you otherwise would have done from hand. That is that aha moment people have where they're like, wow, okay, I understand what I can do now. And then they don't really need to be told after that.

24:06

Yeah, I think it's just the, the challenge that everyone has is you might have tried these use cases. Right. A couple of months ago even, and you just didn't get anywhere. And now with the right tooling, you can try it again. And also knowing which MCPS to have enabled, like testing different models to see what output you get. I think that piece is where you start to discover the unlock. Like, it's a lot of trial and error and it's not like a normal software product where like, you know, you go to Microsoft Documents, write a document and then, you know, export it or print it or whatever you do with documents, email it. And the. Everything's controlled.

25:31

It's.

26:17

It's sort of still, in my opinion, experimental. The Wild west. Like no one knows what they're doing. And I always use on the podcast the. The. That Marge dress analogy out of the Simpsons where Marge gets a dress and then she keeps cutting it up and chopping it up to make it look like a new dress. And I sort of think with the models, they're at a point now where that's almost what everyone's trying to do right now. They're like tweaking how you consume the model, what kind of loops we put it in, how to structure memory. Like there's a big push at the moment around just using markdown files on disk is like memory and for rag and like all this stuff. And like that's. That was never out of the wheelhouse. Like GPT4 could have done this stuff, right? 3.5 probably could have with code interpreter. But it's just how we're now packaging these things up as there's this shared learning in the market. It is improving the ability for these things to get context and then output something of value. And to be quite frank, there's a lot of pieces being filled in like studio quality music, like phone calling, emailing, like, you know, all of the connectivity as well I think is slowly even.

26:18

Even operating software on your own desktop, right? Like these, these things can be added in, in, in different ways. I actually had a crazy experience yesterday that I shared with you. I had asked the system to make a game. Like I was just testing with Pygame to make this old game sync sub that we always used for demos when we were doing create with code mode. And I just wanted to see what it could do writing some code. So I'm like, can you research this game and the history of it and then make a new version of it, please? And then what I realized was I didn't have any MCP enabled except for simlink.

27:28

And.

28:03

And so I'm like, oh damn, it's not going to work with the research. And so I went to a new tab to enable the MCPS to let it do it properly. But then when I came back to the tab, it had done it and I was like, what in the world? Like what has it actually done? And it had actually used my own computer with browser use in background tabs, about five of them or something like that to research the game. It made its own sprites with Nano Banana. And then it, and then it made the game. It was just unbelievable. I was like, whoa. That was one of those goosebumps moments where I'm like, it's smarter than I am. It sort of like worked with the resources it had to get the job done. So it was like pretty cool experience. But on your point around the evolution of working with the models, I think one of the things we've been talking about lately is what you did with that presentation and being able to put that together is something that companies and enterprises would want to happen organization wide. Like you don't want people, you want people to be in this position where they're able to get to that really profound part of their task as soon as possible and not do all the busy work associated with that. And I actually think that this idea of a shared enterprise context, which is where in a secure environment that's like really gatekept from the outside world, your own team is able to access internal data, policy documents, design documents, examples of previous successful presentations, for example, like you did, and other resources that the company is part of their IP really and part of their identity and then having that by default enabled for different roles in their organization. And I really do mean by role based on.

28:03

Can you explain, because we've talked about skills in the pod before you go any further, like, how would this be different to say having like skills per role or skills at a high level in the organization? And the skill might be like, here's how we do presentations. Like, how is that, Is that any different?

29:53

Or that's precisely what I would propose using skills for. Like, if you think about skills simply as like a sub prompt, like a prompt for a sub agent that has to do one component of a task or perhaps participate in a larger task and think about the skill as really being an additional prompt that has a lot more detail. So let's talk about say when I do a presentation, I need to adhere to our corporate guidelines. So the corporate guidelines is a skill MD file that has all of the corporate guidelines that must be adhered to and that is injected into any prompt when you need to work on that. Or if you're working in an agentic mode, it might be a sub agent task which is evaluate Mike's presentation to see where it it falls down in terms of our corporate guidelines. And so that step will then report back to the main agentic process saying Mike's really screwed up here because he actually discriminated against Geoffrey Hinton fans in his presentation and we might get sued. Right? And then that goes back into the main thread. It's able to take just that little piece of context and understand, okay, now I'll iterate and I'll redo the presentation based on that and then it might hit the sub agent again and go, are we in line with our corporate guidelines now? So that's a way that the skills will work with sub agents, but it doesn't have to be subagents. It can work in a normal chat paradigm where it just becomes part of the prompt when and if it's appropriate.

30:12

And that's, I guess there's another piece to skills is like this, like shared knowledge across the organization. So if, you know, if someone's making a discovery or you know, in a particular department, like they're in the marketing department and they're working on a project and maybe this is something that's like switched on, where it's storing project and organizational knowledge globally within that department. So if someone else comes along that, that AI assistant's now aware, like, oh, you know, Sarah from marketing is also working on a similar product. Like is like, is that another paradigm? You think that is important?

31:36

Or I too, because think about some of the example use cases we've been working on with cloud files, for example, in say a OneDrive. And you've got this complex directory structure of all the different files that might be used, like logos, documents, policy documents, past presentations and stuff like that. When you first ask it to do a task, it's got to really go through that file index and work at, okay, these files seem relevant. I'm going to scan these files to find the relevant sections that might be appropriate and things like, like that. And that process maybe takes four or five iterations for it to gather all of the stuff that it needs to do. Like I gave you an example of a task last night that actually coordinated across three separate projects because each one needed small updates in order to coordinate the update. I won't go into the details because it'll take too long, but my point was that discovery process probably took. It was like a page of tool calls. You know, it's probably like 60 tool calls or something like that. But in the end it gathered all of the relevant context and was able to do it. Now what we do there is we actually then have the AI say, okay, what did you learn about the project here? What did you discover that's repeatable that you will want to know for next time? Dump that in an agent Stored MD file or other similar file, right? In our case, agents MD file. So that knowledge is now available about your structure of your project. So the next time you ask for a similar task to be done, it's like, bang, I know exactly what to do. I, or at least, at least what my starting point is to do that. And so I think this is where an enterprise can build up this corporate knowledge where it's like, well, okay, this is something that applies to everyone. This is a shared drive. Everyone has this. So let's start to litter this project with these context files and memories around the best way to work with this stuff that actually benefit everyone. So the next time someone else in the enterprise goes to do a task. It's like, I know precisely how we do this. Like this is how we do it.

32:14

I guess it's also like an attack vector, right? Like if you're installing software or something, or loading files in, you could just drop these like malicious MD files everywhere now and the agents will be like pretty fooled.

34:16

Yeah, exactly. But I mean, that's why in an agentic process we have quality assurance steps. Like a big part of the agentic loop is quality assurance to say, does what we're doing here actually align with what the goal is? And if not, we need to replan and reevaluate. So we were talking earlier, like some of our actual chat loops that we're working on with the different new tools we have enabled are so good. You're just like, well, hang on, how is Agentic even different? Because it's basically doing that anyway. And I think the answer to that is Agentic is going to have steps in there that are verifying and really that that relentless pursuit of the original goal. It's like we made a plan, we're going to stick with it and we're going to evaluate against that plan at every step of the way. So yeah, there might be a malicious step in there, but it's going to look back at the original goal and say, hang on, my goal wasn't to remove all of these files.

34:29

I think this is a test we should do on the show is like try and like hack an agent.

35:21

Let's, let's dump files in our share.

35:26

Like, can we get it to like call call, like use the phone call skill to like call the police? Again, this would be a cool experiment like to really like real high stakes.

35:28

And divulge company secrets.

35:39

Yeah, yeah, but we'll just make like, you know, Hinton Incorporated or something.

35:41

Is the issue a libelous press release? That would be pretty fun. But yeah. So my point is that I think that what people need to be thinking about is we are creating organizational IP with our workflows. How do we share that organizational IP with other people so we all become more productive. And I think that there's a real way of doing that now. And I think as you point out, that way has probably existed all along, but now that we've discovered it, we need to make the most of it. And I really do believe, like even just in my own work, I've seen the benefits of that. Like as the system gets a real feel for it, it can do stuff. And I think another major, major point about this is I would argue that your Lesser models, right? Like a haiku, a Gemini Flash, a GLM 4.7 Deep seq models like this, they benefit so so much from really good context and tool calling that if you as an organization have people say, using the greater models to build up these context trees and skills and other elements and around it, you can then have the rest of your organization using the lesser, cheaper, affordable models that you can actually afford to run across, say, a thousand people. And they're still going to get the same sort of profound output as the big models because they're benefiting from all that accumulated context.

35:46

I sort of start to wonder here about the. It's like, like, I don't know, like cohesion of context or whatever you want to call it, where you have all the. So you're accumulating this enterprise. I mean, essentially it's just like markdown files with structure for like knowledge and then you've got your skills and I guess at a high level, because this is really just a bunch of markdown files. The benefit to a large organization is it's highly transferable. Right? Like you're not, you know, because you would, you would think you would kind of say, oh, okay, well, the moat that Claude or Chat, GBT or Copilot or whatever it is might have in the future is, well, they've got all this context stored and so therefore it, like the cost of switching providers, you know, it just becomes too much of a pain. And I would argue at the enterprise end that's probably still true just because of like privacy and security and all that kind of stuff. Like you just sort of get stuck with a vendor because of procurement really. But to me, there's also the play now of you sort of think a bit longer term, like, and I don't necessarily want to say like copilot, but you would start to think back to what we've talked about many times before on the show, where the operating system or the. Maybe it's not the operating system, but maybe it's. Even the cloud providers have a bigger moat on this than people may realize. For example, you know, you look at like Azure or Amazon, most people in the enterprise now have all their data in several databases in the cloud. Right. And a lot of them for things like etl, run applications like databricks. And so you kind of start to think like those solutions if, if context is king here and storing that context is, is king, these cloud vendors and these products start to have even maybe more of an advantage over the labs longer term.

37:17

Yeah, exactly. If you can build a really solid map of that data and how to work with it. That becomes an enormous advantage. Like almost like it could actually be a product that these people produce. It's like this is our AI layer on top of your entire cloud. All of your databases, all of your files, all of your roles, everything that you guys have as an organization, we have built the ultimate context thing that you can query. And your AI provider, whichever one it is, can interface with that and actually benefit from that context. Now they're not going to do that because they don't want it to be portable, they want it to be just them. But as you say, they do have this major advantage, which is that they have it all sitting there. And it's funny because so many of the examples you see people giving with Claude code around these cloud file systems or local file systems are like, oh, I'm going to organize my files. It's like, oh, look what it did. It took all my receipts and made them into an organized folder. I would argue that is totally unnecessary because if you map the data right and it understands where everything is once it's done that, it already has a view of that data. It already has a way, it's a little bit like document based databases where you build a view which will extract all the relevant documents into that view and then run a second layer view. Just to be clear, we use cloudant for years, so I know all about this. And it would have a second layer view on that which would then give you like second order data that's then organized. And I'm proposing that this is what all the cloud providers should be doing, is giving this ability to make views of your data that give context to give specific kinds of tasks you're doing. So for example, rather than an enterprise having to say, we've got three Oracle databases and we want to be able to access these columns and rows for this role. So therefore we're making an MCP with tool calls that allow them to access that stuff. Instead. If you had a sort of AI level view of that data that had the right permissions and stuff like that, you could almost have a generic layer where you can build up these contexts and then make them available in a shared way to the specific roles in your organization. So they can then just hit up that context anytime they need, either by referencing it directly or having the AI infer that it's needed for the particular task at hand. And I really feel like this is going to be the next evolution, which is not, it's not working with the traditional systems, the way they're intended to be worked with, it's. It's extracting what's needed for the tasks in a cohesive way and then sharing that amongst the organization. And I think just to labor the point a little bit, this is what we've talked about, where ultimately you could probably sub out some of the underlying systems if they're expensive for you, because ultimately it's all going to come down to how do you produce that view of the data, and it doesn't really matter what the sources are.

39:18

I think the overall positive trend here, right, if you're just looking at it from the point of, like, consumers and also enterprises consuming this technology, is it's working out pretty well because ultimately, if this is just a series of, like, MD files and structure around, you know, existing cloud data and databases, and then they're building MCPs to interface with different, like, permission control data. Right. All of a sudden there's no, like, you know, like, everything can happen on your cloud or your file system, and then the inference is just like, when required, not necessarily like you, I guess. You know, I get what I'm saying is you're not giving all your data, you're not handing over your data to an AI lab, which is like a huge concern for people.

42:13

Yeah, we've had it. We've had a few people we've spoken to ask about that. It's like, do we need to build a new unified data source that we give the AI access to in order to do this? And my answer is always, no, you absolutely don't need to do that. You really just need to give it the ability to access it in a way that will give it just what it needs. And I think that's far superior. We know that less context, but more concentrated context is so much more important. And I actually had a really funny thought. You know how there's people like my wife, for example, my old business partner, who will open like 300 tabs in their browser and just go, oh, no, that's my mental memory of the thing I'm trying to do, the trip I'm trying to plan, or, you know, something that they're trying to accomplish and they'll open up all these tabs. Well, with browser use now, that can actually legitimately become the context to actually get that task done. If you think about it, I was doing something the other day. I was. I was doing some. Some research on suppliers again, and I had about six tabs open. And so then I asked it, can you please research this and it's like, oh, I see you've already got these six tabs open that have the relevant information. I'll get the information from there. And it actually completed the task based on that information. So in a weird way, maybe the tab people were just always right. It is right to have so many tabs you can barely even know what they are and click on them.

43:01

Yeah. But it's just like a weird mental model of their, like their world. All the crazy.

44:26

Yeah, they just never had the mental capacity to synthesize it themselves all in one shot in their head when, when.

44:31

To dig into that. And it's not that relevant. I'm just interested in the sense of like If I have 100 tabs open, is it is the way it's working looking up like all the page titles or in URLs and then going, oh, I'm going to fetch all the stuff off that. Like how?

44:38

Well here's the beauty of it. The way ours works and I say this having been the one to make it, but it's just the most beautiful thing ever. But no, what, what it actually does is have a local agentic loop using a really cheap model. Right. That will go in and work out. It has a goal, it's given a goal to do with the browser and so it's aware of the tabs and the tab groups that are open and it can say okay, based on the tabs that are open, these ones seem relevant. I'm going to go in, try to scrape the context. It can actually run JavaScript if it needs to extract things like if they're hidden and things like that. Because it's your browser, you're logged into everything. So it's able to get into systems that normally would prevent like web scraping and block servers from Amazon and put captures up and all that sort of garbage. Because it's your browser, it can just do all that stuff and so it's able to then traverse the DOM if it has to run JavaScript, if it has to take screenshots of the page if it needs to and then do inference off the screenshot. So it's like a sort of all out attack war on. I am getting this information out of here whether you like it or not, browser or website or whatever it is. Like there is no defeating this technique when it comes to gathering information, even extracting videos from web pages. Like this is the ultimate. Like if you think about MCP tools like that try to get say YouTube videos and YouTube transcripts that get blocked all the Time or you know, websites with like, like information they're deliberately trying to make hard to scrape. There's quite a lot of websites like that and web apps like that. This just bypasses all of that. You can get anything. And the cool thing about it is it can be a coordinated attack from you and the AI because you can open up the tabs you think are relevant and group them if necessary and it can open up tabs that it thinks is relevant. So the actual power of this is quite profound. And you might think, oh well, you could use browser stack or one of these cloud things or you know, headless Chrome browsers running on a server. But they have all the deficiencies I mentioned earlier like IP based blocking, rate limiting, anti scraping techniques or expense. Like it's expensive to run those tools, right? Both on a server level, a time level and a cost level if you're using a third party tool. So actually just running it on your computer or another computer you have just laying around bypasses every single one of those things because it just thinks you're an aggressive consumer browsing the web at some incredible rate, which is totally fine. So I really do think there's a huge unlock with this sort of local work on real computers because you just overcome all of the technical challenges that normally exist there.

44:55

So pivoting away from all of this slightly, but I think it's just such an interesting like topic around something that now that the models are getting really fast and cheaper and more competent, it like has been discovered in my mind. A lot of people talk about SAS software and in the public markets right now SAS software is like tanking because people are worried what impact AI will have on it ultimately. And like you know, it. I think the future, like the models and the progress is moving so fast now it's really uncertain what the future looks like. Like do you know, is this a good investment? Are these people ever going to be able to make money? Because do these layers of AI just get to such a point where software is totally throw away and in context and it just changes our whole worldview of things? And that's a completely separate debate. I'm not so sure that it's the best.

47:38

We're not known for our investment advice, that's for sure.

48:35

So I'll stay away from that. But one thing I did learn not to bring this about email emailing calendar, like I have the last two episodes now. But the one thing I found rebuilding some of this stuff myself was like how? Like the real question was like how, how should this be in an AI world, right? Like, you know what, what kind of things can you now do differently that you couldn't do before with software? And if you think about your experience using, well, I don't know how many people use it, but like say Microsoft Word or Google Docs. I'll focus on Google Docs. Often when I'm preparing for the podcast or previously was preparing, I would cut and paste like links and a whole bunch of stuff into a document just to have to reference. Not that we're great at sourcing material or very reliable or accurate, but you know, just, just to have in there in case we want to refer to these things.

48:38

Like slightly below average.

49:33

And so what I find is like, you know, the styling's off. Like sometimes it'll paste links in, you can't see it. Anyone that's used a web based text editor realizes that it's an awful experience when you're like pasting information and cutting and pasting. And I don't want to bring this, bring it even below average in terms of talking about cutting and pasting. But it got me thinking about this idea of like when I paste something, why isn't the software smarter? Why isn't the software thinking, why isn't the software like, he doesn't want a white text on a white background. That's stupid, right?

49:36

Like that.

50:11

And I think that's pretty valid. And so I started building all these little sub functions so when you pay something, it thinks about it. It's like, well, what, what do they want to paste here? Should I style this better? Like, is this even going to paste in, in a nice style? And you know, similarly one you showed.

50:12

Me that was pretty awesome was you were, you wanted to write an email to someone and you just pasted in the body like the, the, like the goal of what you wanted to say the person's email address and just all the junk you had accumulated. And it actually extracted the email address and put it in the to field and removed it from the body. Yeah.

50:31

So this was the other thing I was thinking, why can't you bark orders to your email client being like, oh, BCC and Chris. And because I BCC and Chris a lot, it just knows, it just goes, oh, he's got 17 Chris's in his address book. But it's definitely this one because that's just what he does. So it's thinking and I think this idea of reborning or rebirthing everyday software that you use where like, like to me that's the best path for these existing SaaS businesses and existing software providers like Adobe or Microsoft to start thinking through because that experience is so much better and more powerful. I mean, in like to a larger extent like Gmail and, and Calendar, instead of just like, just stitching in these sidebars like, ah, hey, I'm email Gemini. I'm really incompetent and have no understanding of your overall context. Like, I, I think that's where that knowledge graph plays into it. And one other thing about that that I figured out in doing this is when you have multiple email accounts, like you might have an Outlook, Gmail and whatnot, and you might have multiple calendars. Each of these require different context. But ultimately you can sort of have a black book or like a Rolodex of contacts, but then these all have different contact managers and different ways of storing data. So one interesting way is also accumulating a knowledge graph about people as you work. So when you email someone, it remembers key points, it connects that to meeting recordings, it connects it to mentions in other chat sessions. So you start to accumulate like this is sort of like huge knowledge graph on individual people. And that allows you to do things that feel very natural with AI. Like if I say CC and Chris, it's able to quickly consult that file or key things from that file. And when I say quickly, I'm talking faster than you can blink like it, it can go so fast like, like.

50:50

Like almost like a flash.

52:51

Almost like a Flash. And now I can book travel competently. But yeah, there's something I think really interesting there where you look at an existing piece of software. Like another example would be Adobe Premiere, right Where you're like, I mean, I don't do a lot of editing for the podcast, but it's certainly top and tail at each week with, with the intro and might cut out like, you know, something we've said that's bad. And so, you know, even just having those, the tooling in there when you make edits to realize like, oh, you know, he didn't want to cut where there's still a huge chunk of audio. Maybe he wanted to cut here instead and he's just slightly misfired with the mouse. I think adding these elements of thinking into all the existing applications we use today could honestly be one of the most productive game gains where everything just starts to seem more intelligent as well in a lot of the applications we use. Like, I think there's a lot of.

52:53

Disruption here because it's stuff that the computer can do so much faster than we can. And so much more reliably than we can. That is really not helping anyone. Like you taking the extra time to do it manually. It isn't like the grass fed human content thing. It's like this is just manual work where you're having to give the computer precise inputs that it could otherwise infer. Right. And if it can do that, then everybody has a productivity gain. And like these things may sound trivial, but really it is a big difference. Like if you can get through 10 more tasks a day that you, you couldn't have otherwise done because of all these little productivity gains, that's better for everyone. The economy, companies, individuals.

53:51

Yeah, I think these are sort of the unlocks and gains. And so my advice at least to these SaaS companies that are worried about this stuff or even like if you, you know, if I was at like Microsoft right now looking at the productivity suite, I'd probably be less focused on, on like co pilot sidebars and all this annoying nonsense and more focusing on how do you bake in these experiences where the software all of a sudden just gets you and feels more intelligent. But I do. That does bring me back to the point of really the tools that win this era to me is still going to be the monolithics because of having all that knowledge graph and context and that core context. And you kind of wonder once those people unify and start to adopt these like AI OSes or platforms, then does the, the document editor just have a, like the ability to consume into your knowledge graph from that product and that like that's what it's, you know, you can sort of connect your, your SIM theory instance to Microsoft or whatever or it's like I connect Claude to Microsoft Docs and now I have all my context from that. So I think that's probably a likely scenario here of where this stuff heads.

54:36

Yeah, true. And even though you say that the monoliths already have this stuff, the thing is that systems like ours can actually get all that stuff too. And we can, we can do it as well. They, they have the ability, I guess they could cut people like us off. But other than that it's like this is still, this is still possible to do externally. Right.

55:58

But can they cut you off if you have things like SIM Link where you just have rude access to the computer? It seems like the answer is no way. And I mean the cloud. Yes, but I don't think the cloud can restrict your data. Maybe Apple Health won't let you consume their data. So Chris, before we go, because we are petty people I'm kidding.

56:16

It's funny.

56:35

We were debating last week whether we should talk about OpenAI introducing ads with ChatGPT and we both agreed, like, who cares? We don't care. Like, they're not really in the zeitgeist anymore. The vibes are off. As Sam Altman said, like, the vibes are going to be bad for a while. But I do, I do think this is interesting to talk about, like the different approaches to companies and sort of like the, the general mood of things versus the reality. So I, I don't know the exact revenue figures and I'm not going to pretend I do, but recently OpenAI announced, like, you know, don't worry guys, we're making a lot of money because a lot of people have been talking about, you know, they're obviously burning a lot of cash to run the, the domineering factor of Chat gbt, which is consumer. And everyone's saying, oh, that, you know, they're going to go broke if they can't raise money. It's, it's end days and, you know, everyone's talking, at least in the circles on X, about flawed code and agents and all these other stuff that OpenAI hasn't really nailed yet, in my opinion, like their tool calling and things like that. Having said that, if you apply models like Codex or GPT5 to really good frameworks, it turns out they're not that bad. So I think it's a, it's a skills issue by OpenAI right now. But I did want to sort of call this out. So there was a post on X like just today saying we have added more than 1 billion of ARR in the last month just from our API business. People think of us mostly as Chat gbt, but the API team is doing amazing work. Now, what I find funny at the moment is you've got Altman out there saying, hey, we, you know, Billy's in the bank from API usage. There's a lot of defense around, like how they'll make money and how much money they're earning all of a sudden. And I think that's because the narrative is out there that like, the wheels could fall off this thing, especially as they lose enterprise share to Google and Anthropic. There's also an information in the, an article in the information where the OpenAI CFO, who seems to, every time she opens a mouth, get them in a lot of trouble, has come out and suggested that one way, and I've heard her say this before, that they could make money in the future. Is that pharma and biotech firms that use their models for discovery would give them a cut of the drug discovery and they would have licensing deals there now. So, so at the lower end you've got ads being shipped soon into an $8 a month chat GBT Go tier and free users. So they're not only going to ship ads to free users but also their $8 a month customers I guess, because they've got to pay for that almost free plan and that, that usage. And so we're getting ads. And Sam Altman, there's a lot of people out there quoting him saying like, you know, our last ditch effort to make money would be ads like you know, if we were desperate and then now that's sort of come true and then you've got the CFO out there being like, don't worry guys about all the billies and our huge valuation because we're going to convince biotech companies that use our models to make discoveries to give us a cut of the license. Like it honestly sounds insane. Like it's one of those like sniff tests where you're like this sounds like absolute. And yeah, the problem with the bio.

56:35

One is I couldn't imagine a legitimate scientists being mono model like, oh, I'm only going to use this model and therefore I can attribute all of the thing I created to this model. Surely they're verifying across multiple models, fine tuning models, doing other things beyond just that single model just sitting there on chat GPT like what if we try horse tranquilizer on people? You know, stuff like that. It's like instead of giving it to cats, give it to people. And so I think that no one's ever going to do that. No one's going to ever admit even if they did that I only use ChatGPT 5.1 and therefore you guys deserve some of the money. They're going to be like, no, it played a small role in an overall project that I worked on. To me, the problem with the ads is not the ads because I understand some people can't afford or won't pay more for it and ads is a good way to supplement the income. I sort of get why they're doing that. To me, the problem with that is so much more profound for their company, which is having ads shows that they are looking at what you're doing on there. They're profiling you and they are looking at precisely what you're using the thing for. And they're categorizing you and labeling you and building a Shadow profile on you to know what ads to deliver. Now there was a thing, I don't know if we actually mentioned it last week, but there was a thing that came out about clip Claude where people had found a way to discover that if you use Claude directly, like not via the API but directly use their interface, it is building a profile on you with very, very specific categories including like sexual orientation and other hyper personal things that it is building on you. Now that really really rubbed me the wrong way thinking okay, if I'm using either ChatGPT or Claude directly and I can imagine probably Copilot's doing the same thing. I don't want to slander but maybe they are, who is going to trust that? Like I. That really, really shakes me when you think about the really highly. Like I'm thinking about companies here. Like if you're putting in really serious stuff like that could get you in trouble personally, get you in trouble as the company or anything like that. Do you really want to do it in a system that is actively profiling you for advertising? Now I know they say they don't do it on the upper plans, they're not going to deliver the ads to you, but that doesn't mean the profiling isn't still going on. And I really feel like if they want to compete at the top level, they absolutely mustn't be associated with that kind of thing. Now I am sure that they don't do it at the API level and I'm sure that they don't do it on their, their top level plans. But it's just the just being in the muck like that. I just don't think you really want that. And it must be a huge amount of money for them to be so open about it and to proceed down this path I think because I just don't like it as the future of AI. I just don't think it's like I get it, I get the Internet's all about advertising, but there's just something about it I really dislike.

59:54

I think what's upsetting about it is it feels like fundamentally this AI is such a huge breakthrough. Like it. It's something. Now I know everyone listening to this show could certainly not live without anymore. Like you, you would really struggle just to get by, I think day to day, like I personally would. And, and so there's that piece of it and you think like, oh God, is this where we're at? Like the only way to pay for this stuff is through ads. Like are people not like it Makes me question everything. Like have these guys not packaged it up in a way where companies and people are getting enough value that they're willing to pay for it. They're willing to just be the product again, like all Internet businesses that go consumers seem to be. And I understand for most people, like because they're using these like free models and they're not investing the time of them, maybe they just don't have the time that, that they're just never going to pay. Right? Like it's just never going to happen unless they see true value.

1:03:03

And these are the same people who are dropping $1,000 for an iPhone every time it comes out, right? Like your average consumer has an iPhone. Like it's really common for people to spend thousands of dollars on something which clearly does deliver a lot of value to them. So it's not like people can't afford to pay for this stuff if they want to. It's just that like you say that they're not showing the value there for people to be willing to pay more than that.

1:04:04

I guess the reality is most people just don't value their privacy at all. Like they just, they say they do, but they mustn't because like clearly this has legs. But interestingly, Google came out and publicly stated they have no plans for Gemini ads. No plans doesn't mean never, but no plans. And that they're surprised by ChatGPT. I mean that's a, I'm surprised our.

1:04:31

Competitors would make such a profound misstep out of desperation.

1:04:52

Yeah, but so you've got this, you've got them now doing age verification. So they're reading your chats to figure out how old you are, which I also think so creepy to see if you're allowed to enable Sexy Mode as I'm calling it or not.

1:04:56

And it just, I'm telling you now Anthropic knows if they can enable Sexy mode too. Like they have the same data.

1:05:09

Like no, I know that, but my point is just, I, I, it's like Altman's like, oh, the vibes will be off for a while. But I feel like, is everyone over at Open AI okay? Like Greg, are you, are you, are you okay? Like, do we have to write another sad song here? Like I, I don't get it. It's like every decision they are making feels like how you would go to single handedly destroy a business. The other, the other thing is it's like they're losing enterprise which I think is willing to pay for this technology because they see value Going after consumer which is loss leading, but long term may actually be the right call. And I'll look like an idiot trying to pay for it with ads because clearly they're losing a bunch of money. And then they've got a lot of people out there saying like they're going to go broke at this burn rate. And also they have to raise gargantuan subs of money at ridiculous valuations. Maybe not ridiculous, I'm not certainly taking any side. And you've got the CFO coming out saying, oh, don't worry, we're going to take a cut of like novel drug discoveries as our revenue model. All of this seems very defensive. You've got Altman tweeting about, don't worry, we're making Billy's from our API. Like something doesn't add up. The messaging feels off and like it feels so defensive right now. Instead of what they used to be known for, which is like we're now deploying the best model for this use case.

1:05:15

We said this the whole time is like what we kept coming back to is they had category brand recognition. Like when people thought of AI, they thought of ChatGPT. And I really feel like the marketing effort that Claude Code has done over the last two months has been enough to say, hey, OpenAI isn't the only thing that is AI. And Copilot's done that to some degree as well. I mean, even though I guess it's using the OpenAI models, but the, the idea that they, the one thing they had which is their brand, they're just taking shots like it, you know, they're taking self inflicted wounds on it. And also getting to the point where people are realizing they're not even the best models.

1:06:38

Yeah.

1:07:23

Like at least, at least if they were doing all this bizarre crap and they had by far the best model. Right? Like, then you'd be like, okay, do what you like, guys, because you're still going to win on the model. But they're not even winning on that.

1:07:24

They should be listening to the show. Remember when we said no one will care about chat, GBT apps or whatever in, in a couple of weeks? Like, have you heard ace single mention of these things?

1:07:36

Like, yeah, like, well, next time I need to plan a trip to Chicago, I'll be right on there using that Glory interface or like buying pet food or whatever I can do on there. Like, it's just not like, I mean, like a hark back to what we were talking about earlier about the actual productivity gains that help you day to Day. They are. They're subtle. Subtle like actual things that the AI is good at. It's not like dynamically building a UI when you could already just go to the website and do that. Like, it's really. That. That was a terrible.

1:07:47

How do you actually use mcp? I use it to gather context and take action. And I like to use a lot of things at once. Like, I want it, like, as you say, reading my browser tabs, looking at files on my desktop, calling perplexity, calling X, getting different opinions on research, being able to read pages, all these things. Right. Not, I don't want to see interface.

1:08:18

Yeah. And it's the. That is how you use the AIs intelligent, because it can come up with a strategic plan of how to use all of those things in a way that will accomplish your goal. You're not directing it at each step of the way. Or even worse, clicking on dropdown boxes to input stuff. Like you said, you're a great example of building up personal contacts of the regular people you talk to, and how you interact with those is so much more valuable than having an interface. Do you want to be a dickhead in this email or would you like to be nice? Let's say, oh, click a button for that. It's like, oh, make it more poetic. You know, it's like this or stuff.

1:08:41

Is having my salesforce leads show up in a GUI in chatgpt. That's not helpful. If I want to say, like, go and ENRICH, like, these 20 contacts and research them, that is helpful. These are like. It's just like. It's almost like, Sam Altman doesn't use AI for product. Like, I just don't get it. I'm like, if these people are sitting doing this stuff, why isn't this product getting better? Whereas you've got anthropic on the other side where it's like, obviously these guys are using it.

1:09:21

Like, on one hand here, we're talking about having cognitive overload by accomplishing too many tasks in a day. It's like I've just accomplished too much today that I have to stop because I just can't keep.

1:09:48

Well, I can't make sense of it all of what it means. Yeah.

1:10:01

Like, my brain actually isn't, like, animalistic brain just isn't equipped to deal with this many, you know, decisions and testing things in one day. And then on the other hand, they're like, hang on, how can we make money? I'll just add advertise.

1:10:04

You know, the funny thing about capitalism is Everyone thinks that people will be less busy and have more free time. And Elon thinks we'll all live in some, you know, kumbaya camp, just like having lots of babies. And I just don't think at least he practice.

1:10:19

What's he preachers like.

1:10:34

Yeah, that's true. Fair. But here's my thinking around it is I think it's just going to put demands on workers to literally be 100 times more productivity where you do feel more exhausted. Like I hate to say it, generally this stuff happens.

1:10:35

You can't really lie anymore. You can't really be like, oh yeah, I'm working. I've worked really hard all day on this thing. It's like, well now you can be like, well, where is it? It should be done.

1:10:49

This is a single, like single 15 minute task now.

1:10:59

Yeah, yeah, exactly. So it's like there's no more long lunches and things like that because it's really obvious when you haven't worked. So yeah, you're probably right. It prob. Probably will lead to sort of like mental decline and ultimately bad things for society. But while it's new, it's fun and we should all be using it.

1:11:03

Yeah, exactly. All right, let's leave it there. I don't know what happened with that episode. I'm cognitively overloaded. But if you do want to join us on the Australian leg of the Still Relevant Tour. Still Relevant Tour. There's definitely going to be like some merch you can buy, maybe some gold pendants, maybe some Still Relevant T shirts.

1:11:20

And if anyone is a singer or.

1:11:38

A band or anyone stop over complicating it, we literally taking the show on the road. So we're going to record the episodes live at all of our Lives Still.

1:11:40

Relevant, one of our songs at each event.

1:11:49

It's hard enough logistically just to be able to record a podcast, let alone have some sort of like AI Sing Fest. So let's not promise it. All right, that will do us. Links in the description. Sign up for the Still Relevant Tour. Also, if you want to support the show and use everything we talk about, SIM theory, AI Again, link in the description. Also LinkedIn if you like that kind of thing. I don't. I'm never there. But you can join that as well.

1:11:51

It's such an emphatic presentation. It's awful and I hate it.

1:12:15

But please join and advertisers will probably use it to target you. All right, on that note, goodbye.

1:12:19