B (3:28)

Yeah. And actually, if you listen, I think you're gonna play the song I made soon as well. And there was something in that song that it pronounced that I was really impressed by. Oh, gdpr Because I was thinking a bit different to you. Rather than trying to, like, craft lyrics in a different model, I'm like, I'm gonna use Suno 4.5, just straight raw. And so I thought, I'll find the most dry and boring email in my inbox, copy and paste it in and just say, write a song about this with no embellishment, no instructions, no styling information, just the email itself. Are you gonna play it?

C (4:05)

Hey, Chris Sharkey, it's time to rise Got an overdue task no need to hide the clock is ticking it's calling your name don't let it go we're playing the game completely overdue Feel the win your GD Our training where it begins Four days late but it's never too soon Light up your day like a rocket to the moon.

A (4:38)

This is the point we're at.

B (4:39)

How good is it? And I was just. I was just thinking, like, I don't know about you, but I love that song. I always love the ones I made. But I was like, that really motivates me to do my GDPR training. I was like, feel the wind. Like, it's so good. Your GDPR training's where it begins.

A (4:58)

It's like, I think it's co pilot that gives you that radio that, like, it's like a your day sort of you can listen to. And I think maybe they should do it in song.

B (5:10)

Like, if you put that song on, like, Nova 95.5 or whatever, no one would know the difference. It just sounds straight up like a. A song on the radio, except it's personalized to me about my GDPR training. And then I was thinking we talk a lot lately about having, like, agents doing your work in the background and, like, you know, having an AI companion throughout the day who's off doing background tasks and actually completing real work for you. Then I was like, why not have another AI companion that's literally producing the soundtrack for the day? Like, you know, we're winning, we're kicking goals, but, like, with really, really specific information in the songs as the Day goes on like a sort of custom gym playlist that's actually responding to the exercises you're doing or something like.

A (5:56)

I honestly think the peak lol is going to be when someone links one of these music makers to their health tracker. So it detects the heart rate and it's. And it's singing about like.

B (6:07)

I die.

A (6:10)

So I wanted to test more of. Of Suno here. So I did a few variations of that song just so we could get a feel for the model. Because one thing they talk about in the release is it's really good at Fusion now, for example, you can do emo plus. Neo Soul was one of the examples they gave. So let's hear a bit of Emo and Neo Soul.

D (6:37)

Here. I am listening once again to two sharkies talking A. But when will they reach episode 100? It's not the question that keeps me.

A (6:56)

So I think. I honestly think this is leaps and bounds better. And they're only doing this as a point release, just its pronunciation. And the other thing I noticed is nearly every song now that you get out of this thing is good enough. Like, you know, good enough in the sense that I would play it on the show, whereas before I would have to go through many iterations and like tweak how the model wrote out the word to make sure it was pronounced right.

B (7:19)

Yeah, exactly. And there was a lot of just absolute trash coming out of it. But I love the idea that now these things have APIs too. I just really, really like the idea of the, like these songs just being created in the background just so you could just have a playlist that's totally custom to you as your day goes on.

A (7:38)

This one's EDM and folk this day in AI.

C (7:43)

Why do I still tune in for.

B (7:45)

Mike and Chris and their average spin?

C (7:49)

The countdown to 100.

A (7:53)

And then I don't know where the.

B (7:54)

Folk came into that.

A (7:55)

Yeah, I don't know. That one didn't fuse very well. This is apparently Korean week after week.

C (8:02)

For mediocre insights I could easily seek. Will they ever. Will they ever make it there? 100 episodes I.

A (8:13)

So, yeah, pretty, pretty good. And then I did a Chinese one. I don't know, it's not in Chinese. And then a little bit of Arabian.

D (8:36)

Cool.

C (8:41)

How can we tell without a boom?

A (8:43)

As I listen alone in my living.

C (8:45)

Room no expertise, no special inside.

A (8:50)

I love how it's just insulting us as we go. But anyway, there's your pseudo 4.5 road test. But I'm really impressed. I. I like such a great product so much fun. I made some kids songs with my kids this morning and I don't know before it was really hard to get one that they would, they would like, you know, love them, get really excited again because they just take it for granted now. But they're really into this model as well, so highly recommend checking it out sooner. Version 4.5 Now Chris, there was some updates to Meta this week that we want to go through relatively quickly out of Llama Con. That's the conference around AI that Met is now putting on and this is everything they announced at Meta Llama Con. Now I think probably one of the biggest things was the Llama API that they say is now in preview and this is essentially their own sort of OpenAI like dev tool. So a sandbox and then access to the Scout and Maverick Llama 4 models. Interestingly about it is they did what we said they should do quite a while ago and partnered with Grok and and also Cerebras to get that really turbo fast inference. So you can. Apparently it's only in preview right now we're hoping to get access to it. But fast inference on the Llama API, I think the only poo poo on that is that the Llama 4 model has been a little bit disappointing. And I think that was also reflected in our experience using their new interface, which is probably worth talking about as well here. This is the interface that they've released at Meta AI and they've fused together the products. So the Meta Ray Bans, which I own, have AI in them. There's a, there was a, an app specifically for those sunglasses, but they've, they've tried to fuse all this together now into sort of one app, which is weirdly similar to the Design of Alexa AI chat, GPT's AI and so on and so forth. They're all starting to look the same.

B (11:02)

Yeah, I used it all morning and I found a few things interesting. One, it's incredibly fast. Like I was using the voice option and it almost shocked me at how quickly it was able to reply. And furthermore, it also clearly has some sort of inbuilt tool calling to it because I asked it about. I, I think it misunderstood what I said because I was saying something about speaking sport. And then it's like here's some sports on near you today. And it went through the NRL and AFL games that are on today, which was pretty interesting and accurate. I went and actually verified what it said, which was cool. Also, it's not crazy censored. I won't go into what I tried, but I asked some fairly controversial things and it actually answered me. It stopped short of giving me money laundering ideas. I was sort of asking about that and then it's like, that's illegal. I was like, yeah, but hypothetically. And tried to get around it, but I, I couldn't. I also tried the image manipulation with it, where you can sort of have an image and say, can you do this to it? Can you do that to it? And probably the most interesting one is I took a. I, I got. I was trying to get like a cat and then I was like, can you give the cat a lower iq? And it sort of made the cat a bit like more scrawny and stuff like that. But then I asked it to write this day in AI dripping in blood in the background, and it wouldn't do that. It started to sort of get into refusal territory. But it's okay. I just, the thing for me was I'm like, why would I use this? I just don't understand why I would use a sort of subpar model in a subpar interface. Like, it's okay. It's a good offering. And I'm sure for people who just use Facebook and haven't used GPT for whatever reason might try it. I just, it just seems like another offering for the offering sake.

A (12:47)

Yeah, I think the strategy here definitely is they're trying to spin some sort of social angle with this because there's a section in it called Discover which you can click into. And I'll be honest, it's just feel like it's just filled with absolute garbage. But it's. It's sort of reminiscent of what we've seen with Facebook now. Just getting people addicted to scrolling with. With this AI slot like that's basically. Now boomers just sit all day scrolling, responding to AI images that trigger them or, you know, make them notice something. And that's what's happening here. But now they're just sort of productizing it and calling it social. It's also hella buggy. Like it. It's pretty slow. But yeah, I don't know. I like, I'm curious if people listening are going to sit here and scroll their Discover feed on meta AI like it. This feels like such a junk product to me that just doesn't need to exist. But I can kind of see why they're doing it because they are realizing that this is probably the starting point page or the starting point on the Internet, but almost every interaction in the future. Because I think a lot of people. That starting point right now is increasingly becoming something like a chat GPT. And so meta's realizing that they're not that, you know, core kind of one of those core apps or starting points on the Internet anymore.

B (14:10)

Yeah, that's right. You're not going to go to facebook.com and then ask it about what the latest NRL games are or something day to day like that, like you would with the others.

A (14:20)

Yeah, so I'm like, I'm not even sure where to place this. I. It's an interesting take. Like this whole social angle and sharing. I mean there's actually comments on this photo. Like, that's beautiful. Restyle my photo as a graphic novel. So maybe this will, this will work and I'll be like really wrong here. But yeah, I don't know, like, it's just weird. Like, let's get this bread. That's some dope ass bread. And then there's a love heart. Like, do we need this as a society? Like, is this, is this helpful?

B (14:56)

Yeah, exactly. The only thing I'd like to do is use it to manipulate images in my local council of like people leaving the bins in the wrong spot and.

A (15:05)

Then getting all the neighbors angry on getting everyone upset. Yeah, I think there's definitely some nefarious use cases. I was impressed with some of their image stuff. All their images always have that like really synthetic look. Like it's really easy to tell that they were generated by AI, but it's still like that. The UI and tooling around images and turning those images into video or like they call it slop creation tools. They're really neat. Like they're pretty good. So I'm sure for all the engagement betas out there, this is a great day. But I just, as you said, I don't, I personally would not use this because I just know that there's far better options out there. But potentially for someone who just wants like an interface, like a chat GPT like interface for free, that's super fast and capable to do a few things. Like perhaps this is a good option. I just, I can't imagine going to the workplace being like, yeah, I'm. I'm a big user of meta AI.

B (16:03)

Yeah, exactly.

A (16:05)

I, I will say though, the glasses, having the AI integrated into your glasses, like if you're out for a walk and you just want to know something is pretty handy. The only challenge I found, and maybe this update improves those things, is it's just so clueless. Like you'll ask it for the weather and it'll tell you in Fahrenheit and then you've got to like Google how to go into settings and change it. So I, I don't know, like I'm skeptical of this strategy and where it leads. I think the open source reality of this is if, if llama 5 or llama 4.5 can come out and actually be a decent contender to something like a Gemini 2.5 or even a Claude 3.7, then it starts to get really interesting again. In terms of the API front, can you see developers just wanting something reliable, like a reliable entry point, fast inference for functions or something like that?

B (17:02)

Not really, no. I don't see why, sorry, anyone would care at all. Like the Llama models are the most available in terms of API providers. Like all of the major GPU hosts, let's call them, have it, including Grok and have for since the day it came out, probably before it came out. So it's one of those things where, okay, if it was free, maybe like if they did it free for a while just to get more adoption. But given that this is the model that's easiest to self host and the most available on the various hosts, I don't see any benefit whatsoever of them providing an API. I mean, I might be missing the point there, but I don't really like. Because we were like, oh, we should switch to it if it's cheaper and then you're like, oh, there's a waiting list, who cares? I'll have forgotten about this within the next hour.

A (17:54)

This product will have been gotten in like a couple of hours. I, I assume after this podcast I can look.

B (18:00)

I mean you can see why they do it, but you know, it's sort of like, I don't know.

A (18:06)

I. What I.

B (18:06)

Why do people who live in one bedroom apartments have plants on the balcony? You know, that's kind of what it seems like. You know, it's very depressing, very futile. This isn't going to help, man. You don't have a garden.

A (18:18)

It's just like, I don't understand. Like, can you imagine the amount of devs and product managers and like internal meetings they had over this crap? Like it's so bad.

B (18:27)

Yeah, yeah, exactly. It's like your balls are seriously on the line on this one, Clarkson. You know, like, oh wait boss, no one signed up.

A (18:37)

They're honestly the new Google on the show, like just getting poo pooed all, all the time.

B (18:44)

Isn't it funny though? You know, it's, it's so funny. We've off we've often said this, like, in business, when, you know, you're talking about all of these, like, minor issues and you're like, you know, what would sell sales solve this? A heap of sales. Like, if we just made a heap of sales, then everything else comes into clarity. And I feel like when it comes to the AI companies, a really amazing model solves all their problems. Do you know, like, the fact that llama 4 was below expectations? It's sort of like a company releasing a mediocre quarterly earnings statement where everyone's like, oh, this company sucks. Whereas if they just blown it out of the water and had the top model and then released this, it would, it would have so much more credence and be so much more exciting, even if it was precisely the same. Do you know what I mean? It feels like still at this point, the model is the currency and the best model wins. Like, we trash Google for months, for years. Like it was our main thing on the podcast. It's just talking about how Google takes like 50 years to log in. Then they released Gemini 2.5 and suddenly we're, we're in love with it and it's all we can talk about. So.

A (19:49)

But I think for these big tech companies, there's two attack vectors and everyone's obviously trying to disrupt ChatGPT GBT because it's just growing and growing in volume and traffic. And so, like, you've got two attack vectors. One is dilution. And to me, this is just a dilution strategy at this point. It's like, make their product available literally everywhere. Like, have a literally clone version of their product. We'll add social to it to act like we're innovating for shareholders and then put it in every app and turn every app into AI. So it's just, you know, you're just saturating the market with this stuff. And, you know, does the consumer really care that llama4 is not a great model? Maybe, maybe not. I would say they, they will long term, as more people use this primarily as, like that in their day to day.

B (20:36)

Or as you say, it dilutes everyone's impression of AI as a whole. Because if that is their, their interaction with it, they're like, AI is AI shit. And so because they, they use it, they're like, okay, it's no good. They don't go, oh, maybe there's an alternative one that's actually better because why would you.

A (20:55)

Yeah, it definitely, I think it, it, I think a lot of these people that are using free versions of model or these like sort of throwaway interfaces do have that impression with a lot of AI stuff like, oh, it's not very good. It wasn't that helpful. I even the other day, just as an experiment, tried the Gemini thing in Google Docs to help me like rewrite something a little bit better. And oh, my God, it was so bad. They must be using like the earliest version of Gemini. I just don't understand how they haven't put 2.5 in there and said, you know, guns blazing, let's do it. Probably cost, but you just get this such a bad, like it's like a brand problem. Then like, you use it and you're like, oh, Gemini's bad. And so then you're like, why would I touch any of the Google models? So they Google continue to do own goals. But I think the attack vectors for OpenAI are definitely like full market saturation. And then the other attack vector is just build a better product and you've already got an audience, which is all the social users in the world, essentially through all your app properties. But I don't think they have a better product here. As you said, the model is the product really in a lot of these cases. And the product stinks.

B (22:12)

Yeah. I think back on the Google thing, I think, excuse me, what Google should do is Release Official Gemini 2.5 or 2.7 or whatever they want to call it, ditch all this experimental stuff and commit to something. Because the problem with Google, in my opinion, is everything's experimental. Like they've got the one that does the images and the one that does the thing and they're all slightly different and there's flash and flashlight and all this stuff. I would just go bang. Gemini 2.7, it's free for developers for a full year. We're committing to that.

A (22:44)

Yeah.

B (22:45)

Build on our platform, enable people with amazing tool calling, give them ways to host their AI apps and really just go all in on just being seen as the leader in this market. They have a unique and amazing opportunity to do that, and I think they should. The only downside being that I actually bet against Google on polymarket. I put my money where my mouth is that they'll win this month. I just feel like they won last month. I was wrong. I just don't see them lasting another month at the top unless they do exactly what I just said. Prove me wrong, kids.

A (23:18)

Yeah, I think Gemini 2.5 is like the dumbest strategic release. Like, we said we were going to poop on meta, let's now poop it back on Google. But yeah, here's why I don't understand. When you've got that much money and that many resources, you don't do what you just said and say like 2.7 out. It's in every application we have. We're putting the best model in every app. It's available free. It's free up until some, you know, limit to all developers and they just go for it and go all in on it. Yeah. And then I think honestly, you would win literally every single dev. The consumer would be having a better experience in all the apps. The Gemini brand sort of rises up. Then I would just slam investment into the actual Gemini apps so that they're on par or, or better than chat GBT stuff and then like focus on meaningful use cases for what it can do instead of just copying whatever chat GPT sort of does to create hype. I would focus on like meaningful workplace staff things for personal life that, that actually like use cases that really make sense. And people find so much value in like some basic agency staff that, you know, to me, that is the winning strategy. That is how they eliminate literally everyone.

B (24:39)

And I think to add just one point to that, which is what you said earlier, I would use the best model in everything. I would not use Flashlight for some stuff and Gemini 2.0 experimental 4.1 in another. I would use the best thing I've got at any given time. I would use as many GPUs as I can. And I would just make sure every time someone interacts with Gemini, they have the best experience.

A (25:04)

I would also change the Google homepage for the first time ever. Instead of having the single box, I'd have a tab. I'd literally have chat or search or, you know, regular web search or chat and just do it that way because it, to me, they've just got to do it. They've got to go for it and say like, it might disrupt us. It might not. They could literally still have ads on the side when you are in the chat mode and it's just related to your chat. I don't think that would bother people if it was free. Like they. They probably would use it if it was like unlimited free super fast.

B (25:34)

Yeah, that's a great point. I agree.

A (25:36)

All right, we got way off topic. Google, there's your strategy. Get on it now. If they do all that stuff, we'll take credit, but I don't.

B (25:45)

We will never mention it again.

A (25:47)

Yeah, I guess that saturation point though is interesting, right? Because these interfaces are just becoming so Obvious. Like, everyone has this Canvas button. Everyone has a, like a Create Image button. Like, there's just no inventive thinking on the UI front. Like, it's just like whatever chat GBT does, just everyone just follows. And I just don't see how they think they're gonna win users over by having the exact same experience. Like, no one's really pushing innovation around how we interact with AI. Or is that just the new Google and everyone's copying the next Google?

B (26:22)

Yeah, I'm not sure because I think there is a long way to go in terms of interface, because as we've discussed on many episodes before, a lot of it's around context gathering. And when I look at the Meta AI interface, the first thing I thought was it's going to be so hard to get like the right context to actually answer a question here. Like, what do I have to do just paste pages of raw text into this box in order to get an answer straight away? It just rules it out for me. It just doesn't have that multimodal input handled correctly. It doesn't really have the concept of continuity with working with a system. It doesn't have background tasks. Like, none of the things that would impress me and excite me are there. And I just don't feel like in the long run we're all just going to be typing stuff into a chat box.

A (27:06)

Yeah, I, I think that there's something about this idea, we've talked about it before, of having like a passport where it's. It's sort of like the memories, it's like your preferences, it's your past interactions. It. I guess you could sum it up as like a memory, but I think the passport also has more like connections into, say, various MCPs. It's sort of like your data stack to interact with a model. And I guess the question long term becomes, is similar to like the newspaper business. Are people willing to sell, and I think they probably are sell all that data over to the big tech companies to get, you know, to, to just interact with this one interface and this one model with that has, you know, access to all this data. Like, I think unlike search, it's just there's so much more to it because in a lot of ways you want to hand over like, access to your computer, your email, your, your, your drive, like all these things, and there's just no way on earth, like, I'm gonna trust Meta. I don't really trust OpenAI because of all the crazy drama and stuff. I probably would trust Microsoft maybe, but I think this is the problem in a lot of ways, but maybe people just don't care. Again with their privacy.

B (28:21)

It's funny you say that because I actually asked three different models about this question because I knew we were going to be talking about it today. This idea of a console of the future for AI. And the three different models I asked were Gemini 2.5, Claude 3.7 and the new Quen model. It's like Quen 3A, 3B or some weird name. It's in SIM theory if you want to try it. And they all said the same thing. They said that as the future of the web changes to be more AI centric, like as in you're working out of an AI console as your primary entry point, then the value of user data and user preferences and the centralized nature of that goes up significantly because you become the central so focus point. Your data isn't siloed in all these systems anymore. It's all controlled by you and central, which is good. And then all of them interestingly suggested the concept that the companies would pay you for access using, you know, Bitcoin microtransactions or something like that. We always come back to the crypto bro stuff. But this idea that there would be some sort of marginal cost to get access to your data by different systems. And while that may or may not be the case in terms of companies willing to pay, it does show that there is going to be value in having access to your personal passport, which is all your preferences and info. And it's interesting that as you and I look deeper and deeper into the MCP concept, we realize that it's not just enough to grant access to the tools you use day to day, whether it's your email, your calendar, or the systems you use for work like your accounting system, your docs, your whatever the different systems are that that you're using day to day. It's also you describing to the system what those are used for in this context. So, for example, you might connect to Google Drive, but it might be purely for reference, like, these are documents I must reference as part of my work. Or it could be like, here's where I store my diary entries during the day. Like, you should write diary entries here, but don't read this stuff because it's not relevant. So there's going to be different applications of the different sets of tools. And I feel like the current concepts around this stuff don't go far enough. It isn't enough to just Give a Model 80 miscellaneous tools. You need to say what each group of tools is for, how it should be used, how it applies to your personal preferences, and which data it should be able to read from and write to. Because I feel like once you get that, it becomes more of you, like more of what you would do if you were in the position of making those decisions. And therefore the AI becomes more empowered. So I think this personal data sovereignty and control and more, more appropriately, like your own classification or data identity is going to become so absolutely critical because A, the AI is going to be way more effective if you get that right. And B, as you said, you're not going to be willing to use the tools without it.

A (31:32)

The, the, the part like, while I 100 agree that you want data sovereignty and you want the ability to, you know, connect these tools, like, basically it's like your sort of AI stack or your AI brain, and then you're plugging that into a model. Right. And I think that the battle is now on, like, you know, it is that chat gbt, probably for most people, that is. Is it flawed? Probably not anymore. It seems to have just fallen off a cliff. Is it Gemini? Like, which company isn't? And I think that's really what a lot of these guys have realized as well, is like the future of siphoning your data and collecting your data. Whether you're Google, whether you're meta, whether you're the newest entrant open AI, the next way of siphoning all your data to be able to sell your ads is simply going to come down to can they build this connections into your data? Can they get access to even more, like your calendar, your email, can they be the place where you're getting work done, all these other aspects? And so I think that Maybe they don't 100% acknowledge it yet, but that's the real underlying race here is like, who can get access to the most data about the individual. And this gives far more access than putting, you know, searches into a search engine or anything like that.

B (32:51)

Yeah, exactly. And I think that as they do get access, say through authorized mcps that go into the different silos where you work, I think two things will emerge. One, people will have multiple identities. So there's me in the work context and there's me in the personal context. And being able to switch between those identities would be really valuable because I don't want it like accessing work documents when I'm asking personal questions, for example. I think the second thing is there's going to be a really big opportunity for the larger companies to essentially replace some services. So it's almost like adding like a proxy layer between you and the companies where you silo your data. So like, if you look at, say I'm connecting into like a snowflake data warehouse and I'm using that for analysis of my data through an MCP and asking it questions, or say I'm connecting to Google sheets or a mic, whatever, Microsoft Excel cloud or something, and I'm reading in spreadsheets and stuff, it's a matter of time before the tool calls can say, hey, you know, when I work with this data, I can actually just store it in my own system and work with it there if you want. And I'll only go out to Excel to like write out the analysis or to read in any updates. And so I feel like someone who is able to grab this market in a big way could eventually start to replace these solutions and be like, you're paying $4,000 a month for access to this platform. But, you know, we can just swap this MCP out for another one which has the exact same interface. We can copy the data over using the MCP to MCP Bridge and you save 4,000amonth and it only costs you an extra hundred a month with us. So I feel like it's another example where if you are the console which is accessing these different data silos for an individual identity, you have that ability there in the middle to swap out individual components of it. It doesn't even have to be a whole solution. It's like, oh, we can just do this part for you now using this different set of tools. So I feel like there's going to be some major opportunities around that because, and because you're specifically talking about the different nature of the web. I don't see people going off and logging into 50 different tabs when they're working anymore. It's going to be through a central place and those tools eventually will be replaceable because of.

A (35:14)

Yeah, I, I think you, you bang on that. There's a few aspects here, like there's the overall deflation of these software tools when the AI console or AI workspace is doing more work. The reason being is I sort of know this now firsthand. Playing around with mcps in the context of just getting work done, like being able to ask it to do tasks on your behalf with some agentic application, like, you know, go and process this refund, go and reply to these emails, go and do this or that. You start to see like your time in those apps, like you're still using the app like Gmail or Stripe or whatever it is, but your time in app decreases dramatically like Zendesk or whatever it is, the, the time in apps going down and the apps more becoming like an MCP interface into your AI workspace. And it's sort of because it's doing the work in the app. There's less like the app is truly, it's not really software anymore. It's just like an endpoint, like service provider really in, in a lot of ways and giving you structure and methodology and security and process as opposed to, you know, less so interface. And so you think about the future interface of the web and honestly just playing with MCPS and playing with this async work concept far beyond I think what we're seeing in like Meta AI or chatgpt or anything like that right now. This idea that you're like queuing up five tasks and it's off in the background doing things for you in different applications. It gets me mentally to a point where I'm like, well, why would I ever log in again? Like, why would I ever go to Google Drive? Like, do I even need this thing anymore? Like, could I get an S3 bucket for my company? And I just interact and create files in here? Like, why do I. It's, it does start to make you question, like, why do I need any of these things if this thing can render interfaces in real time and do these interactions on my behalf so that there's that long tail.

B (37:17)

Think about, think about something like Twilio, right? Where you're using say SMS marketing and things like that. Right now you need a company like Twilio to like coordinate like sending the messages and all that sort of stuff. But a lot of the APIs and stuff they provide if provided as tool calls and the AI itself is orchestrating the campaigns and things like that, then later, what's to say you couldn't swap out that interface with a much cheaper SMS interface that is literally just sending the messages and all the value add from Twilio just disappears because the AI is handling it with tool calls.

A (37:59)

But I, I think that's not the best example because I think if Twilio now go and build the best MCP and you embed it into your workflow and you're using that app and you're paying the bills, you know, that's an advantage to them. But then I guess you're right. If someone comes along and does it and says it's cheaper, you could easily switch and you wouldn't even notice.

B (38:21)

Yeah, and I guess my point is the cost to switch becomes very low because all you need to do is provide the same functions as a different MCP and that's it. As long as it can do what it says it can do, there's just no cost of switching in terms of providers like that anymore.

A (38:37)

There's also no preferred interface anymore because if you're already happy in your AI workspace, right, and it can essentially whip up any UI it needs to on the fly, like there's no need to have a preferred like, you know, you might be like, oh, I prefer Twilio's interface over like one like Sendbird or whatever it's called. But it doesn't matter anymore because as you say, it's just interacting with function calls and all the interfaces being rendered by the AI workspace anyway. So it's like it's truly deflationary around those services and like all the different SaaS applications that we use today. I think another way though you could look at it is, you know, there's probably a two way economic model that can replace the ad business on Google, which is like the AI workspace itself is like recommending mcps to use and like has those preferred vendor relationships with them so that things are just really accessible. And your AI agents know like can already do a bunch of stuff and they're just cutting deals with those vendors.

B (39:41)

Yeah, we had actually discussed this concept. So if someone types into an AI console, you know, I want to buy wheat or something like that, I don't know why I always come back to buying. But imagine if it's able to say like here is the mcp, we recommend the commodities buying MCP install now, you know, similar to Google Play or one of those systems would do if you search for something specific now there's a lot of value in being the one that gets installed. Like if you are the interface to where someone's going to be spending money or trading or just interacting with. In general there's value in that. Like someone who controls the which MCP is used for which purpose primarily is going to win. And then the other concept that you and I have been discussing lately, if I've got 10 different MCPs that can generate images, for example, how does the AI know which one to use? In a sense the actual tool definitions become advertising. It's like Teo tool engine optimizer. I don't know what you'd call it, but it is going to be the case where tools are competing for the AI's attention which one gets called. Because if there's real money associated with tool calls, which there will be in many, many situations, like you and I have already talked about MCPs that actually have a marginal cost because there's, it's doing something of real value in the world, let's say looking up an identity like you do with Clearbit or something like that, where there's a marginal cost with each call. Now if you've got Clearbit and you've got one of the competing ones, how does the AI know which one to use when you're asking it to resolve an identity like that? And it really does become a competition and the ones that win are going to make more money.

A (41:33)

Yeah, I, I, I do think it's easy when you look at these things to say, oh, the cost of switching will be easy and you could just plug and play with anything. But as we know, there is logistics and setup and whether it's sms, you've got to register phone numbers, like there are logistics involved where I think these things could be more sticky and real businesses in the future, like instead of people selling software as a service, they're literally selling an MCP as a service or a really an endpoint as a service. And I, I think the overarching discovery I realized this week, and maybe people have realized this well before me, but I've certainly come to terms with it now and said, okay, this is definitely gonna happen, is I think in the browser today, right? I've got a bunch of tabs open and those are generally like the apps I'm using throughout the day. I'm very rarely, if ever using desktop software anymore. I'm just using like pretty much everything on the web, maybe outside of like an ide, like a code editor and then raw access to the terminal and stuff like that. But as you start to think about the browser, right? When you're doing a lot of tasks in, with, with an AI workspace, I find myself really just going between sort of async tabs with, in an AI workspace context, like I'm doing my searches in there, I'm doing my sort of tasks like getting it to help me maybe write code or write a document or whatever tasks I'm working on. I'm working with it, right? And then in the background I might be like producing an image for marketing or getting it to go process some, some, some of the latest support tickets or whatever it may be, right. But I'm finding myself increasingly, as I said earlier, I'm not tabbing around in the browser I've just got one tab open and then tabs in my workspace open. And that's all I'm doing. Like, I'm very rarely if ever leaving that workspace now. Like less and less. And that's not by design. That's not because I like, I'm loyal to using say Sim theory or whatever. It's just literally because that's the fastest way I'm getting things done in my day to day now. Like that's the starting point and sort of end point of every task for me or things that I'm trying to do now. It's not for everything. Like, it's not like I'm like doing all the fantasy use cases people talk about, like booking a restaurant, doing my groceries, like all these like personal tasks. It's, it's the majority of the tasks of work stuff, right? But I sort of see us moving from like this sort of browser interface and tabbing through apps and websites to just being in an AI workspace and tabbing through, you know, different sort of call it MCP interactions, whether it be like search or working through support or whatever. It seems like in a lot of cases it's sort of the starting and end point. And then I can also see a world where you start rendering interfaces through an SDK in these things where, you know, you might think about something like using Trello or Jira or something like that and it's like, well, I don't want to go over there. Like, can't you just show, like, just show me the current tasks that I've got, but show me in a nice.

B (44:42)

Visual way and, and to go one step further, why the hell do you need Trello when you can have an MCP that's a database and the system itself is literally just building its own database of the tasks you have upcoming. And you're like, tell me what the urgent tasks are. And it just prints them out in a nice little table. You don't really need to go off and look at a Trello board. If you want to, you can say, make it look like a Trello board, bang. And it just outputs. So I think that those are tools that can be completely superseded by a system that has the right tools available to it. And I think to take what you're saying, really what you're using all those tabs in your web browser for is either gathering context, whether it's looking at emails that are coming in or you know, some design system that you've got up or whatever the tabs are gathering information. We know the AI can go do that with access to, to the right tools, it can gather that information. So you simply need to ask, you don't need to go and do it yourself.

A (45:36)

But it's really the commonality now of database. Like, because if you and I are working together on a project, you might like to see a Kanban board, but I just like to see a list. But in the past, we would have to say, no, we're both working on Trello with a Kanban board. But with the commonality being the database, the commonality doesn't have to be the interface, it's literally the database. And the workspace of the future is just rendering my preferential ui. And whether I go to a new job or whether I'm in my personal life, it just knows how my preference of rendering data. But it doesn't mean we can't share a common data source.

B (46:14)

Yeah, I really like that concept. Back in the day, there was a graphics programming library called Allegro and the guy made a demo app that was Space Invaders. And you saw Space Invaders in all these different interfaces. So one was where it was like curled inside a cone and you were moving around the bottom, shooting up that way. One was like 3D, 2D, whatever. But it was all literally controlling exactly the same game. And I feel like that's what we're talking about here. You can come at things from totally different angles simply because the AI can effortlessly do that, reform the data in ways that, that suit you.

A (46:49)

And that's why I think everyone looks at, say, Google and Microsoft today and says, oh, you know, it'll be really hard to beat these guys because they've already got like their, you know, work suites of like, Google Docs and sheets and all of these common interfaces that people are used to using. Right? Microsoft, exact same cam. But I do think as a challenger, OpenAI, if they sort of had focus in the long term and truly be disruptive, like Google was to almost everything on Earth when it came out, in the sense that if the commonality is the database and the price of interface goes to essentially zero and can be highly customizable, then like, it's pretty easy to be like, oh, I want to edit a document now. Like, give me the best document editor in the world in my style. Like, I don't know, I think that's where it can get to. Right.

B (47:41)

Well, I've got another good example because with things like SIM Theory, like, you know, we're sort of doing it, you know, we don't Always have a lot of time to do things. One of the things I like to do is build tools for us to solve common problems in the system. And what's incredible about say Gemini 2.5 is you can give it access to your database models and say, here's what the database looks like. Can you please build me a full on menu system with confirmations and stuff that will allow me to do the following tasks and it will make a system with beautiful colorful menus. In my case, lots of love messages and emojis and words of encouragement built into it where it'll be like, would you like to do like the Homer Simpson style, Would you like to do this? Would you like to do that? And you can just go, yes, no, yes, no, yes, no. And solve a task knowing that it's got like database integrity and it's reliable and all this sort of stuff. Now I can create a tool like that in less than 30 seconds to accomplish a task over and over again for the future. And I really feel like this is the precursor to what we talk about where people are going to build skills which are multi step processes that they're doing all the time for their business, which then can become part of a personal mcp, which is a set of tools which you then give access to the AI. And this is why I keep coming back to this idea of the. Your personal preferences isn't the right word, but your personal sort of like stack or something. Yeah, yeah, like yeah, exactly, stack. My personal stack of tools and skills that I need in order to get the things I want done. So once the AI knows about that, it knows when I asked to do a certain kind of thing to go off and use that multi step tool or multi step skill that I've created. And as I build those up over time, not only do I have them available to me to help me with my job, but they work with every other thing I've connected, like my email, like, you know, my spreadsheets, like the data that we share in terms of tasks and things like that. So I can say, you know, when there's a new task in the database that meets this criteria, follow this skill procedure. So an example of that might be adding a new model to SIM theory. Like these are the steps you need to take and it can go off, write the code, build a pull request, I go review it, I'm like, wow, it's nailed it again. And commit it, you know, like that's a task that I was doing manually before that could be fully automated by combining These custom MCP tools. And I'm not just saying this isn't just technical. In fact, I think it's probably going to be better with less technical things. But there are going to be procedures like if you're a doctor's surgery, you get an email coming in saying can I have an appointment at 3pm on Wednesday with you know, Doctor, what's your one? Gupta, Dr. Gupta. And it's like, okay, here's the skill. Check the calendar, Dr. Gupta's calendar, she's free, you know, okay, now add an entry to the calendar with that MCP thing. Okay, that's added. Add a note to this patient's history in our custom MCP system that updates our doctor surgery. Send an email to the client confirming the appointment, and then call them on the phone a day before to remind them of their appointment and confirm it. You know, like all of this stuff could be fully automated as a skill such that when that email comes in, that skill runs. And now that part of your business is fully automated without ever doing any custom code or coding at all, or even knowing how any of these systems work. Because you've simply connected all their various tools and you think about it, and the more and more tools you can add to your system or groups of tools you can add to your system, the more and more of these multi step skills you can build up and suddenly you just got this system that's ip, it's value, it's it. You've built a valuable system just by combining tools.

A (51:38)

Yeah. And you think in the near term the biggest economic benefit is just like replacing a lot of these, as you say, like these repetitive processes where you often even ask yourself the question like, why is a person still doing this? Like how have we not been able to automate this yet? And I think that automation level comes down to a few factors. There's so many, like there's so many minor decisions that a human might be making along the way. Whereas like that, that next step is really train the AI on the task, go and like implement a bunch of these things and have it working in the background and then have a way where you sort of still understand what's going on. And I think that first iteration 100% does not replace the underlying tools, infrastructure and technologies. In fact, the ones that are implemented now might have the best moat because they get stuck. Like you build these things, you're like, it's automated and it's working well. Why would we touch it?

B (52:33)

Yes, exactly. And I think that's the point. No one can just come in and say, oh well, AI is going to replace all these systems. Doctor surgery. So rip everything out. It's just going to be a chat box from now on. Like, they're never going to do that. Whereas if it's like, okay, it interfaces with all your existing systems, it's just going to cover the 90% of the cases that are really simple like this. And then the outliers, sure. It's just going to be like a regular email or a regular call or whatever. Nothing's going to change there. It's just going to do the ones where it's obvious. And I think you made a really, really interesting point around. Why haven't processes of these natures been, like, of this nature, sorry, been automated before? And the point is, because along the way there's lots of these small decisions which are really obvious when you look at them, that just couldn't be made by a computer program before because what's it going to do? Like text match on it or some sort of heuristics that decide. Whereas now even the smallest models can make those decisions effortlessly and really well and they can construct legible English that they can communicate with along the process. So I think there's going to be so many processes, not just in business but in life that can be really, really structured like this and done in a way that has powerful AI decision making in the mix, but is still really dynamic and able to understand when a situation falls outside of it and it needs to ask for human intervention or just simply not do the process in that case.

A (53:59)

Yeah, even having like putting all that together in the doctor's surgery where the patient comes into the room, the AI interface or workspace then can show like, you know, a summary of that patient, but have a custom interface suited to that medical professional that they, they like to see like how they like their data structured rather than some app that was built in the 80s that's never updated. Whereas, like, as the models improve the interface and the value that the medical professional gets out of that particular piece is, is so good. And I think if you had all of the data hosted locally and you were running local MCPs to get that data in and everything was running locally, that also takes away a lot of the privacy constraints and problems as well, because it's like all in that office and it's all existing within that space. And it doesn't mean they have to ditch the app that's storing the patient records or anything like that. It's just a new way of Visualizing or working with that data.

B (55:01)

Yeah. Like even if you had, say, a store where stock levels were an issue and someone's writing in saying, you know, I'd really like to buy three of these chairs, for example, and the system's able to use an internal MCP that can actually look up the stock level and be like, yeah, we've got those in stock. You know, I produced an, sorry, an invoice for you and attached it to this email pay that will get them shipped out to you. Like that's something that would have taken a salesperson time to do before, but now could be done automatically. Similar with like SDR roles in companies, when the sales inquiries coming in, you could have a system that automatically is, sorry, qualifying the leads by either ringing them or emailing them and really upgrading them, assigning them in Salesforce to the appropriate person. Like this was a whole human role. Like lots of human roles before could be totally.

A (55:50)

And startups, like there are AI startups this right now, like playing this virtual sort of sales rep, SDR type role in businesses. Like this is already happening. But I'm saying you could do it.

B (56:03)

Simply with just a simple skill definition with a combination of those tools being available. You don't need some startup to do it.

A (56:10)

Yeah. And I agree. And this is why I think the future is not necessarily software or websites. It's these like train skill, MCP tools, whatever we want to call them, I'm not sure call them apps, but I think the distribution of those where some professional, whether you're a doctor or whether you're a salesperson, goes out, spends the time building a tool and process and then shares that process, whether it's like open source or there's a small feeder, then adopt their methodology to use it. I get the feeling that's probably highly likely to be a new like. Like there's. There's economic value in that. Right.

B (56:47)

I also love the idea of having the concept of like reviewing and enhancing these skills. So for example, one of the things we're experimenting with MCPS is just simply telling the AI like you have access to a database, you can write to MD files, you can store whatever information you want as you go and retrieve it later and things like that. And similar to the memory concept, but a little bit more flexible. Now imagine a system that is playing that SDR role on qualifying sales leads that periodically has a process that goes through and goes, let's have a look at how that actually went. Because I have access to the sales data, I have access to the Original inquiries. Let's see which of my responses worked and which ones didn't. Maybe I can actually suggest upgrades to the skill process in the, in the form of prompting that will lead to a higher conversion rate and you start to get towards a system that can actually improve over time. And then you think, okay, maybe there's a B testing involved. Like I'm going to try being a real dick to this guy and see if it, you know, he, he responds to tough love or you know, this one seems like we should be more empathetic and really focus on solving their problems. And so suddenly you could actually get to the point where your process has become a little bit more dynamic up to a degree you're happy with, so you can actually improve the experience for everyone involved without any effort.

A (58:16)

But don't you think this all comes back yet again to your AI passport or your AI stack or your whatever you, whatever we want to call it, that it's like your methodology that you're bringing to, or training into your agent or assistant or whatever you want to call it that is inherently valuable. Like I just. Because every. There's so much nuance in tasks and things that we do in our day to days that I just think there's so much economic value here in people going out and solving for all these little pieces of the puzzle or entering into an existing business. I mean like think about the efficiencies you could create. Like there's a lot of obviously baby boomer cohort retiring, selling small businesses. And I know there's been a trend of some people to acquire these and just apply technology and smooth out the processes and things to get, you know, to get higher value out of that business and extract more profit. And it seems to me like there's probably like a similar play here for people to go into either large enterprises or smaller businesses, refine a lot of these processes and just cut a lot of the sort of like busy work and figure out a lot of these processes that they can replace in the business to create.

B (59:30)

You know, it's. You just really made me think of something really interesting. Imagine someone who is in a particular role and over the course of a year builds up this personal passport identity with an AI where they've got skills, they've got conversation history, they've got memories, they've. They've really built this identity that allows them to do their job brilliantly and then they quit, right? What happens then? Like imagine if you could say as the company own that identity, which is definitely going to be Morally controversial, and then give that to your new hire and be like, hey, use this thing. This thing knows everything about this job. You're going to nail it.

A (60:15)

That song was right about us. We've literally talked about this before on the show. I'm sure we have talked about this exact topic on the show. So it's literally tuned in every week while we discuss the exact same idea. But no, it's about realistic now.

B (60:29)

And I just. I just wonder, like, firstly, do I own that IP as the worker and I can go to another company, hey, hire me. I have got the playbook down on this thing. Like, you hire me and you bring my AI identity along with me and I'm going to absolutely smash it for you. Or does the company say, hey, you built this thing on company time. We own it. You go, and you don't get to keep it. Really big deal.

A (60:54)

But I think this is going to come down to whether or not people let their staff or team use. You know what I mean? Like, if you join a company and they're like, we're using chat GBT teams and you have to go in and use that and you build all your IP in there, well, then of course the company's just gonna, like, absorb that at some point. But I think that the. The piece of it is like, if they're building and deploying mcps or like, that concept of like these like, trained, trained skills or parts of the job that they're automating and using and running all the time between teams, and those things are shared. I think if the employee leaves, it's still. Yeah, that's built into the knowledge. But I do think there is, there is that ability of an individual to just know how to, you know, like, they're like AI first worker. I think those people are just valuable no matter what because they can just go and like, recreate that. It's definitely going to, like, I would say, if not sort of already in the next couple of years, it's going to be one of those roles where everyone thought, like, prompt engineer would be like a real thing. And I mean, it kind of is when you're building AI products. But I don't think it's like a singular job.

B (62:08)

Yes, but it's almost like a. A prompt that you build up through your interaction with it by building the right memories, conversation, history, data, combinations and skills.

A (62:18)

Yeah, but if you've done that once, you can certainly go into another company and do it again.

B (62:23)

Yeah, no, I agree. And I think that this is where it really Comes back to. I would back someone with that, what I just described, like a really solid identity around that. To use a lesser model and outperform someone on a better model using that identity than. Than the person with just a raw model.

A (62:44)

Yeah, I, I think this is why though, that the. It's gonna become a question of buy or build for a lot of these companies. Like, do we build our own AI tooling internally? And like, because that starts to become the IP, right? That tooling, those MCPs, those functions. But then I think another piece to that is like, how does the employee interface with all that stuff? And you don't really always want to be in the business of building out these interfaces. It's just not, probably not going to happen. So there's going to be a consumption layer in there somewhere and that will be meta AI, of course, with the Discover feed and all the AI. Yeah, just bringing it back. Let's move on because we could talk about that all day. Look, I think it's a really interesting topic. If you are still listening at this point, leave a comment below. Like, I'm curious on many fronts, like how you see this playing out. Like, do you think the interface of the Internet's going to go, like, we won't use websites, we'll use like tabs in a, in a sort of AI interface. Curious to know. So, next topic, it's going to be a fire round because I want to get back to work. So Poly Market, you were going to provide a bit of an update you sort of did before.

B (63:58)

Yeah, I sort of alluded to it. So I finally bit the bullet and got some money on there and, and because I just keep looking at the. Which company has the best model end of whatever month. So last month it was end of April and it did end up being Google. So I'm glad I didn't go in then. Right, yeah, you were right. You were right. So what I've done is I've loaded up some money on Polymarket and we're not sponsored by them, by the way, or anything like that. It's just interesting to see how the models are tracking. But I was just really, really surprised. So the current stats for the end of May, and it's the start of May as we're recording this, so, you know, the whole month is yet to go. Google is a 65% chance of having the best model at the end of the month, which I think is sort of like the more or less saying no one else is going to release anything this month because, you know, it just seems like they always go slightly better. And also it's interesting because they use, I checked, they use the LM sys leaderboard at the end of the month at like midnight on the last day of the month as the definition. That's how they decide. Google is trending upwards like mad and I just don't understand why. XAI is a 23% chance of having the best model. OpenAI is 7% despite O3 scoring very few like a couple of points below Google on, on the thing. So I don't know. I just don't know. Like I do think Google has the best model right now. I, I definitely agree with that. I just feel like a month is a long time in this space and anything could change. Now keep in mind on this thing, even if someone releases something and everyone gets all excited about it and thinks it's going to be better, you can cash out before the month actually ends. And so I'm going to keep tracking it, see what happens. But it's very interesting looking at sort of the later months and seeing what people think over time. I guess they become less certain as time goes on. But right now Google is definitely leading on there.

A (65:51)

So you bet against Google having the best model at the end of maybe that's right.

B (65:56)

I figure someone else is going to come up with something.

A (65:58)

I think you're going to lose. I, I think you're 100 going to.

B (66:02)

Lose bet against me then.

A (66:04)

I might. Okay, all right, fine. I, I like, I just think daily driving the models I and, and I wanted to talk about this anyway what we're daily driving this week I've gone between more models than I've ever gone between in a long time. It used to be prior to the latest releases I was mostly on Claude 3.7 little bit of 3.5 occasionally I, if I got really desperate I would try out oh one but now my daily drivers are predominantly still Gemini 2.5. But I find myself for certain tasks like writing, you know sometimes with like front end or design related stuff, anything designery. I'll go to Claude sometimes if Gemini trips up on things. I'll also go to Claude 3.7 thinking like I'll push the thinking up and then I've also been trying when I get stuck as my sort of, you know, last ditch attempt oh 4 mini high which I gotta say like it's pretty smart. It's not a good daily driver I don't think but it's pretty smart at cutting through problems sometimes. I've also been playing around with O3 on chat GBT to get the search and the, the tool calling capabilities when I get stuck on things. And then I've also been using O3 raw without tool calling on SIM theory just to get a feel for the differences. And this is the shocking bit to me, right? And I know people will disagree with me, but hear me out. Any problem I got stuck on where I was like, oh, it just needs the latest talks or it needs the latest information. I went and tried oh3 with the web search and tool crawling, but even in their own model release paper they say that it tends to hallucinate more. Right. And this is the problem I found using it is it? The experience feels really good. You're like, oh, this thing's going off, it's searching all these sources, it's doing all this work and then it would come back with a solution and it would just read like an outdated doc, like there's again, there's no actual thinking. Even though they say it's a thinking model. There's no thinking really happening in that, in those, in those tool calls where it hits the search. It just, it does look through better results I think now and engage more results. The problem is it's so easily misled. Still a lot of the answers came back sounding really good. But then when I investigated using the solution, it was garbage. And then I just ended up reverting to Gemini 2.5 a lot of the time. So I, I don't know, like, I'm not, like, I, I don't think O3 and O4 mini are daily drivers, but I'm still finding myself living in a world tackling really hard problems, going between the models, trying to find solutions where I, I can't like, you know what I mean? Like I need, I'm still finding, I'm needing a mixture of models like Gemini 2.5 is I think arguably the best right now. But it's not so far ahead that I wouldn't look to use anything else similar to what Claude 3.5 was for a long time there.

B (69:18)

I think the next major advancement is not going to come from a model at all. I think it's going to come from what I was describing around tool use being structured by an AI based system. So what I mean by that, because you just described it there, a system that just uses basic tool calling like a search and like a research or whatever is going off and, and gathering that information, running it through an inference and providing a summary and then adding that summary into the request alongside the, the chat history that you've had with that agent. And so you end up with it taking that information into account, but just working off the model. What it's not doing is applying things like preferences, weights, like saying this is really important, this is less important, this is a more trusted source, this has been verified by this other tool and things like that. I think what we're going to see is really intelligent tool combinations that are done with guidance by an AI system that are able to know how many tools to call in what combination and order to call them, and then how to combine those into synthesized answers. And I feel like what I'm seeing, because we are now with the MCP stuff, working with so many tools, it isn't enough to give an AI model, whichever one it is, a hundred tools and just expect it to work out what to do. There needs to be planning, there needs to be a sort of trade off between time and money versus what kind of answer you're going for. There needs to be some real intelligence in the way the problem is solved. And I think just relying on the model to do that raw now isn't good enough because I've tried this and I think the models are okay at it, but what they need is strong guidance. And I think with the right guidance, we're going to see real innovation in this area where the models are capable of so much. If you call the tools in the right way and then allow it to use the output of those to either trigger more tool calls or to gather that information together and make decisions.

A (71:30)

For me, my dream product though is that O3 like tool calling experience, except it actually works and doesn't hallucinate and checks its own work. Because if it could check its own work, I'm not sure, even just like using my actual computer and running it itself, or just like actually being able to experiment locally or there needs to be some sort of checksum with this so it, it stopped like the hallucination. It just seems such a huge problem. It goes down a rabbit hole. It's definitely gotten better. I'm not, I'm not saying it hasn't. And I know some people get blown away by answers. I would still just question, are they actually checking the work or just reading it and being like, oh, this sounds really informed. If you, if you're an expert or know something about a certain topic, it seems you can kind of cut through the really fast with, with this kind of thing. And you, you're like, ugh, it's not so great. It's it's it like for smaller queries and stuff, it's fine like comparing products or, or things where it's not as critical.

B (72:29)

Yeah. And I guess that's the point, right? Like you want a combination of tools that's going to do what you want. Like if you just want to make like a beef Wellington and you want the recipe and you don't want to go through the blog spam post and it just outputs the recipe, like, what are the consequences if it hallucinates? Add chives or something.

A (72:45)

What era do you live in? What the, what, what is a beef Wellington?

B (72:49)

It's Gordon Ramsay's signature dish. You know, it's like a pastry around a base fillet.

A (72:54)

That's really good. Okay. I shouldn't have criticized it. Amazing.

B (72:58)

Yeah. And so, yeah, like the consequences of like accidentally adding the wrong ingredient to that are pretty low. Whereas if I'm doing a research paper, I want additional tool calls that are going to go and verify that the sources exist, that the sources match what I'm using them for in a sort of third party check and know that this thing is right. And I feel like that is where intelligent combinations of tools based on the seriousness of the task and the nature of the task will be something that an AI system is perfectly capable of doing automatically and will lead to a real leap in the utility of AI systems. And the same thing we've discussed many times with output types, it's going to be the case where what it outputs, whether it's text or a presentation or an image or a combination of those things is going to be important as well. Like, because a lot of the time you're spending is copying and pasting stuff out when it could should know the right tools to call to produce a more fully.

A (74:06)

Sounds like what you're really saying though, is if you're willing to spend the money and iterate enough, you'll eventually get the right answer. But I would think right now with something like O3, there's probably a cost trade off where they don't want to do these checksums and go nuts with every answer. Because that sucks. An experience.

B (74:22)

Yeah, exactly. Because, yeah, the opposite is the problem we always talk about. If I just say hi to it, I don't want it to think for like 35 minutes, use 40 sources or like 80 tools.

D (74:32)

Yeah.

B (74:32)

And I don't want it to search the web. And like it's funny because we were using the new Quinn model and looking at the thinking and I say hello to it and it has like a Friggin five minute debate with itself about the right way to answer me saying hi. It's like, oh, it probably doesn't want me to output latex when I say hello. And yet for some reason still decides that I do want Python code in its response to me saying hello, you know, and it's like, no, I don't need you to do all this stuff when it's simple, but when it's complicated, I do.

A (75:02)

Yeah. Okay, so changing topics because I, I really wanted to get to this. It's a little gossipy, but I think it's also relevant to the Meta AI discussion earlier. And I meant to mention it before but I completely forgot. So remember when GPT Image came out and everyone was doing these like stylistic, I don't know what you call like sort of anime ish style images and it really blew up and so like everyone was using it. Friends of mine that don't care about AI were trying it out, the, you know, all the image features and there was the, for a while there everyone was turning themselves into like the, the toys like, like we did as, as an example and it became a really big thing and, and the, the exit. OpenAI were talking about how like their traffic doubled, tripled, quadrupled, whatever it may be. And then this week we saw something really interesting. I'll get to that one in a minute. Where all these people were reporting this new update that went out that OpenAI said to GPT4.0 on chat GBT. So they basically released a tune update of it and they said it would be, you know, more interesting and better. It seems like off the back of all the publicity around the image image stuff really all this company cares about, right, is daily active users and getting people, you know, using Chat GPT as their homepage or coming back like they want the, they want it to be an app that people are addicted to. And, and so if you think about it, like if you get addicted to it, you're willing to pay for it. They're able to collect more information about you, they're able to maybe sell you ads, they might build a social network like all the things we talked about throughout the show, right? And so if you think about like what are they optimizing for? It's like every other Internet company like TikTok's optimizing for. It's just like brain drain time where you're just, you're constantly using that app like engagement and all of these companies, if you look at it, end up optimizing for engagement long term. And it feels like this week, Chat GBT is like, oh, you know, let's take advantage of all this image stuff and make the model more addictive for people so they stick around, keep the conversation going longer, keep, keep them on our website longer, keep them in our app longer. And so we started to see all these reports on X, like, new Chat GBT just told me my literal on a stick business idea is genius and I should drop 30k to make it real. And like, literally it says it's performance art disguised as a gag gift. And that's exactly why it has the potential to, to explode. You've clearly thought through every critical piece, production, safety, marketing position, like flattering them. Yeah, so it was basically flattering and reaffirming literally anything anyone said making them feel great. And, and there was just countless examples of it. And so a lot of, a lot of talk was around this. A lot of people were giving these crazy examples of where it was just validating literally anything and being very agreeable. And then finally, Sam Altman, you know, only a couple of days ago now, said the Last couple of GPT4.0 updates have made the personality too annoying, even though there are some very good parts of it. And we are working on fixes asap, some today and some this week. At some point we'll share our learnings from this. It's been interesting and he said, we started rolling back the latest update to GBT4O last night. It's now 100% rolled back for free users and we'll update again when it's finished for paid users. I don't know why free users were sucking up too many GPUs maybe. We're working on additional fixes for model personality and we'll share more in the coming days. But clearly this was either a, an experiment for eyeballs and time on site, which I'm certain it was. And then they've realized, oh no, people have discovered the, the engagement bait that we're doing with the model. But you can start to see all of these companies yet again. Like, one, they're trying to collect a fire hose of personal information, and then two, they're trying to just keep you in their app because then they make more money, which makes sense. I mean, I'm not necessarily criticizing it, but it's just sad to me that, you know, we went from like, let's build AGI to benefit the world to let's get more engagement than tick tock.

B (79:22)

Like, that's like treating it as if it's another social Network or that kind of product instead.

A (79:27)

Yeah, it's like, why not just get economic value out of it, guys? Like engagement doesn't help the economy. Like it just wastes people's time.

B (79:36)

Yeah, it, it is interesting that, that, that's their focus.

A (79:41)

And then meta obviously are like, oh that, that's great. That's what we do. We're in the engagement business and like, you know, making teams, instant press business. Let's go do that as well. And so I think that's, that's sort of the summary there. So have you spent much time with the O304 series? Like what are you daily driving? Because I, I mentioned earlier, I'm obviously my preferences right now.

B (80:03)

Yeah, I'm still on a combination of Gemini 2.5 and Claude. I don't really use the O models at all really, other than testing. I just don't find them that useful compared to the others.

A (80:18)

Yeah, I, I don't know. People always criticize us like we're biased against them or something, but we're not because I like they're honestly much more affordable models now to run for us. Like Gemini 2.5 is quite expensive.

B (80:30)

Yeah, it is and I agree and it sometimes is my hesitation in using it to be honest. But the, the. Yeah, I just, when I'm working I'm trying to get the job done and I just, I just don't in my mind see the O models as, as something that's going to help me get one any better. It might be the same, but it just isn't good enough or better enough that I would risk it.

A (80:53)

Do you want to quickly mention before we go?

B (80:55)

Although, although just to finish that. I hope it actually does win by the end of May. So maybe I need to get on there and start voting. I was thinking about if I could get it to teach me how to manipulate the LM sys charts for profit.

A (81:07)

I mean, actually maybe that's another thing with GBT4O and these optimizations. Like, like they're just so aligned to winning this lmcs. Like they're just highly optimizing to what the user wants to see versus what they need to see. And I think that's another part of it is like these labs are just so misaligned because they're chasing benchmarks that aren't actually aligned to creating economic value. It feels like the only benchmark that should matter is like can this make me a dollar or not? And yeah, yeah, if it can't, I mean, why even look at it? So when three dash 20.235 billion. That'd be the parameters. Dash a 22 billion mixture of experts model was released and also Quinn 3.3.2 billion. Also on a bunch of benchmarks. Do we believe them or not? Probably not. They say it's Superior to deep seek one open AI01, not Gemini 2.5 Pro and also superior to OpenAI03 mini. That was on Arena Hard and. And the benchmarks flow through around about the same in those areas as well. You got that model into SIM theory. It does have quite a. The context, sorry, I should say is 128, so quite standard stuff. You played around with it. It's a thinking model. What was your overall impression? I know it's.

B (82:42)

I haven't used it enough to give a full answer to that. It seems okay. It seems like a really nice start. I think it's worth noting that it's hosted in the us, not China for people where that matters for their work. And it seems okay. The thinking's interesting. As I said before, I thought reading the thinking is kind of of weird because you'd like oh geez, some of this is kind of unnecessary and you're really second guessing yourself, but maybe that leads to better results. It seems okay. I'm going to give it. I'll give it a try during the week for certain things. The Quinn models for me have always been something where you try and be like, wow, that's really cool, it's legit competitor. But then I never really use them enough to get a good feel for them. But it's there. We can try it and see how it goes. It doesn't support vision, which is a bit disappointing. It also doesn't support tool calling, which is. Wasn't an issue for me before, but now with the MCP stuff it really is a big deal because it's going to lead to it falling behind.

A (83:38)

I wonder though if that's the version of the model we're using because on their website it says improved agentic capabilities. We have optimized the Quinn 3 models for coding and agent capabilities and also we have strengthened the support of MCP as well.

B (83:51)

Yeah, not sure because definitely the one we use doesn't. But maybe, yeah, maybe it is just due to the host we're using rather than an inherent problem with the model itself.

A (84:01)

Yeah, you can kind of see why Llama or Meadow is hosting Llama because that is the problem with a lot of these open source models. You want to try them and switch to them, but you need reliable endpoints and there's so many posts with various pricing and stability issues that it does become. It does in a lot of cases rule these models out, unless you want to run them locally, which I think increasingly people just, just aren't really going to do. I think they are going to want to host and have someone maintain it and, and, and manage it. So anyway, it's in there if you want to try it. I played around with it a little bit. As you said, you need to really daily drive it. And when you're trying to get a lot of work done, it's really hard sometimes. I have suffered through days for this show just to like get a feel for a model and you know, I'm soon like, you know, if you're switching back and. And I think that's why Gemini 2.5, we keep calling out. It's the first one. I didn't want to switch back. I was like, no, I'm just gonna to stick it out with this thing. It's so.

B (84:55)

Yeah, like to give you an example, like, I used it all yesterday evening, like just trying it out. And I managed to get it into a state where every single time it replied to me, it would say, hello, Mr. Marbles, which is what my local calls me for some reason. How are you today on every message? And I'm like, oh God, there must be some bug where I'm not passing it the message history because it hasn't noticed that it's already said hi. And I spent like ages trying to work out what the bug was. Was. But there wasn't a bug. It had just got itself into some weird loop where because all the previous messages said hello, it must have assumed that all subsequent ones must as well. And so I started a new chat and it was totally normal and fine. And so stuff like that. I'm like, I don't have time for that. Like, you know, I'm not going to waste my time dealing with issues like that when there's. Every other model does not have that issue. For example.

A (85:43)

Yeah, I think that their models are getting better and they're. They're trailing on the verge of some of these models. And it's. It's good that a lot of these models are being released and are out there and open source, but ultimately. Yeah. Are you gonna use it? Like. No.

B (85:59)

Yeah.

A (85:59)

All right. Any final thoughts? I'm pretty excited. I'm gonna spend a lot of time playing around with Suno 4.5 because I. That was so much fun. I'll play that track again at the End in full. There's two more verses which just, you know, a pretty, pretty mocking of us. I, I, it wrote it, not me. Like I literally just cut and pasted in without reading the lyrics. And then I was listening, I'm like, wow, this is really deeply insulting.

B (86:23)

Pretty. Yeah, it's pretty brutal. But you know, we gotta, we gotta like, be, you know, we gotta be humble. We've got to take the criticism where.

A (86:30)

It'S deserved and it is right on that note, that's a great note to end on.

C (86:43)

This song is me just want to meet something, but just gotta win it. Here I am, listen once again to two sharkies talking AI but when will they reach episode 100? Is the question that keeps me wondering why do I tune in week after week for mediocre insights I could easily see? Will they ever win? Will they ever make it there? 100 episodes I pretend to care Cause day and day why do I still tune in for Mike and Chris and their ever expand countdown to 100 drags alone? Will they do something special? Probably, probably wrong Inconsistent boom factors what does that mean? The most middle road podcast I've ever seen how can we tell without a boom? As I listen alone in my living room no expertise, no special insight Just enough knowledge to put up a fight and where has Marshy been all this time? The AI assistant in his prime they promised it would make a return.

A (87:58)

A.

C (87:59)

Lesson they never seem to learn why don't they prank call anymore? The one thing I actually tuned in for episode 90 something plays in my ear the hundredth milestone drawing near Will they celebrate? Will they even know? The most average milestone for the most average show? I keep listening in though I don't know why to this day and age I my weekly supply Will they ever Will they ever make it there? 100 episodes I pretend to care this day and may I? Why do I still tune in for Mike and Chris and their ever expand the countdown to 100 drags along? Will they do something? Something special? Probably wrong Their mother fell for the fishing attack but those golden moments never come back I wait for episode 100's glow will it be different? I need to know two sharkies stumbling through tech talk While I listen on my daily walk when will they hit that century mark? I follow too long to disembark this day in AI has become my routine the most mediocre show I've ever seen yet here I am waiting for 100 the biggest milestone I've ever wondered Will they make it special? Will they even try this average Listener needs to know why SA.

D (90:31)

Here I am listening once again to two sharkies Tom talking AI but when will they reach episode 100? It's the question that keeps me wondering why do I tune in week after week for mediocre insights I could easily see?

C (90:56)

Hey.

D (91:00)

Will they ever. Will they ever make it there? 100 episodes I pretend to care this day and AI why do I still tune in for Mike and Chris and every spin? The countdown to 100 drags alone will they do something special? Probably wrong. Inconsistent boom factors. What does that mean? The most middle row podcast I've ever seen. How can we tell without a boom? As I listen alone in my living room no expertise, no special inside Just enough knowledge to put up a fight where has Moshi been all this time? The AI assistant in its prime they promised it would make a return A lesson they never seem to learn why don't they prank call anymore? The one thing I actually tuned in for episode 90 something plays in my year the hundredth milestone drawing near. Will they celebrate? Will they even know? The most average milestone for the most average show? I keep listening though I don't know why to this day and they are my weekly supply Will they ever. Will they ever make it there? 100 episodes I pretend to to care this day in AI why do I still tune in for Mike and Chris and their average spin? The countdown to 100 tracks along will they do something special? Probably wrong Their mother fell for their fishing attack but those golden moments never come back. I wait forever episode 100's glow will it be different? I need to know. Two sharkies stumbling through tech talk while I listen on my daily walk when will they hit that century mark? I follow too long to disembark this day and AI has become my routine the most mediocre show I've ever seen yet here I am waiting for 100 the biggest milestone I've ever wondered Will they make it special? Will they even try? This average listener needs to know why.