Claude Mythos Changes Everything

What if instead of giving AI a task, you could hand it an entire project? Not just write me an article, but run my blog. Well, Anthropics' new model, Mythos, suggests that reality is just months away, not years. It's scored 93.9% on coding benchmarks where every other model to date has been stuck at around 80%. You can't use it yet though, but when distilled versions hate your subscription, the way you work with AI changes completely. I'm Mark Webster, I'm joined by Gail Breton, co-founder, authority hacker, and today you'll learn what this means for real-world businesses using AI. So Gail, Mythos is here, or rather it's been announced, not specifically released. Almost here. Tell us about what this actually means in terms of what we're going to be able to do with AI this year. It's kind of hard. What I know is that if this was released today, if everyone had access to it, and the model did what you asked it without resisting, which we don't know how good it is at resisting, and Anthropics doesn't seem to trust themselves in having the model not do bad things, it could break most of the internet. This model is so good at coding. How specifically? Well, this model is so good at coding that it's able to find and exploit vulnerabilities in any software you use. So this model, they were able to, for example, find vulnerabilities in all existing browsers, in all existing operating systems, and that could be used to hack your computer, steal your data, etc. Let me actually show you one graph that I think is interesting. That's this one. And so that is basically the model's abilities to hack into Firefox. And so you can see that Cloud Sonnet is at 4.4%, or PUS that you're using every day is at 15.2%, a success rate. And the new mythos is that 84% success rate of hacking into Firefox. That means it almost always is able to find a vulnerability when it tries. So nothing is secure, theoretically. Exactly. Could be, let's say, probably the biggest banks would be okay, but let's say a bank in Ecuador, or some kind of third-party country that's been coded by local developers, or whatever. Like, things could happen, right? And it's like, imagine you're hacking to that, but that's interconnecting to many other banks and country operations that it shouldn't. It could break the internet, basically, more than Kim Kardashian ever did. And so like, Entropic decided not to release it to the public, and decided to release it to only a handful of companies that maintain software that essentially makes the internet run. So Apple, so Oracle, so Amazon, these kind of companies. Are they doing this? Because the cynic in me is like, well, they're going to IPO later this year. They're just kind of making this amazing AI hype around this. But it does seem like there's some legitimacy to their kind of staged approach here. So they're releasing it to all these big companies so that they can secure everything before some version of this gets released. Like, in the release video, they had a guy who has been kind of like a software security specialist for many years. He looked pretty old, older than us. And he was like, in the last two weeks, I found more bugs than the entire rest of my carrier, since I have access to this, basically. So yes, that's pretty much what it means. And what they're hoping is that these companies will have enough time to fix at least the biggest issues so that the internet doesn't fall apart when inevitably an open source model with this level of capabilities one day arrives, and you can train it at being a hacker, and it can try to break things for you, which will happen one day. Now that we know it's possible. Now that it has happened once, you know, it's kind of like world records, right? It's like they happen once, then everyone just starts matching them world record and beating it. And so that's kind of like the world record of AI that just happened. And that means that right now it's only in the hands of entry, maybe open AI, we don't know. But eventually it will be like in Chinese labs and everywhere and with much less security around it. So are we going to be able to use this many times soon? No, that's the answer. I don't think they will. First of all, it's very expensive. So they released the pricing for enterprise, right? Pricing for enterprise is $25 per million token in and $125 per million token out. If I remember correctly, basically it's five times the price of opus. It's very expensive. And so, you know, your max subscription would not begin to cover it. Second of all, they don't have the compute. Cloud is already going down regularly due to lack of compute these days. And you have API issues and all these things, they have to reduce the limits. That's more a popularity issue. So like lots of people are using it right now. But there's just not enough compute in the world right now. If you were giving a model that is that big, because the pricing really tells you how big the model is and how much RAM it needs to run, et cetera, VRAM and so. And so like there's just not enough compute in the world right now in the data centers. If everyone using a coding agent was using a model of this size, it just doesn't exist physically. And yeah, I don't think they want to. I think they kind of are afraid. Like we're hitting the point where AI was fun and now it's time to get a little bit scary because it can start doing some harm and actually do serious things basically. So they released a model card. I think they're calling it. It's like a 100 page PDF with like 250 lots of research insights and like what it can do. Again, to go back to the IPO, which they are trying to do that this year, that is the plan as I understand. Especially if they have this level of breakthroughs, like is very likely. If they're like, hey, we have this amazing AI that can do all the scary shit, but we're not going to give it away because it's too powerful. Like that is a very appealing thing. Like, oh, I want to put my money there because these guys have figured this shit out. Do we know or will we ever know that this is not like just some overhyped marketing play? Well, there are people who have benefited from this model that started speaking out, right? So for example, FFM peg, it's like a library that people use in a lot of sass, you know, that allows you to like convert video and do all that stuff, et cetera. That model found a bunch of vulnerabilities. And, you know, they're like basically like see when entropic tweeted about this new project of like helping companies fix their issues, like that's one of the companies that benefited. And the public decided they received stuff. It wasn't bullshit. And, you know, people even commented like, why aren't you mad about AI sloppy pull requests? Pull requests is when you ask a change for the code, right? Or when AI makes a change. And FFM peg replied, because the patches appear to be written by humans. So like basically it's like that has reached a level of our human devs and it's just that good. And Tarik is also an employee from Entropic as well here confirming this is not bullshit, basically. So, yeah. To step back to like the initial question, though, because I think when Opus 4.5, 4.6 came along recently, it kind of set this new barrier as all new models do for like what's possible. And I would say like people have kind of gotten used to like, oh yeah, cool. AI can now do all these cool things that we didn't think it was good enough to do before. We have conversations with people and it's like, oh yeah, you know, AI can do images. It can write content for you. And like, oh yeah, but it's not very good or it's not as good. It hallucinates me. It's like, no, no, no, it's good now. So what is the no, no, no, no, no, no, it's good now in six months time or at the end of the year with the pace of development that this has shown like, what are we going to be able to do at the end of this year that we don't think AI is good at now or we don't use it. For now, we don't have access to you and I. Yeah, we don't have access to it. Right. So speculative. We haven't tested it in our hands, et cetera. But let's still talk about some of the numbers they gave us and the benchmarks. Right. So you can see like, this is pretty easy. Like opus 4.6, you know, people may, you know, complain that it's gotten a bit worse recently. I might agree sometimes. The point is it's still a very good AI model and you can do a lot of things with it, including building SaaS, et cetera. Like people are doing it today. Right. It's happening. And then these are the agent coding benchmarks, basically. And so the little grayed out bars are opus and the white bars are Mitos, the new one. And SWEB bench pro is arguably, I mean, these two and the terminal bench are probably like the two biggest ones together with SWEB bench verified. Just to give you an idea, the models have for on SWEB bench verify, for example, they've been stuck around 80% for like a while. Like, you know, opus 4.5 was at like 80.5%. And then opus 4.6 was like 80.8%. Like we've been progressing like marginal points. And a lot of people have made the point that actually AI may never go further and the benchmark is already kind of saturated and Mitos came in at a 93.9%. So it's like, you know, that's as big of a jump between opus 4.6 and Mitos as there was between sonnet 4 or maybe the current haiku. Haiku is about the level of sonnet 4 and opus 4.6. Like that's kind of the level of difference of how good they are at coding. Put that into reality for a second. Okay, we're going to tell you. So you remember how we talked about how I built Brent Snap? Maybe like, you know, it's almost a year ago now, maybe. So for those who don't know, this was a website, a tool that we've built that you could put it, describe what your business did, and it would find you an available.com domain with the social handles and trademarks and open source stuff like that. The steps was like an AI model with brainstorms or domain names, then you would check against an API and I would just kind of like run the front end display the domains that are available and you can click and buy them. And you'd have a button to check the social handles basically that would call another API, check if the model that you've selected has the available social handles for an API and then check that. I built that with sonnet 4, which is, you know, the model that I'm telling you has as big of a jump from that to opus 4.6. It took me more than two weeks to build, right? Then when opus 4.5 came out, which is a little bit worse than 4.6, but not much, I one-shotted it. Like I one-shotted Brent Snap basically. With one single simple prompt. One prompt. And it made it look like. One plan plus execution, you know. So maybe two prompts you could say, but the point is like I one-shotted it, it worked. And it found better APIs. It made it run cheaper than when I built it with 4.6 actually. So it's sonnet 4, sorry. So that's the level of jump that I have experience going from sonnet 4 to opus 4.6. Now imagine a similar level of jump from opus 4.6 we have today to the next thing. So what you can imagine is much more complex system. Like now you're making like dashboards and things like that, but potentially you could, you know, spend a day in plan mode and one-shot a much more complex SAS, for example. Or like build, like, you know, now we build kind of like, we one-shot kind of like astral websites that are pretty simple, etc. Imagine like a website with like a complex database behind the scene and like interactive tools and, you know, payments built in like authentication, login system, account system, etc. One shot like we make HTML websites one shot today, for example. Like that's probably what you can expect from these models in terms of like being an entrepreneur. So like something that today we probably even with opus like building something with like an account system, a payment system, like a full e-commerce system, a database in the background, connected all to your accounting system, etc. Probably takes you like a week or two to build with opus today. That's similar to what it took me to build brand snap with sonnet for. Probably that's a one shot with a model like that. Again, we're speculating, but like that's kind of the level of what you can expect in practice. And again, much less bugs, much less issues, a table to fix itself, that kind of stuff. And again, like when you have this level of jobs, there's kind of like new emergent capabilities that come with this model that are hard to predict. And so probably there are some there that probably I can predict, but that's a story that helps. I'm going to push you on this. Let's try and predict it. So like if you could say, you know, of the work which we do at the moment in our business, or take an agency business, if you want to like simplify it, what types of things are people not using it now that they will be or they could be potentially at the end of the year. You know how again, we're speculating, I'm going to repeat that a lot, but you know how I have my notion system where I can give a task to our cloud code now and we'll just go and do that task for me. Yeah. And it's pretty good at it, right? I'm expecting with a model that, you know, again, like if you look at the reasoning benchmarks, it's also significantly smarter. Like humanity's last exam is a hard benchmark and you're talking like almost 50% more performance here. I expect to be able to give it a project rather than a task. So what if I can, if I'm able to, instead of saying go write me an article, if I'm like just, you run my blog now, you use all the tools available, you come up with a strategy, you execute it, you learn from it, and you adjust as you go, for example. Like I suspect if we are hitting these levels of intelligence, this may be possible. I won't necessarily run your business, but it might run projects in your business and handle subagents creating the tasks, longer horizon stuff, potentially learning from its mistakes, that kind of stuff. So yeah, potentially by the end of this year, which is when I expect models will get these level of capabilities as consumers or prosumers, then potentially instead of giving tasks to a, I'll give it projects. That is a pretty fundamental change to all work, all knowledge work. And that's going to be, I would imagine if that comes through very disruptive to hundreds of millions. It might not be perfect at this, but it might be the early days where you can start giving projects to AI. Like I'm not saying it's going to nail it and be AGI and do it perfectly, but if it can match a two years experience employee in running a project, that's already changing a lot of things. Right? And if it can start taking, like thinking long term, writing down stuff, reading its task, etc. Like that's going to change a lot of things. So yeah, that's the level of jump we're talking about with this model. So I hope with these two examples, with the idea of one-shotting these things and project running, that gives you an idea of what's potentially coming our way within three to six months. And you said we're not necessarily going to get our hands on this model specifically. What will the reality look like? Are they going to distill it into opus 4.7 or 5.0 or something? So, okay, opus 4.7 and summit 4.8 were already in the cloud code leaked files. Like it was already mentioned. Which we covered last week on the podcast. So you can check that out. So they probably exist, but I don't think they're based on that because from their research paper, this model has been available internally since February 24th, which is not that long. Like, you know, then they need to red tape it, they need to distill it. And so what I think is like, this is probably a distilled version of this is probably going to be the 0.5, the 5.5 versions of their models. So probably they're like quite close to the five version of summit and opus, which a lot of people are predicting are potentially launching this month. After that, we have the like at the end of the year, the same way we got opus 4.5 towards October, November last year, we'll get a 0.5 version towards the end of the year. And my guess is there will be a distilled version of potentially an improved version of this model then. And what again, Anthropic has said that there's an upcoming version of opus that's coming very soon with new security guardrails that they're going to use to test to see if they can release a model this powerful basically. So we know there's an opus model coming very soon. We know it's coming with new guardrails and they're happy enough to release it. Well, they're not happy enough to release this. So my guess is it's not based on the same architecture or whatever. And this is kind of like the gen after that. You know, you talked about the cost of Mythos for those who are able to access it being quite a step up. What does that mean for, you know, the future of opus? Like, are we still going to be able to like use these kind of future models or future versions of opus on our $100 or $200 subscription? Or do you think these costs are going to go up? I wanted to show you this graph to answer this question actually, which is essentially a benchmark that they've run on opus 4.5, sonnet 4.6, opus 4.6 and Mythos. And that's like a browser benchmark. And, you know, the model is given a task and they calculate, you know, the success rate and how many tokens are used to do that, right? And if we look at even opus 4.6 here, to achieve the task at 83.7% success rate, it uses a little bit over a million tokens. So like let's say 1.1, 1.2 million tokens. Mythos to achieve 86.9% or even 84.9%, which is still higher, uses around 170,000 tokens. So provided this efficiency translates, even though the model costs five times more, it might actually be cheaper to run the task. Again, that's one benchmark. That's three times more efficient. Well, that's one task on one benchmark. Not all tasks may translate to that. And usually what we tend to do is when we get better models, we give them more complicated tasks and the token usage kind of ramps up anyway. So like, you know, if you want to achieve what you achieve with opus today, it may actually not be very different in price. But what you're probably going to be likely to do is, you know, when I talked about giving you projects, for example, if you do that, it will probably just ramp up to the millions of tokens usage anyway. And then you'll get more, but you might need to pay more. So do you think small businesses need to kind of make some fiscal headroom here for, you know, AI? Well, you know, employee costs, essentially. Because we think about this as like a software tool. But then, you know, you have no issues about paying employees thousands of dollars per month to manage projects and things like that. Well, you know, are we moving to a world where we're just going to be having, you know, like a 50k a month Claude bill and just going to do the work of like a whole team? Yeah. And also the distillation will do a big job, right? So this is like a big unoptimized model. But imagine that they distill it and they managed to make it like four times cheaper. So it's a little bit more than a piece right now, but not a lot. But also they maintain, let's say 80% of the intelligence. And then let's say it's like, it's slightly less token efficient, but like, you know, you move your graph a little bit to the right, basically, and a little bit down. It will still be like a very, very good model compared to today's office, right? It would be very cheap. It would still be much more token efficient. And then you achieve like something that works well. So what I think is just the usage is going to increase. But what I think as well is like, this is not very far away, right? This model is now rolled out to Apple and other things. It exists. People are using it as we speak. Yeah. It's being used right now by Apple to patch iOS and stuff like that, etc. So it's like, very likely the tools and surfaces that we use today, they will still be the same. Like you probably use a model that is a distilled version of this inside the cloud code that you're using today plus minus some features. And so what that means is like, you need to kind of start building your systems so that you can actually leverage these models when they come in and, you know, like the notion system I was talking about or like managing your website with AI, that kind of stuff. Can you imagine when you paid human developers to maintain your website and then there's infinite AI that can look for vulnerabilities and then you have to pay human wages to actually fix them? Like, you know what I mean? You're going to be forced into that world for anything digital because the cost efficiency of having humans do things is just not going to work anymore. Just think about, so, you know, we obviously come from an SEO background, you know, had an agency for five years, ran an SEO course for 10 years. Like, we know a thing or two about that. We've run many websites, many blogs with teams of people in the past. And I sort of see this not as a choice, but like as a... Yeah, you have to. You will not be able to run a blog successfully unless it's fully automated with AI quite soon. But it's like you're talking about a blog, but it's a lot of aspects of business in general. Like once your competition has AI running the project and then you're paying someone 10K a month to do the same thing, it's like, what the fuck? Like, yeah, you can't compete. So like they win and then everyone else who's automating wins. But then that's my point is what's the next level up from that? Because you say, okay, well, they run a project, but like instead of run my blog, just like do my marketing, grow my business. No. Where's the button to like make money? Like mind boggling in a way. It's like, well, if it can do that, like... That's what I told you yesterday night. Does that negate every business owner, every knowledge worker, every agency, like, you know, it's unthroughing. It's just going to be running the world with all this. I don't think so. I still think like building the systems, the harnesses, the structure, the strategy will still be quite important. And I don't think this model is like, you know, there was a survey that you checked, right? Where it's like they asked AI researchers, like, is he's going to replace you? And only one intent said that. So it's like, I think what's happening with these models is they're becoming super humans at some tasks, like coding. Yeah. But they are also like well below a human or like, you know, they may be like the level of a 14 years old at some other tasks. And then it's like, you kind of need to come in and fill in that role. And so like, I do think you become that system builder, you become that architect of, you know, throwing some agents here and there and doing things for you and making up for the model's weaknesses while leveraging their super human abilities at other things. And I think this world might stay that way for a while. And again, I always take the example of self driving because we had the first demo a long time ago, 2016. And again, in many aspects, self driving cars are super human, like their reaction time, et cetera. They beat us. But actually we still don't trust them to drive for us most of the time. Now we're getting close, but it's been 10 years, right? And so probably there will be that longer than people expect period where the models will still be kind of wicked some stuff and the best combo, like again, business, it's like, I win, you lose quite often type setup. And so you need to go for the optimal setup. And the optimal setup will still probably be for a while a human operating machines without like not just the machines operating everything, which means you'll rely heavily on these AI companies that provide these models. But you'll probably need that it's a new level of vision and system building things. And that's essentially what we're trying to do now is like learning to build all the systems with code code in notion and projects and all that basically. So it's just new entrepreneurship if you think about it. This reminds me of a blog post from Wait But Why. It was a very popular blog in like 2014, 2015, way ahead of its time. And they had this series of articles about the AI revolution and the road to super intelligence. And they had no idea that, you know, large language models are going to play a part in this and how that was going to proceed. But the theory is basically, okay, we have this like slow incremental human progress over time. And then as soon as AI starts getting good at improving itself, you know, very quickly, it kind of hits that exponential curve and just gets like so good so quickly that it's like, you know, in seconds it's getting like, you know, all of human knowledge times 10 times 100 times infinity basically. And this release and that graph that you showed before where it was like, okay, here's our progress. What does this look like? Like look at this little breaking trend here. Like compare this graph, which is essentially all the recent releases of anthropic models, like from Opus 3, and they all follow a fairly similar straight line. And then you have this Mito's model that was just released and you can see it breaks that line. It would be like, you know, around where my mouse is here. If it was following that line, I just breaking it. And so if you look at this graph and you go back to wait, but why? And you look at this, you're like, oh, are we just like on this angle here basically? And is this happening? It's worth pointing out this is like a 10 year, 11 year old prediction. Like, you know, this isn't. It was cool. It was Elon Musk, I think back then. There are many folks that sort of say, you know, like there are some physical limitations to large language models. There's only so much better they can be, but you know, like it's. And also, I think the argument here is like, this model is not publicly available. And by the time we get this level of public availability in terms of like, you know, intelligence, like it may be further along that line to the point where that line may not look as broken as it looks now. That's my argument as well. Like when these were all released on day one and you had access. I would also argue that even if this is as far as it goes, right? Imagine Mythos is pretty far. Like it can't physically get any better than that for whatever reason. I think that's already enough. Like once people learn how to utilize it, that it's like it's hugely disruptive. Like coding, like, you know, it solves like 94%, I mean, 93% of coding. I don't have the benchmark here, but like it solves coding. Like it's coding is sorted and there's no coding without these tools going forward. And if you don't, you will get hacked by one of these models. So it's not even like you should use it for productivity is you will die if you do not use it. Right. And that's the best marketing argument. So yeah, it's like, it's going to change things significantly the same way opus is already changing the world. The Claude Code adoption has grown massively. We might talk about entropics, annual recurring revenue that has been basically growing by $10 billion in one month in the March, which is absolutely insane. Let's talk about that now if you want. Yeah. Okay. Well, I mean, I think this ties back into what I was saying about the IPO is like, they're trying to show revenue like hard this year to pump that valuation. That's the graph. And yeah, I mean, it looks like the wave of a wire graph doesn't it? Yeah. And most importantly, I mean, open air has not reported their revenue for a while. So it's like, you know, they may actually be closer than I'll. But it looks like they may have overtaken open AI right now as the leading AI lab. And the trend is vertical. I think if someone did a projection, if it kept going that way, which is quite unlikely for a few years, but it would be like 274 billion in a year or something like that, it would be like kind of insane. And so like, yeah, it's vertical. And so that's why they have compute issues. Like that's what's happening. And that's why they can do your mythos. That's all to say is that there are millions of users who are paying them a couple hundred bucks to get thousands of dollars worth of credits, essentially per month. So like they're still subsidizing us. And that's what they're taking in. So I guess the argument is, you know, they're buying market share and all that. But like if they stop subsidizing or when this eventually the gravy train runs out, like do we convert to being a 3k a month user? Not necessarily. Like think about it. If you distill mythos to the level of a Haiku, you probably get a better model than opus 4.6. For example, which for many people's jobs today is enough and provided, let's say you have a mythos level model that is kind of like your main operator, but it can call sub agents, right? Provided it just calls like, you know, an equivalent of opus 4.6, but that costs the price of Haiku like to do sub tasks, for example. Potentially they can turn your subscription profitable and you're still happy. So that's kind of like entropics bet, I think it's like as things are progressing. And I'm telling you, teaching people every day, people are progressing slower than the models. It's harder to move people than to move the models. And so my guess is like, they will turn that profitable if the capacities increase at this pace. Still, are you not financial advice? Are you buying stock in the IPO then? I don't know. I mean, I'm not even sure I'll be asked to be honest. But no, no, no, so in the IPO, it's like, okay, before I know that on the private market right now, you know, like there's a private market in San Francisco for people like employees with selling their shares, their vestibule, etc. People are fighting to buy entropic shares and there are open AI shares that are still left on the market right now and people are not buying them. So it's like, it seems like entropic is the bet right now. But the valuation right now is going to explode. But like my guess is right now, the big growth here is actually enterprise. It's not even your subscription. It's like they're doing a, like I think I read an article that say that eight out of 10 of the Fortune 50 companies are now entropic customers. So, so that's a very, very strong market penetration. So, 440 out of 50. Yeah. Yeah. And that means that, and I think that's where the growth comes from because they charge per token at the enterprise level. So they charge the API prices that are very high. And now an open area is pretty much the challenger at this point, which puts them in a very interesting situation with all this middle story actually. So, what is their response likely to be then? Okay. I want to show you a tweet. Like we're doing a Twitter episode today. So you can see this tweet from this guy called Adu, who I don't really know, but he said, it will probably be months before we use a model of this level of capabilities and share the benchmarks for Mito's. And there's a guy who says, but who is this guy? If you mouse over his avatar, you'll see that he's the hell of codex at OpenAI. OpenAI has also been bragging about having an internal model that is incredible and will change the world. Sam Lutman has done an interview with Axios last week. Greg Brokman, which I think that's his name, is the CTO of OpenAI. Did an interview on the Big Technology podcast. I listened to both and they're basically saying, like, they're basically telling people, like, my son, employment is coming. That's basically what they're saying. And so, like, they're feeling quite confident and strong about their upcoming model called Sput. And then you see that. You see this, which is a response to it will be months before we can use this, which tells me that that's the interesting proposition. Right. I think OpenAI is probably the only lab who can compete with this Mito thing. I think the Google models are not there. They can barely code already. So yeah, I think, I think OpenAI is the only lab at this point that is positioned to throw something out there. And so the question is, assuming they can throw a model out there that's competing with this, how do they release it? Do they follow on topic in saying this is too dangerous? So we only give it to big companies to patch. But remember, they're the challenger, right? Let's go back to the revenue number. If I can find it here, they want to catch up. They are also valued at an insane valuation and they're going to collapse if they don't make money. So they have a big incentive to cut the grass under entropic speeds. And really a model that is going to be closed, at least to the performances of this earlier, to gain market share and gain revenue and get back some of these enterprise customers. So if OpenAI does that, then Entropic will be forced to release something close enough as well to maintain their lead. And so we're either going to go into that fight between the AI companies and the consumer wins provided the internet doesn't collapse. Or we are going to go into that two-class system where enterprise has access to these best models first for like six to 12 months. And then the prosumers, the people with the subscriptions, etc, get kind of like a watered down version of that later on. And so that's what's playing out in this industry right now is are you going to be able to have access to the frontier with just your credit card in your hand or do you need to be a big enterprise? It's interesting you mentioned about jobs and stuff because they've, Entropic actually has this economic index section on their site and they show throughout all their like which jobs are mostly automating tasks, augmenting tasks and have tasks that don't appear in their data. So a cashier for example, most of the stuff that they do is not, they're not using AI for, but there's a few things there. Whereas a software developer or a web developer, half of the stuff that they do now is augmented or automated. I think the disruption is likely to be quite focused in certain areas unless so and others, which is obvious, but also they have this global usage as well. So you can see in the developed world primarily, the colors are darker green, the usage is much, much higher. So we might be moving towards this kind of like two-stage world where there are countries, companies, people that have access to this because they can afford it, right? It's not cheap, 200 bucks a month to use this is a lot of money in some countries. Three-stage world. Like you're going to have the four people, the kind of prosumers in the West and then you're going to have enterprise and then you might have three levels of models as well. Like they use the mini models, then you use the kind of like opusonet levels and you kind of get a tier above for enterprise. That was kind of the promise of AI is like, and that's one of the things that's interesting with open AI is like their ethos on that is very different from Entropic. Entropic is like the corporate overload, like the suits and ties. They want the big money from the big companies. Open AI, because they come from this non-profit background, etc. Like the company's culture is a little bit different. And one of the things they keep hammering is like, we're not necessarily going to make decisions, we're going to put these things in the hands of people and then like we think it's fairer if everyone has access to it basically. And so like, you know, broad access, cheaper API is in general as well and trying to do the best while distributing as much as possible, which could potentially mean that this spot model may be a bit under, but also available earlier. And then, you know, Entropic is forced to distill a version of mythos that competes with that or something like that. And we get much closer. That's my guess. I think that's what's going to happen. I think there will be a struggle between security and these models, but they will be forced because the force of the market is just too strong. Like they have the EAP with this year, etc. I think my main takeaway from all this is that everything we assume that would happen, you know, you kind of have this based on how things currently are, like where the trajectory is. It's like that our understanding of that is poor. And I think we need to really expect a lot of disruption and unexpected things in the coming months. And a lot of people are like, oh, I'm just going to wait till it like figures it out for me, etc. But the point is, I think the interfaces are not going to change very much. I do think like those CLI tools will stop you around. There will probably be a desktop app like the codecifc, etc. But you need to learn these tools. You need to use them. If you would like to learn how to use cloud code, which is probably, in fact, definitely the single most important work tool that you can learn and use in AI right now. We actually have a free guide. There's a few videos up there. So if you go to authorityhacker.com forward slash learn cloud code, you will be able to see how we're using it in our business to make sales, do our marketing, automate all our operations, these types of things. And we'll also show you how to set it up as well. So authorityhacker.com forward slash learn cloud code. And you can get access to that today. Let's move on though and talk about image models because there have been some leaks. I guess you would call it that. It's been tested in open arena. Like, you know, the art sites where they test things. And so, yeah. Yeah. So this is GPT image two, which is OpenAI's new image model. And I've got a website up here, the AI corner, and they've kind of summed up a bunch of these leaks right here. Starts off with this image of Sam Altman in a NASA spacesuit in space. And, you know, looks decent-ish, like kind of AI. It looks to be too cinematic for me. Yeah. I don't know. But if we actually look through some of the things which are shared here, you know, we're seeing pretty good, like real life images of the inside of a store. We've got a kind of like infographic showing the body. I can't zoom in on it. I'll show you some images then. I have some. This one I just wanted to show though is it's a photo of an IKEA building. But it's not a photo of any key buildings, an AI generation of it. And you would not know looking at it. It looks like a photo. So like the photo realism is pretty damn good. Yeah. Here's this one. I like the mundane photos because they're like the hardest thing to do, basically. Like, you know, make it look like a bad iPhone photo. It just feels like it. But like this is a Medicare sheet, for example. And you can see that maybe the handwriting is a little bit too clean, I would say. But I like the kind of like scribbles, for example, here. But overall, like, I think if you post this on social people will believe it. Even this, this kind of like technical documents is very hard to do with AI, to do something like this with the writing, et cetera. This cornflakes one, I would say maybe the hand, like the texture of the skin still maybe missing a little bit. But I mean, again, if I posted this on social and mess up, pixel peeping, it's very hard to tell it's AI. And you know, the box of cornflakes looks very good, which is quite interesting. It's like the world knowledge of this model is excellent in the sense that it knows what a box of cornflakes should look like. Now, it's not perfect. Like I think, for example, like, you know, the little like rating and whatever like the text starts falling apart, like the logo with the leaf here is also falling apart a little bit. But again, it's incredible. And look at this one, for example. It's like Obama and Trump taking a selfie together. I mean, honestly, like, it's pretty cool. You look at a lot of AI images and things like that. And it's really at the stage with this where it's, you cannot tell. Like, you don't even get a sampling. Trump is a little bit more orange in real life, but other than it's pretty cool. You mentioned the handwriting. I found this one, like I called Blake Robbins posted it, and it looks like actual handwriting, not kind of AI done. It's a bit messy. It's, it's great. Yeah. I mean, you cannot tell that this is AI generated handwriting, which again, poses issues for things like people signing documents and signatures. And yeah. Generate an ID photo for kids, you know, imagine, you know, a lot of places now that just you can show a photo of your ID and they're like, oh, it's fine. Imagine if your kid generates an ID that's like two, three years older than them, for example. We used to do that. But gosh, this is dating me now. Like back in 2002, 2003, we'd like scan in my passport and then go into Microsoft Paint. And like edit the numbers to make me look 18 instead of 17. And then print it off. And yeah, that worked. But I kind of imagine like what the fake ID landscape looks like for kids these days. So yeah. If you know, drop us a comment. But yeah, it's like, so again, that is also pretty scary. And this model again, for ads is going to be incredible. None of it is already pretty good, but it looks like we're getting another wave. And obviously Google probably has seen that. They will have Google I.O. in one month exactly today. And they basically haven't been releasing anything recently. So expect all the releases to come in exactly one month, including new Google models and everything. Anything else to say on OpenAI's image model then? No, I think it's going to be great. We'll wrap it up there. If you want to stay up to date with the latest developments in AI and how this is going to affect small businesses, online businesses, knowledge work, then make sure to subscribe to this podcast so you don't miss future episodes. We're releasing new episodes almost every week. We stay on top of all this stuff so you don't have to. And we just distill it into exactly the stuff that you need to pay attention to. So we'll save you from all the noise out there and just give you what you need to know. So thanks for listening to this episode of the Authority Hacker Podcast. Please, please, please do us a favor and whatever podcast app you're listening to this on right now, just go in there and leave us a review. Ideally five stars if you liked it. That is the single most important thing that you can do to help us drive engagement on all these audio platforms. So go ahead and leave that review. And if you know someone else who might benefit or might like to listen as well, do us a favor and just copy the link to them, send it over, and we'll try and grow the audience that way. So thanks again for being a loyal subscriber and listener. We really do appreciate it and we'll see you next week for another episode.