B (3:55)

Yeah. And look, I might be naive, but we talk about the daily active users and their brand recognition as their main advantage. Right? Because, like, you can't argue only losing 6%. I mean, that's nothing like it's a. They're still got. What did you say? 800 million daily active.

A (4:09)

I think the other thing is you've got to look as well, like the pie is growing. Right? So, like, if I bring it up on the screen now with similar web, it's like generative AI tools and the market's growing, so it's getting bigger. And yes, their share is technically less, but the market is bigger.

B (4:26)

Yeah, yeah, no, I get that. But in a way, like, what worries me or what would worry me if I was in that business, I'm sure I wouldn't be worried at all. I'd just be on my yacht or whatever. But like, let's say I was in there and I still cared for some reason I'd be worried about the fashion nature of it. Like it's fashion, right? To use ChatGPT, it's, it's a fashionable thing to do and like use it in my work, use it in my life, whatever. The problem is that that kind of cheap AI where you're not really invested in it, like you're not building detailed agents, you're not, you know, working with AI systems, you're simply just using like a chat prompt. It's so easily replicated for free in other products. Like as you mentioned earlier, like the, the Gemini built into Google is pretty good. Like it actually does the job pretty well. Like a lot of these basic AI assistants you can get anywhere are pretty decent. And I just can't help but feel like the casual user in the long run is just not going to bother with ChatGPT and I'll just use whatever's easily available on their phone or their desktop or whatever it is, right? And to me, the real users that matter are things like massive organizations, governments like big industrial corporations and things like that who need controls, who need reliability, who need trust, who need security, this kind of thing. And those big contracts are going to go to someone. And I just don't know if OpenAI is going to be that company now just because of the way they've diluted the, and, and damaged their reputation by going in all of these different directions. Not to mention, you got to remember, if you're a regular open AI API subscriber, there is a single checkbox in the UI which says you can train on my data and you get a million free tokens or something a month or something like that, right? If you check that box even once, if one person on your dev team, and that's anyone who's an admin in your OpenAI checks that box even once, they're going to train on all of your data, right? Like that's a deadly serious security risk. And who, who would mess around with a company who makes it that fickle to suddenly train on your data? Like, I don't know about you, but it seems to me when I talk to people in serious decision making positions, one of their biggest fears is their data being trained on like that. Like that is the big issue. And these guys haven't even really got an answer for that.

A (6:47)

Yeah, and I think you raised two interesting points. One is that chat gbt, it hasn't, they haven't as a product, figured out a way to go deeper into your life than what I think was promised. Like, it's not really a great personal assistant. It's a great chat model and a great chat experience. It's very fast, it's accessible and available everywhere. I think they still have arguably, you know, top, maybe second now image recognition model. So I think as a platform, it's great. There's some features in there that, you know, were very much hype, like deep research and a few of their sort of agentic things like the computer use stuff, the browser, all that kind of stuff, but nothing's actually stuck at all. Like, there's never been a moment where they've gotten deeper into your life. It's sort of like every time you go to chat GPT, it's like a Tinder date. You go there for kind of one thing. Right. You like my analogy.

B (7:44)

And then before my time.

A (7:45)

Before my time as well. But I'm still using it.

B (7:49)

So just for research.

A (7:52)

Yeah. So you, you go there for one thing and you don't stick around. This is not a relationship tool. And so I think that's the problem, right. Is like they never got deeper with the consumer. So if Deep Sea comes out with a model, we'll get to later. And Anthropic has a model or Gemini has a model and people hear about it and that's just as accessible and the cost is free. I think the consumer is lost. It's almost like the newspaper business model, in a way is like, free is a great price and the models are just getting better. So that consumer side of the business, it's always going to be tough and eroded. The only way to do it is to have all the daily active users and sell ads. But if you deploy ads too early, like they're finding out, or maybe about to find out, everyone's just going to go to where there's no ads.

B (8:43)

Yeah. Nothing is going to piss off the hang on the average AI user than injecting ads into the responses from your AI models. Like, can you imagine how quickly you would switch away from that model if that started to happen? It would kill you. The other thing I think is important noting when you look at the financial side of OpenAI, which YouTube is all abuzz about, everyone's talking about, oh, they've agreed to trillions in contracts and then how are they going to pay for all that? Now I think the problem is they're entering into these big boy contracts because they need the hardware, right? A point you made earlier where you know, Google has their own TPUs. It's their hardware, they're making it. So that's a bit of a different proposition. Amazon is the, the sort of king of that sort of stuff, so they don't have to worry about it. Microsoft has Azure, so they don't have to worry about it. But OpenAI does. Like they owe money to people and they have to do these agreements to get good prices. Now the problem is on the other side, if all your customers are consumers, you don't have solid contracts with these people. They can quit anytime they want, right? Whereas if you're doing deals with like Coke or Pepsi or like, you know.

A (9:51)

I love how always your enterprise examples are Coke and Peter.

B (9:54)

Well didn't they use that in the early days? All right, fine. The government of Uruguay, like Manchester United, the soccer team, right? You're doing contracts with companies like that. You can do two three year contracts where you might not get the cash up front, but you've got that guaranteed revenue to match it against what you owe these other people. Like that's how real businesses operate. They don't screw around with like, okay, now we're announcing this feature for free to burn a bunch of capital.

A (10:19)

But I, but I think the anticipation from OpenAI was like, we're the next Google, so let's try all these different projects and not worry about the core experience. But the core experience was never delivered upon or figured out like they're, I think this is the problem, right? And a lot of the statistics cite the sharp decline on weekends of chat GBT usage to show it's not really a part of your personal life as much as it is professional, where people are just going to chat gbt, throwing like, I guess company data or whatever into it and getting it to do stuff or if you're a student, getting it to write emails or essays or, or maybe a bit of chat or learning. You know, there's a ton of use cases obviously. But I think that's the thing is.

A (11:06)

To be honest, I think this stuff's way overblown. Like they have the lion's share of the market still. They're only like one good image model and chat model away from being back in the zeitgeist and everyone hyping them up again and the market roaring and that.

B (11:21)

I mean, yeah, like one, just one thing I go off though, like I remember like in our, in our sort of regular live company, right? I used to look at the Open AI bill and it was escalating and escalating. I'm like, this is going to be another big thing because we're always thinking about costs, right? Like what is, what is the cost of this stuff versus the benefit we're getting from our customers? And it was growing and growing and then suddenly it stopped and reduced and now it's not even a really a line item worth worrying about. Right? Like the use of OpenAI is just not a significant line item in a company where it kind of should be and was, you know, like I just, I just wonder how that is going to affect them in the long run when there's so many legitimate alternatives. And when you've got companies in, in markets where they're really looking to cost cut and you get models like deep seq v3.2 that they can host themselves and have a static known cost, right and just do a little bit of extra work to optimize for that model. How does OpenAI justify a company continuing to spend 10,000amonth, 20,000amonth with them?

A (12:30)

But I think this is the whole all feeling of the entire market, whether it's enterprise or it's consumer. Like if you think about the enterprise, well yeah, cost is a huge constraint having like basically zero inference cost or very low where you're just paying for the bare metal in your own cloud or your own environment. That would be important. And then the other challenge is like you said earlier with the principle and the tools, like people just pick the best tool for the job. Ultimately I think a lot of people would put everyone in this camp of like, oh, you know, the plebs just go to chat GPT. I think most people that have an interest in AI or at least hear about better models and things that it can do will just go to the right tool for the job naturally.

B (13:20)

Yeah.

A (13:20)

And I think, I don't think, I.

B (13:22)

Don'T think when it comes to AI tool usage, there are plebs as we refer to them. Is that an Australian thing? I don't know. But like, you know we're talking about like you know, regular joes, right? Like I just wonder if a lot of the people we refer to as plebs or regular joes are just children. They're just children using it and they're just using it because it's available. Because I think that anyone who is using AI seriously in their business is either trying to do it because they want to be better at their job and do a good job, or they're trying to champion in their organization. And people who are trying to do both those things need to use the best model because they don't want to be like the lawyer who gets done for citing fake references or the person who does a PowerPoint presentation that has some sort of crazy nonsense in it and they get found out. They want to use a model that makes it seem like it's them and therefore they're always going to gravitate to whatever the best model is. And ideally not some mainstream model like ChatGPT.

A (14:19)

This is, this is what I think, right where they've gone wrong is just not by focusing on that core of having the best image model, the best model for inference. They sort of control that narrative of like OpenAI is just the best at this. Don't even bother trying, guys. I mean Sam Altman said it originally in India, remember, Like don't even bother. You'll never catch us or beat us. And it turns out pretty much everyone has. And I think the challenge now is they've got to get back to their roots. So to me, this code Red, if it's going to be about anything, should be about building models that people actually are climbing over the walls to use because they're notably better. And I think that's the problem is a lot of these companies now for the sort of general day to day usage of models, they've all become very similar. And so you've got two parts. It's like do we fight in the consumer, do we figure out ads and do we have the best models so that and the best personalization and get deeper with the consumer and is that the direction we're going or are we going to turn to like Anthropic's model, which is say okay, all the token money is in coding use cases right now and generating code. So let's build the best agent coding agents and go after some sort of enterprise platform which I think is obviously anthropic strategy. But the numbers aren't great for them. Like in the enterprise they've gone from 48% market share 18 months ago to 24.4%.

A (15:43)

And the other thing is Gemini's monthly active users are growing pretty rapidly. But I think they do sort of cheat because they just put it everywhere. Like I don't know if they're including the AI search, like if you include.

B (15:54)

Every Google search you ever do, but.

A (15:56)

It already be at billions I think if they did that. So I'm not sure how they calculate that, but I would say Google strategy now is to Sort of have the best or better models than them over time and then just bleed them dry. Because if they can't charge a monthly subscription to a consumer and if they add ads, people flock to Gemini. Where do you go? Like, honestly, where do you go? Like, if you were them right now in this code red meeting, even though they still control, you know, a vast majority of the market, what are you, what are you doing?

B (16:33)

It's a tricky one. I don't really know. But like, all I know is right now I just look at it in my mind. It's like I use GPT 5.1 occasionally if I'm desperate, if I want another opinion, something like that. I'm never thinking, like, if OpenAI cut us off today hard, I probably wouldn't even fight to get it back, honestly.

A (16:51)

Remember early on when they did cut people off and they were so strict on limits.

B (16:55)

And I remember our first few podcasts on this channel, which was quite a while ago. Now our. My fear, the thing that we talked about a lot was what if they, what if they just take this away? What if they just make it for the elites and they take it away? It was a genuine fear. They were so much better than everyone else. That was a real fear. Now I don't care. You could take away, you could take away like most of the major players and there would still be legitimate alternatives like Grok, like Deep seek, like Kimmy K2, like there's, there's really good alternatives out there to. You don't need them. I like the Frontier models, I really do, but it's not a thing anymore. It's just not as big of a deal. And in fact, some of the lower tier models actually have advantages depending on what your use case is, because you can control them, you can run them in your private cloud, you can run them without a marginal cost on tokens, that kind of thing. It really is a different game now. And I just think that for them to actually wield that sword, they need to come out with something that's just so much better that all this other stuff goes away.

A (17:59)

Still push back on that because like the reality is we have access to every single model and pretty much every tool available. Right. In Sim theory.

B (18:07)

Yeah.

A (18:08)

What do you use right now? What do you use?

B (18:11)

I use Opus 4.5 and Gemini 3.0.

A (18:14)

What do you use for images?

B (18:16)

Nano Banana.

A (18:17)

Like this is what I mean. Like, people just want to use frontiers.

B (18:22)

Oh, sorry. I was like, it's like a cross examination. You got the truth out of me immediately. I'm hopeful.

A (18:29)

No, but I guess what I'm saying is like we keep banging on like people like. And I agree with you, if they took all those other things away today, sure, my fallback right now would be those other models. Right. But I still don't like, I just think people want the best tools. And this is why the enterprise strategy is a bit strange to me from anthropic and potentially OpenAI as well. Because you're essentially going to an enterprise and saying, hey X to our model, our one specific model that constantly, you know, is this like pull, pull. Like sometimes it's the best, but then you'll not have the best model for like maybe a year and then it'll be good again and they're asking for these big commitments and then as you said, there's all these like, oops, I checked this checkbox. Now they're trading on my data.

B (19:17)

Yeah.

A (19:17)

To me, the smartest companies, well, like this assumes the enterprise is dumb as well. And the enterprise is going, well, hang on, we can just access all these tools through APIs. So why, like, why what? Why would we cut a deal? We'll just go to the cloud provider. So then you think, okay, well OpenAI is screwed again because they're just going to go to like Azure or Amazon. And I guess the only defensive move then is like pull your models off those clouds. But since there's so many good models now, it doesn't really matter, so you just lose market share.

B (19:46)

I would argue people would struggle to tell. I think people are using this stuff every day would know. But if I transparent, if I was using Azure as the back end and I transparently switched your OpenAI model for an anthropic one, you, you, you would know seconds.

A (20:03)

Yeah, very. I still disagree. I think people do know. Like a good, a good real life example is my father in law is in his like mid-70s. He talks to me when he comes over about the different AI models. Like he has taken a strong interest in them. He runs like a sort of charitable organization now in his retirement and he uses the models to which is kind of bad, but replace employees or contractors. He used to have to do stuff like write things, create images, create video content, like all that sort of stuff, right. And he knows his model tunes and he knows when to switch models. And he also recognized when Gemini 2.5 Pro got dumb. Like he said to me, like, you know, it feels like it's gotten dumber. And so this idea that people, I think people's brains just naturally wire into the best tool for the job, if they're available, obviously.

B (21:01)

Yeah, I tend to agree. I think it's gone beyond being a tech thing. Like I always used to be proud that, you know, I'm a tech person. So I understand this stuff better, like in terms of understanding the model's performance. But I actually don't think that's true at all. I actually think that most people who use it seriously for something they understand do have a feel for the models that's just as good as mine or anyone's. Like, I think that you know your own work and you know when it's doing a good job and when it does a different job and that kind of thing. It's a universal thing. It's, it's. You're dealing with a form of intelligence. It's not like dealing with just a tech person where you need to understand the, like, how it works to be able to use it properly.

A (21:39)

And I think, but I think this is where it's all going right, is like the models, the models are just like sort of like having a chip on a computer in a way or several chips on a computer. It's probably not a great analogy. And you like or maybe like different apps on your computer and you just know which app to use for what purpose and what app yields the better result. And so everyone can use a computer and everyone can use apps. So I kind of think it's similar whether or not long term people want a model switch or not. I'm not sure. Like I during the week experimented around. I had a pretty good idea for a model router. Finally in SIM theory that I've been testing locally, that works quite fast. Like I've got the latency down where it's. You don't even notice. And it works pretty good. Like to the point it's, it's picking from my current choice, like Mike's picks of which model to use for what. So it's completely somewhat manual. Like there's a classifier on top and then the routing is just like controlled by myself. But I use it for a while I was like, this is good. But then I was like, no, I just want, I want the manual stick, control stick.

B (22:51)

As soon as you, as soon as you need to do something real. Right. Like I need to get results.

A (22:54)

Yeah, like I need to actually know I'm definitely using this or I want a second opinion or as soon as it unravels very quickly. So I'm not, I'm not sure I'd love to actually experiment with it and put an automatic model in ride and just see how many people use the automatic mode over switching. And I would argue or I would predict that most people will still end up picking their model, not using the auto pick. Maybe I'm wrong. If I'm wrong, drop it in the comments below.

B (23:24)

Let's try it and find.

A (23:26)

So yeah, I think the last few weeks of the year are going to be really interesting. Does OpenAI just blow us out of the water and they've got some, you know, rabbit in their hat. But generally when they have a rabbit in their hat there's so much hype and it's been planned for ages, but it's almost like they had nothing. Like they just assumed Gemini 3 would be bad, which I don't think was a good assumption. And to think they can now stay ahead with models is probably not true because I would argue Google, Google and Anthropic are just going to keep like, it's not like they're going to stop now and give open a high time to catch up.

B (24:04)

Yeah, I think they, they need to do more than just announce a new model. It has to actually be better. I think that's the key. It can't be just like benchmarks and you know, a nice event where they've got like a Christmas tree in the background and they bring out their nerds and, and talk. It's got to be a model that we all universally say, okay, this one is better. And it's obvious that if they did that now, it would be enough to retake the crown. Honestly, like it isn't, it isn't like Google has that much of a lead or Anthropic has that much of a better model with Opus that they couldn't come out with something that does that.

A (24:38)

Yeah. Interestingly too, if you look at open router.

A (24:44)

This is the market share and it's token, it is open router data. Right. So you got like this specific use cases for open router that people use it for. So it could be slightly misleading but if you look at it in the green on this chart you have XAI with, with Grok 4.1 fast. And we talked about that. Was it last show or the show before? This thing is so cheap it's brilliant at tool calling, so it's great at agentic use cases. It's just a great all round model and it's, it's so cheap it's essentially free. And so if you're building an application right now and you're doing tool calling and you want it fast and everything. Like it makes sense to me why it's number one for the use cases that people use open router for. Then you've got Google second anthropic and OpenAI is sitting in it at fourth on open router. So when given the choice to.

A (25:45)

Like consumed from different APIs. Even though Anthropics models are more expensive than Open AI and Google is about on par with OpenAI models, they're still fourth. And like, let's ignore XAI for now even though I don't think it should be ignored because it's showing what people will consume when the price is good and the model's great.

A (26:07)

So anyway, we'll, we'll see how it plays out. Do you think they're going to have some model that, that like blows us out of the. Like excites us?

B (26:16)

They have to. I mean, I think maybe they will because they have to. Like if they're going into code Red, the code red should not be let's release GPTS2 or you know, let's, let's come out with the 12 days of bullshit web UI that we're actually not that great at. They need to be like, the code red is. Let's come out with a model that is just so much better. And here's my criteria, here's what it has to do to be the best. Right. I'm coming up with this on the fly. However, it has to be at least a million. Context, it can't be less than a million or I just don't think it's in the game. Right. It has to be at least 64,000 tokens output. It has to be phenomenal at parallel tool calling. That has to be its wheelhouse. Absolutely amazing at it. It has to be good at long running, repeated processes. So if it's like, okay, I've done phase one, I'm going to come back phase two, do more research, whatever it is, I. E. Agentic loop kind of thing. I think you need to see a big improvement in vision. I think a big improvement in vision from any model right now would leapfrog them to number one because it's an area that really hasn't improved at all in the last year, if not two years. Like it's gotten a little bit better, but it really hasn't changed much. Do you agree with that? I. I don't know.

A (27:35)

I, I think GPTs have always been great at image or good enough.

B (27:41)

Yeah, but I mean, what percentage have they improved in the last year?

A (27:44)

Oh, nothing. It's flatlined completely. I, I think Gemini 3 is a, a tad better.

B (27:49)

So I think a big improvement in vision would be, would be absolutely massive.

A (27:53)

I mean this is your wish list and like that, like these are, I reckon these are panic tunes of GPT5. Like they don't, they can't, like maybe they're new models, maybe they just have all these models sitting around, but I don't know.

B (28:06)

And then my final speed, my final one is speed. If even if they have, they're already making a massive loss, I would at the API level just throw every piece of hardware I've got at it and just make it lightning fast. Like just so fast, it's unbelievable. And then the other one and again maybe not economically the best decision, but in terms of like mind share, cheaper, like if they could get it half the price, twice as fast a million context and better vision, they would absolutely crush it. And the model probably intelligence wise, only needs to be 6% faster, 10% faster, something like that. Sorry, more intelligent like, slightly more intelligent like, or at least on par with Opus, at least on par with Gemini and bigger context. And also, sorry, God, this wish list.

A (28:58)

Is so long am I. How long is this going to go?

B (29:00)

I'm going for a wish list. The other thing is some of the great API features that Anthropic has added. Automatic context management. So in other words, it's crushing down the token usage from old tool uses and things like that. So just everyone understands when you've got a big context building up from a long session, you've got all the results of all the tool uses in there that need to be there for continuity. So for example, if you've crawled a website, it's got the results of that. If you've, you know, gotten downloaded your emails, it's got all the bodies in there. What Anthropic can do is discard the ones that are no longer relevant, that have already been taken into account automatically. It's a great feature, especially when you're doing agentic looping and things like that. It's very nice.

B (29:46)

Better caching, that kind of thing. Like Open A already has transparent caching, but it isn't the best. If they could get some of those things going as well, suddenly to me they would be number one. And if I was in a listen.

A (29:59)

To everything you've described, what, what I hear is this Google and Anthropic are building tools that people actually want for real life use cases. And OpenAI's APIs have basically reached this weird point of of yeah. It's almost like they're not listening to what people want and they've just got their own sort of agenda with their models or they just can't figure it out because they're noticeably worse at tool calling and they like even the way they're tuned the output format. The feel is off in a lot of these models now when people get exposure to the other models anyway, we should move on because we could rant about this all day.

B (30:40)

And just to finish that, I'm obviously looking at it from a unique perspective because we provide a system that allows disparate companies with different needs to get the most out of the models. Right. So we're in a position where we need the best of everything and we're always going to be fighting whatever the best is. I think it would be different if you had bespoke AI use cases where you're just targeting one thing, you're trying to do the best. I think a lot of this stuff goes out the window then and you're really just looking at like a cost perspective or something like that. But if open I want to win people like us and win people who are building stuff on top of AI they need to do some of this stuff or it's just not going to happen. No one's even going to care.

A (31:22)

Here's the other problem for Open Air long term. Let's now introduce into the discussion deep seq version 3.2. So this is a new model, a new deep seq deepseek R1 being the thing that kind of like blew up the market originally because everyone was like there is no moat. Remember that phase.

B (31:42)

Yep.

A (31:44)

So deep seat 3.2 is out. Said to be a reasoning first model built for agents. Yada yada. Benchmark Max looks good on paper.

A (31:58)

Open source out there available, you can run it wherever you want. We, we've tried to tried it out. Unfortunately the provider we use in the US the token are just so slow. It's hard to get a true feel for the model yet. But my initial impression just from the few queries I did do on it is that it's pretty good. Like it. It seems very comparable to like even Opus I would say in a way.

B (32:26)

Absolutely. Like if you again if you did like a blind test on me with this, I probably wouldn't have noticed that like you say it is disappointing about the provider and we'll look for a new one because. Or just do it ourselves because the when it was. It's like when it's. When it's bad, it's good. And when it's good, it's sensational because it's very fast. Like when you get the request through, it's very fast. It was just that we are clearly part of a queue and the queue is not being picked up fast enough. However, I got it to using in sim theory research the Ashes Test, which starts today, the Australia versus England cricket match, and said, find out about, you know, injuries, player performance, historical things, stuff like that. It did about 20 or 30 tool calls, like, in terms of research. Then I said, make an infographic, a song and a. And an image of the match. And it made this incredible image of Steve Smith, the acting Australian captain, batting it correctly. Got that? It's a pink ball test. That is a day night test. They use a different colored ball for the game. It got an approximate score of where it thinks it'll be. Partway through the match, it wrote an amazing, albeit American country song about the game. And then the infographic it made is just unbelievably detailed. I mean, I know Nano Banana deserves a lot of the credit, but, like, look at that. It's got player images, it's got stats, it's got historical information, it's got comparison figures. It shows genuinely accurate statistics about this game. And I said to Mike earlier, like, can you imagine, like a newspaper or online magazine or anyone who would want to present this information to their customers in the past, how much that would have cost in terms of employee time, research, graphics designers, all that sort? Like, you couldn't have done it. It would have cost, what, hundreds of dollars, if not thousands of dollars to produce an image? Like, there's probably thousands, right? And if you look at the actual, like Nano Banana cost, It's like what, 15 cents or something to produce an image like this. And then you've got the deep SEQ inference, which is like, cheap as anything. Like, it's pretty amazing what it can do.

A (34:46)

Yeah. I think going back to the model, though, with deep seek, like.

A (34:54)

This is further erosion, though, of. Of OpenAI. Like, because if you said earlier about building for a specific use case, like, obviously we use all these tools and, and not necessarily optimizing for specific use cases because we want to make all those tools available. But when you go like, if you're building a product on top of AI, it's like, what are your constraints? Well, one of them is your OpenAI bill or your anthropic bill, or, you know, it's essentially like an electricity bill or a utility provider, and it really ruins the Economics of these companies where you a gross margin in a SaaS business of like 70% or 80%. And so every dollar a customer gives you, you keep 80% off. It only costs 20% to run the business. With AI businesses, it's sort of inversed in a lot of ways. And I think what can happen here or is happening is all of the startups that are building around AI are using Chinese models. Right. And deep seq v3.2, I think further erodes for both OpenAI and anthropic in a lot of ways because it's just, it's so much cheaper to use. Like even if you're consuming it through the services we use. Just to put some stats on this, it's what, $0.56 per million tokens? I mean I think Grok 4.1 fast is still cheaper.

A (36:20)

And probably better and faster at 1 million output A$68, so 10 times cheaper. This particular iteration of deep seq 3.2 has 163k context window, but I think it could be higher.

B (36:36)

Yeah, if you ran it yourself, it could be higher. And another important clarifying point is people here Deep SEQ Chinese model and think, okay, the Chinese get all your data. They don't. The version we use, for example, is hosted in the us. There's no training on the data, it's secure. Additionally, these are open source weight models. You can host it yourself in your own private cloud, local to your whole system. So you can get way, way better in terms of privacy and security with a model like this than you actually can with the major providers. So there's a lot of advantages to it rather than disadvantages. And as we've said, the trade off in terms of lack of intelligence or tool calling is minimal at most.

A (37:18)

Yeah. So if you're building an app that relies on AI, like you're building one of these like, you know, toy robots everyone's really into right now, like lamps I saw this week. And you just need to run inference. You would be far better to just run a model like this in your own private cloud. And that's going to be, but obviously.

B (37:38)

Is if you ran it like there's a minimum fixed cost, you're probably looking at between 500 and 1,000amonth to run it, right? Like in terms of the GPUs you'd need to rent or if you had to buy a GPU to do it, it'd be like 20, 30 grand, something like that. So there is a base like cost to it, but if you know the demands there Then you can easily justify it and then it's not a marginal cost, it's a fixed cost. So it's a lot better.

A (38:03)

And I think right now the argument early on and the advice I give to a lot of people is just use the best models to figure out the use case. Don't bother around setting up your own infra, just take the hit on the, on the cost until you've figured out the business model and then, you know, go like once you've got to get to that point in the business where you're like, this is working, this is a long term thing, then I can optimize around a model like Deep Seek in my own cloud and have fixed costs around this and get my economics looking right. And the question is, is, are we there yet? And I'm not, not entirely sure. I think for some businesses maybe.

A (38:41)

You would be closer to that right now where you're like, there's just these streamlined use cases in the business where it's far easier to have fixed costs around this than just pay for continuous inference from various providers. But I think in terms of raw intelligence in an enterprise or a business today, you do still want to be working with the latest models. You don't necessarily want to be locked into like some, you know, stale model that takes 12 months to provision.

B (39:08)

Yeah, and I think also there's, there's expertise involved. Like for example, you couldn't live with 160k context window that's really small by today's standard. Like I think the minimum is 200 you get now with something like Opus 4.5. And honestly that's weak. Like it should be higher than that. So there are downsides, definitely. But like you say, with specific known use cases, a model like this is hard to beat.

A (39:34)

So interestingly, I think we were talking about this last week or the week before this idea of AI first people, like if you were hiring now, this is what you would be looking for is these skills where people understand how to use AI models and tools in the workplace to be more efficient or just get a lot more done and be a lot more productive and automate processes and things like that. And there was this article that stood up a lot of, a lot of commentary in Business Insider this week. Matter is about to start grading workers on their AI skills. Meta will assess employees performance by AI driven impact starting in 2026. The company is shifting towards an AI native culture and incentivizing AI adoption through rewards. Better is also rolling out an AI tool to assist employees in writing and performance reviews.

B (40:28)

Do you think maybe Zuckerberg listens to our podcast? And if you do, Zuckerberg, please buy our merch. God knows we need it. Our store is may or may not be a scam, but please just buy some of it for your staff. Whatever.

A (40:41)

I'm amazed that's still operating. But it, it really got me thinking, right? Like, like there was a lot of commentary around, like, this is stupid. People still need hard skills. Like, you don't want a bunch of slop out there, like slop code and you call co code base slop sales proposals going out. And there's this narrative, this like any AI narrative I think around this, where people are like, do we even need this? Is this the right way to lead in this era? Like, should we be forcing people to use these tools or not? And I think even in our own direct experience, there is something now, or at least in 2025, where it feels like workers or employees that are not adopting AI are starting to actually fall behind. Like they're starting to seem like dinosaurs that haven't necessarily embraced these tools. And I think it's becoming more and more obvious. And then you're also in this strange era in education where if you look at schooling or university, university in particular, the financial burden people face when they go and do a degree now, only to come out now and not even get a job. And all of a sudden the AI skills are vastly more important.

A (41:59)

Is leading to a lot of people not even going to university. And the value propositions almost question there as well with if I want to go and work at Meta and all they do value is these AI skills. Why would I bother going and doing a course that's already a sort of dinosaur course? Getting myself in a horrendous debt and you know, and then not having the AI fluency. I need to go and get a job in these companies. So I think this disruption, it's still very early with it, but you can totally see it coming. And these tech companies tend to be the early adopters of the trend, like the fact that they want AI fluent workers.

B (42:40)

Yeah. And I think you asked a question the other day about a task and you're like, well, how long should that take a modern front end developer to complete? And we look at what we would have thought, say three years ago and look at now and you're like, if it's taking this person more than an hour, they don't know how to use AI basically because there's no reason to do the grunt work associated with that job anymore. Like there's literally, if you're using say React, which Facebook invented, right, and you have a well known components library that's fully documented, which they do, right. And I ask you to build a new control that does something. Why in the world should it take you more than an hour to fully build it and fully test it in place when you've got the access to these models? Even, even Llama could do this stuff? Like, you know, there's, there's absolutely no reason why a developer in that role should be doing it in a traditional way. And if they don't know how to competently use AI. If I was Facebook, I'd fire them immediately or retrain them obviously.

A (43:44)

But even, okay, but even outside, like our software engineering or developer bubble, this applies to almost everything. Like if you're a lawyer and you're writing, like trying to come up with an angle in a case, like going back and forth with the AI on this is going to be much faster than just sitting there manually typing things out. You still need agency and you still need some core, like hard skills there. But even data analysis skills, like now you don't really need a data analyst. You can get it to write SQL queries, produce charts, produce insights, at least raise these things. And sure, like, it might be wrong or inaccurate sometimes, but so are data analysts in real life if you've ever actually dealt with one.

B (44:28)

Well, also think about like someone producing like replying to RFP documents or making RFP documents or producing slide decks or, or writing, you know, business proposals or purchase orders. Like all of these things should basically be seen as like rough clay where you work with the AI to refine it into the finished product using your expertise. They should not be documents that you're ever creating from scratch anymore. And therefore anyone who can't use the AI skills is by definition less efficient than someone who can.

A (45:01)

Yeah, and, and I think this, this is the thing, it's like AI fluency or not. And I think Zuckerberg, despite, you know, being seen as like evil sort of alien spawn villain, he, he is probably right. Like this is the way many workforces are going to go and be expected to implement, whether it's automation or just work with AI to be, you know, way more productive. And I think that the challenge is a lot of people in the workplace still see AI as a threat to them, not a tool that will actually get them promoted or like empower them and make things better. And then on the other side you've got people and I know because many of them have reached out to us, whether it be their parents or them directly and said.

A (45:51)

Should I even go get a degree anymore? Should I even go and study and will what I learned be out of date by the time I get out? And. And I think interestingly, that concept requires a lot of self motivation, right? Like if you're going to sit with chat, GBT or, or Gemini or whatever and teach yourself and learn stuff, like you still need structure and you still need to understand what it is you're learning in order to be valuable in a workplace or to get a job or, you know, like I think this idea that you'll just magically teach Yourself with a ChatGPT subscription is a bit bonkers.

B (46:29)

No, because it requires taste and it requires understanding where it plays a role and where it doesn't. Like you really need to know how and when to use it and how to respond to what it gives you in order to be effective. Like, it's a totally different way of working. It isn't that the future of work is just copying and pasting into a chat window. It's like how and when do I leverage this to do the pieces of my work that it is better at and faster at than I am? And I think that that is about context building it. Soon it's going to be about coordinating agents and there's a whole big technique that is going to be needed and the people who know it are going to do better regardless of where they learn that training.

A (47:09)

And this is the thing, right? Like there's a, there's another statement Zuckerberg made quite a while ago. It says, mark Zuckerberg says college isn't preparing students for today's job market. This and the debt burden will create a reckoning for higher education. And I mean, there's a lot of data that supports this as well. Where outside the sort of like, you know, doctor, lawyer, where you need the professional qualifications and that, you know, like you can't, you're just not going to be able to pick that up. There's so many jobs, or I would say the vast majority of jobs where you would traditionally go get your degree, maybe you go back and get an mba and then as a result you were recognized, you were promoted. Like there was a lot of value placed on that, so certificate and that training and the importance of it. And now it's getting to a point where a lot of this stuff is just overnight out of date. Like certain courses are just not landing people jobs in this era. And really what needs to happen is they all need to be rewritten and rebuilt for this AI fluency.

B (48:13)

That's right. The AI, the universities need an AI component to basically all of their degrees. Like they need compulsory subjects in them. That is like, this is an AI certified degree that has like you say, AI fluency attached to it.

A (48:28)

And interestingly.

A (48:31)

Spoiler alert.

A (48:34)

With SIM theory very early on in Australia, a university who I think gets this. And this is definitely not like some paid advertisement for them at all.

B (48:44)

No, it is. Mike's lying. It's a paid advertisement.

A (48:46)

A second sponsor it was carrots and now it's universal.

B (48:49)

I'll censor it. Hang on. Yeah, yeah, so if you're into carrots or universities, you want Bolthouse Fresh or une.

A (48:56)

Yeah. So the University of New England here in Australia reached out to us pretty early on and have, you know, also rightfully identified this. And I think they, they have taken this as an existential crisis. Like how do we get AI fluency written into everything that we do and how do we invent new degrees that are actually aligned to what employers want now and that AI fluency people expect? And I, I think there's a lot to this. Like it's going to become like, you know, when people, I guess the, when the Internet was a thing, like people like, I don't do the Internet. Like I just, you know, there was a subgroup of people and the same group of people are like that with AI today. With like, I don't do AI and.

B (49:42)

So like florists, Florists, like no one will ever order flowers on the Internet. Are you crazy? And then they just got crushed by like Roses Only and Interflora and all those companies.

A (49:53)

Yeah, now they're just like distributors for those websites. But yeah, so I'm, I, like, I think the backlash to this is a bit ridiculous. Like people, people seem to see through this lens that all AI is slop and they just want it to go away. And there's this backlash anytime someone makes a statement like this.

A (50:15)

And now, I don't know, I think increasingly this is just becoming such an important skill.

B (50:21)

I've got a funny story about this. So at that same speech I mentioned earlier at my kids school, the big principal was talking about AI and he's like, oh, I think in this world of AI, we need to learn what it means to be human. We're all humans and there's things humans can do that AIs can't. And then he's like, for example, an AI can't love you. It can't feel love. And then I turn to my wife and said, patricia loves me.

B (50:47)

She's so angry about it. She doesn't like it. But yeah, it was so funny how this school, for example, has. Sorry, we'll listen.

B (51:00)

This school, for example, has recognized the pervasiveness of AI like, it's in everything. They can't ignore it. Because I was really annoyed at first. I'm like, this, this assembly is about the student and their achievements for the year. It shouldn't be about, you know, the, these. These existential crises you face as a school. But then I'm like, but it's so. It's so pervasive. Like, it's in everything they do now. Like, students are using it casually. The students are using it for their work. The teachers are using it to help the kids. Like, it's, it's everywhere. And it's so capable that they have to acknowledge it. They have to recognize it in some way. And I think you're right. I think something that UNE did so well is recognize this, what, two years ago. They're like, this is going to be a big deal for us. We need to do something about it. And I think that this is going to happen in more and more educational institutions where they're like, we can't ignore this. This isn't something that we can just say is slop and will go away. We absolutely must acknowledge it, talk about it, and not just come up with a solution for it, but win the hearts and minds of the students, in this case, the parents or, you know, the, the. The stakeholders in the situation and say, well, this is our approach and this is how we're going to leverage it. Because I feel like if you don't do that, people are going to go, if they have the mobility to do so, to, to people who do acknowledge that and, and have a strategy for it.

A (52:21)

Yeah, I, I honestly think that there's this transition period too, where these, these AI native kids will come through, and it'll just be whether they're being trained in a specific role of, like, how do I then apply these AI skills in a way that's relevant to this career path I've chosen. And then there's also going to be these, this poor, like, in the middle generation where it's like, well, it's kind of partially here, but it's not. And some companies are adopting it. Some companies are saying, oh, it's completely banned or whatever. But the value proposition to this generation right now of, you know, potentially spending all this money on a degree where they're not learning these AI skills, going into debt as a result of it. And then right now, with the way the market is not getting employed either, it's. It, it's ripe for disruption. And I think there's definitely job roles.

B (53:18)

That I can see just being completely either eliminated or just like one powerhouse is going to be doing that job for 10 companies.

A (53:26)

Yeah. And to me, there's opportunity here, which is. Yeah, like for the AI fluency market. Like that. To me, that's what education needs to be about, is how do you get people up to speed and also how do you teach them to use the tools and use their own judgment and question the outputs and give the model some agency because it's certainly not capable, at least today, of just magically doing the job. You, you really.

B (53:54)

Yeah. Even like building a tool stack. Like, what's your tool stack? Like the combination of tools you use to get the job done. Because really, MCP and tool calling is such a big part of it. Like, even just being able to fashion your own tools to make your job more effective is going to be a useful skill.

A (54:13)

Yeah.

A (54:17)

I think it's so interesting to see it and live in this time where you're. You've got kids in the school seeing AI just infuse itself everywhere, whether they like it or not, and then seeing how these people actually react to it. And funnily enough, I think schools have probably the worst internal tooling I've ever seen with AI today, at least they do here in Australia. So it's somewhat troubling because then it means students are just going everywhere.

B (54:42)

My son's school just banned SIM theory and I was like, really proud of that because I'm like, finally we're in. They banned us. Yeah, because he was, he was handing it out to all these students to buy bypass their chat GPT ban. But then we got banned, so I'm like, proud, but also disappointed.

A (54:58)

So they banned chat GBT at the school even though they preached about it?

B (55:02)

Yeah, that was from the start and now we're on the list.

A (55:06)

Isn't that like banning Google? Like, because Google's got all the AI in it. Maybe this is why their traffic's down.

B (55:12)

Yeah, it could be. But yeah, like they, I guess the IT administrators ban it and anytime I try to help the kids bypass it, I get in trouble. So I, I don't now, but yeah, I was pretty proud when I heard that's a justification.

B (55:26)

They didn't even say anything. I just, I just hear from the kids I don't know.

A (55:30)

Yeah, it seems ridiculous.

B (55:32)

It's sort of like, yeah, it's ignorant. It's straight ignorance. Because I don't think that that is really the way to handle it. You got to remember now, a lot of kids in schools get their own laptop given to them, at least in Australia. And those laptops adhere to the IT policies of the school. So it isn't for them. It isn't just banned at school, it's also banned at home, effectively in terms of doing their work. It's weird because, like they use tools like Canva, for example, which has AI stuff in it. So it's like it's not AI generally being banned. It's almost like, well, this specific AI is banned. So it's an odd kind of censorship in a way, in terms of the kids and how they use the technology.

A (56:12)

Maybe it's a conspiracy. Like big Australian tech is like, you got to ban it because we want Canva.

B (56:18)

That's right. We've got to do something to finally get to this ipo. Guys, come on. We've waited so long.

A (56:25)

So, Chris, we do a lot of songs on the show, right? And.

B (56:29)

Oh, is this a surprise. I didn't know there was a song today.

A (56:32)

No, there's no song today. Oh, sorry. But we do a lot of songs on the show and I've always thought I had a certain knack for the. The AI songs.

B (56:40)

You did. I mean, you made what in our musical, which is wildly unpopular, I must admit, but one of my favorite things, you made almost all the songs. I think I only got one in there and it was probably the least popular.

A (56:53)

But anyway, you've. You've finally beaten me in the Spotify charts. It's now number one.

B (57:01)

Fatal Patricia.

B (57:07)

What would it take to get it played on the radio? Like, could we do it somehow? Like if we rally as a community?

A (57:12)

Someone in my. Our audience must have some sort of connection at some rate. But why radio? I mean, no one listens to radio anymore anyway.

B (57:22)

It's just validation. Just like for my self esteem. I'd love it. Yeah, but imagine like some, like American radio said, there's so many, surely one's going to play.

A (57:31)

It would be good to have some sort of rural US or Canadian like radio station. I think that should be our goal. If anyone in the audience has any connection to any radio station, we would, we would like this played. I don't know why though, because no one would like, you know, it's probably got more distribution on, on Spotify with our 300 monthly listeners.

B (57:53)

I'll take it. That's good.

A (57:55)

All right, I have one more segment for you. It's called WTF of the Week or.

A (58:05)

And it's this video. And I know for people listening, this is going to be kind of annoying, but I'll explain it as best I can. So Tesla Optimus, which has its own X handle.

A (58:18)

Poster, just set a new PR in the lab. And I thought this was AI and totally fake. I mean, it might be, but I doubt it. I mean, it is AI but not the video. Look at it run like this.

B (58:31)

It looks unfit.

A (58:33)

I don't know, it looks beautiful. Like it's the first. It's just so fluid and like it looks like a human now.

B (58:39)

Yeah, it looks like a 50 year old woman running.

A (58:42)

Okay. I think, I think it looks better than that. It's probably like me running, like, I'll give it to you. It's a little unfit.

B (58:49)

There'd be 50 year old women fitter than you, I'd imagine.

A (58:52)

Yeah, wave it up.

B (58:53)

Definitely.

A (58:53)

It's a little bit unfit. I can't even walk upstairs, but it looks it. I don't know, it gave me shivers. You got to watch it. I'll link it below in the description for those that haven't seen it, but.

B (59:04)

It'S the first one of these Android videos I've seen where someone doesn't like kick it in the legs or like try to trip it over or something.

A (59:11)

All these videos of people fighting it.

B (59:13)

Throwing bricks at it. Look at this. Best AI in the world can survive a brick attack.

A (59:18)

Yeah. All right. Any final thoughts? A bit of a nothing week. A code red. The deep seat model is interesting. I think it's a good model.

B (59:29)

Yeah, it's. It's all a reputation thing. It's like, can you motivate yourself to stick with Deep Sea for more than a day? You're just not gonna do it. That's the problem with it. I don't know what it is. It's a. It's a vibe thing. Like the, the whole thing of AI is it's a vibe art form. And no one is going to be like, I'm going to vibe with something that might be the worst one out there. Like, you just, you're just not giving it a chance. Maybe that's the problem.

A (59:53)

Yeah, you're only taking a chance on the big brand.

B (59:55)

Please, please give it a chance.

A (59:59)

All right, we will see you next week. Maybe with some new open AI models. Maybe not.