EP16: Nobody's seen AI UGC work, world models & AI agents go desktop

Like your DNA is just a set of code that puts out the program that is your body and your mind. Everything is code. I still feel like this guy is a massive fraud. They all look the same dude. So circle is round again. Coding in App is the new starting a podcast. We are three dudes, we just started a podcast. It's incredibly reckless in this world to not learn AI. Everything is code. I recently realized everything is code. Hello everyone! Welcome to a new episode of the God Mode Podcast. We are on episode 16, four months into this journey. And for the first time in history, we're all wearing a white shirt without actually discussing to wear a white shirt. Feels like it's summertime in AI land and we have two guests that are in the same country. I don't think we've had this before. Luca, Ben, same country, same building. Did you guys discuss to wear the same shirt? Just a coincidence. Luca picked me up from my hotel and white shirt gang, I guess. I don't know. I got the memo somehow. But yeah, Luca, it's good to be in Malta. Indeed, indeed. I think we're going to have a great week where we're going to share some skills and see what workflows everyone's been working with. I know I'm excited to show off a couple of things, especially with the genome analysis. The last week I've been diving even deeper into it. So it's getting more exciting. The more I learn about it, the more I realize that having a bit of a pharma background and now a tech background and crossing them over. Yeah, I think it's a good idea. Yeah. You have a pharma background? Yeah, I got a degree in pharmacy, man. What? What? That's new. That's new to me. I didn't know. Never. Not as impossible to learn if you have the motivation, I think. Wow. Yeah, that's new information to me. DNA is just code, guys. Like it's literally just that. Like your DNA is just a set of code that puts out the program that is your body and your mind. The more I dive deeper into that reality, the more I guess it does feel like a simulation in some way. Everything is code. Everything. Everything is code, Ben. I could read your lips. Let's dive in, guys. Let's see what we have on the docket here. First topic of the day submitted by Ben. So I'm going to let Ben do the introduction. Icon.com, what's happening there? Yeah. So Kenan Frost, he was the founder of Icon. He used to run Skio, which he sold for $105 million in cash. He then announced Icon's relaunch this past week. But the pivot is super interesting and it's kind of playing into the ethos that I think a lot of people are seeing. There's a decent amount of falling back to reality where people are wondering whether or not AI is actually providing the value that has been kind of sunk into this industry, right? Anyway, Kenan bought Icon.com, I think for like $12 million. He claimed from $0 to $5 million ARR at one point in like a 30-day period. But then in March 2026, the site went completely down levels on the left-hand side. He actually screenshotted their website. This was of March of last year. And look at it, right? It says the AI ad maker, but it pretty much exactly matches what is now, Rik, if you show the current website. The only one thing is the human ad maker instead of the AI ad maker, right? Like that's the only noticeable difference. Like even the kind of the story, the investors, the customers, customers and I guess have changed a little bit. But it's fascinating to me that they're kind of going back in this other direction. They're basically walking back. They're saying AI UGC is not a thing. At least it's not something that's really providing realistic results. And I called this too. I think we were all talking about this. Like I have not seen AI UGC actually work. At the end of the day, like a lot of my business is run on human UGC. Now there's a lot of optimizations around the editing, the payments, the operations, the communication, the outreach, everything around finding UGC. Sure, you can automate or you can use AI to help with that process. But the actual recording, the reality is that human UGC still works. And I think it's really only getting harder to make AI UGC at least today because people are catching on to these little intricacies that make a UGC AI versus human, right? Like I can pick it up super easily now. Like the more you look at an AI UGC, you just know that it's not. So yeah, total walk back. Kenan basically said, look, we're willing to offer six UGC videos for I think it's $50 each. So we'll see how well it works. But I've seen a lot of companies that are doing something similar. Jenny, Clueli, all of these, you know, kind of UGC first companies basically, right? Like most of their business, like Clueli, for instance, started as a, you know, the cheating out for everything. But my understanding now is that a lot of their revenue also comes from actual UGC and becoming an agency of sorts. Same with Jenny. Jenny AI, if you know that company. They were kind of like a writing assistant for universities. And essentially my understanding is that now they're making most of their money just from being a UGC agency. So all of these companies are kind of seeing that there's like a super lucrative opportunity there. And I think Kenan is kind of capitalizing on that as well. Yeah, I have two thoughts here. First thought is, yeah. First reaction to you is I sometimes see successful AI-generated profiles that I'm thinking, oh, this is actually such a good idea. Did you see these ads of the Down syndrome guy that gets thrown at who is selling like a product on the street? And then it's like a guy with Down syndrome and he gets thrown at like a water balloon or something like that. And then he goes home and then he shows how he created the product. Like it's totally AI-generated and it's like playing into the emotions of the people. But then like actually giving this product. This was, I think, the best AI-generated. I know which one you're referring to. And I guess it's the UGC that's being created that is almost impossible to do in real life that is maybe being successful. Whilst the, then maybe we need to mute whilst, whilst just whilst trying to imitate real life and to create something that's, you know, very achievable for a normal person to do. That stuff's not really working out. I don't know, because at the same time, I'm seeing a lot of AI UGC or AI content that is doing very well. So I don't agree that there's no space for it. I think there is a space for it. I think actually the hard part of creating a platform that is selling this UGC is that people or agencies are able to tap into the same frontier models that these platforms are trying to resell or compete with. And I think that's why platforms like Icon have not really kicked off. But then when you look at platforms like Weezy, more Higgsfield, a lot of people are actually subscribing to those and using those platforms to create all kinds of content, whether it's UGC or not. So I think there is a space for it. I just think Icon didn't really land the market very right. And I don't know. I'm looking at their page and it's like they swapped from going the AI ad maker to the human ad maker. And sometimes I'm looking at it and it's giving me doubts. How do we know that they're even going to be selling you humans though? Maybe they're actually selling you AI content just anyway. It's this shift from like I think their problem was that they try to be 100% AI generated or like 100% AI from start to finish. And then you can generate everything with AI. And this just doesn't work yet. We've been discussing this for like ever since the podcast is here. Video creation is 80% there. But the last 20% you need a human in the loop. And UGC needs more humans in the loop because it needs a human to actually record to feel human. And it needs a human to distribute in a way or edit in a sense. And yeah, to have this taste as well. And I think that AI and AI ad maker is just not there yet. And that's the positioning they're changing. They're adding a human in the loop which is probably going to increase the quality significantly or at least gets rid of these little clues that Ben mentioned. Where he said, I noticed that if it's AI generated, these little clues are there. And if you have a human in the loop, they can either make sure these clues are gone or they can actually have... Now they're becoming a two-sided marketplace where they have their creators and they have brands. And they're connecting the creators with the brands, which is an awesome business. I've been thinking about this too. Like two-sided marketplace is hard to build, but this guy has a following. This guy has been going viral a few times, maybe not with the best ways. But I think this is still the best way to apply AI in a way. Better to have 80% AI, 20% human in the loop than to have 100% AI product. That's like, I think that probably gives better results. And that's also nicer as a company to sign up for something where you know there's actually a human in the loop. Instead of... Same for like, Luca, if you would want to do UGC ads and on the original icon idea, the AI ad maker, you will have to do it yourself, right? You will have to go into the platform. You will have to change and choose settings and all of that. You will have to define and choose everything yourself. So it's more like you do it yourself with tech and you hope the outcome is good. Like Hicksfield. You buy a bunch of credits and you hope the outcome is good, but you might burn through all your credits. While with a new version of Icon, I guess it's still you give ideas, but then there's a human in the loop that then makes sure these ideas are executed correctly. So I guess it's like even... Like it's definitely a step in the right direction, I guess. Just a scalable in my opinion. So whatever his investors put in is not going to come back out. Yeah. I also just like, I still feel like this guy is a massive fraud, Kenan. I mean, scroll down, scroll down even further. This landing page is just has crazy accusations and testimonials. Keep going down further. And you'll see there's like the... Keep going. Keep going. Yep. Keep going. Yeah. So like the us versus them, first of all, this like table. Like why are they comparing them to ChatGPT? ChatGPT has a three-day free trial. Like I don't know, maybe they do, but full refund if you're not happy. Like they do too. Made for meta ads. Like they're comparing apples to oranges, right? And then if you go down even further, what is this? This AdMaker 3.0. You have the premium banana real. And then you have the enormous cucumber AI. Both are five stars, 200 plus happy brands. Like shout out appreciation, shout of frustration. You're buying it with like Amazon Prime. I am. Is that a thing? Like, I don't know. I don't understand any of this. The landing page of this company has always been... I mean, Levels pointed it out on the left-hand side. It's always been so perplexing. It's like all that they are playing into is the fact that a lot of famous tech investors have backed them. But beyond that, there is no credibility on this website. It doesn't make any sense. It's... There's something really odd that's going on here. For sure. It's very odd indeed. That's so funny. The more I look at this page, the more I'm like seeing red flags. Red flags, red flags. Is that the League of Legends logo? What the fuck? There. And explore more. Story. League of Legends. I guess who's the plague? Well, yeah. Dodgy stuff surrounding this. No social media posts. No press releases. And no communication yet. Keeps rebilling customers just days ago. That's been going on for months now. So... This is a story I would love to continue following. Very interesting. Icon.com. Yeah. But I believe maybe one takeaway here. I think like UGC actually is the best way to market apps. And Ben, I think you've been learning this firsthand and seeing this firsthand that like your apps or your revenue grew by 4x since you actually took UGC serious, right? Yeah. It's definitely a... If you do it well, it's way more... The upside is so much higher than going the meta ads, the Google ads route, right? Because you're essentially playing into the fact that everyone's glued to their phones and reels are pushed so heavily and TikToks. If you can, you know, partner with creators and find a strategy that like, you know, works repeatedly. Yeah. It's all about just kind of negotiating with creators and finding that strategy. Whereas with meta ads, it's the opposite. You're locking into a certain cost per thousand impressions. And that's kind of, you know, how much you end up paying, whether it's $3 per thousand impressions or somewhere up to $15, $20, $30 per thousand impressions in the US, you know? Yeah. I think that the move is combining the two. Using the UGC creators to partner with them, see what content actually works and then pushing the ones that work on meta ads. I think that's the winning strategy. It is. But then again, you're locking in a certain cost per thousand impressions, like I said. So my experience advertising to US markets, let's say you find the best Instagram reel that performed really well. And then you try to boost it in the US, in, you know, New York, LA, like certain markets, you're going to be paying $30 per thousand impressions, right? Whereas with UGC, why don't you have that person recreate that reel and then spit it up again and again and again, and you may be able to get them to hit it, hit it again. The goal in UGC, like a really good UGC campaign could be performing at a dollar per thousand impressions, right? I've seen a lot of those as well. So we're talking a dollar versus 20 or $30 per thousand impressions. Yeah. Hmm. But then again, like if you find these real winners and I've heard strategies also where like you, for example, you have your app, but then you have your network of creators as well that you partner with these creators that they then, that you can even run the ads via their accounts. So that like, that's, is that a way, like how would you approach it that way? If you have these real outlier content, because if you have a hundred UGC posts, you don't have to push 20 winners, even like just take the top five or top three and put some, a little bit of money behind that. And that might outperform the other 95 pieces that have been posted, right? Or are you just paying on a pay-per-view in a way with these UGC creators? Yeah. Most people do like a base payment and then a bonus versus based on how much, how many views they get. So for instance, I think for my creators, I'll do like somewhere between four and $600 if they hit a million views, right? Now a million views, if you were to again, pay for it, that's a thousand, a thousand, what would that be? That would be 10, 10. Yeah. Yeah. A thousand times a thousand. So if we're talking about again, a CPM of 30, you're paying $30,000 for a thousand videos that got a thousand impressions. You know what I mean? It's, it's, it's way more. So yeah, that's, that's kind of the flip side here. I think that a good strategy is using retargeting when it comes to meta ads. So retargeting can help you target those folks that you've already had seen your video. And then you're able to give that to meta, give that to, I think Google maybe as well, but meta is better at the whole retargeting game and say, Hey, these are the users who've seen my organic content. Let's start serving them ads again. And I've seen some success there too, but UGC is still the game for me. Yeah. What's nice is, is it's kind of, you can drive traffic with UGC from TikTok, from, from meta platforms, but then being able to retarget all of them from Twitter. But then being able to retarget all of that traffic through meta, for example, is, is, I found quite useful. So having some campaigns that go viral on X, that go viral on TikTok, but then retargeting that campaign, that audience through, or even your organic search through, through meta. Those, those give me my best returns for ads. Yeah. But funnily enough, I think as well, like this is something I've been doing locally, right? It is creating a lot of fun reels that go kind of viral. And it's for my country, they're reaching about 30% of the population. So what I think is happening, and this is just, and I, I don't know if it's true or not, I don't know how to track it, but I think that maybe that when I'm then running ads to this same population, my face is being recognized from like dancing on TikTok or, or making a funny reel, you know, and that's maybe lowering the cost of my ads. Because when you see a face, you recognize you're less likely to keep swiping, even if it's an ad. And I'm wondering if that has an effect on my, my final ad spend. Yeah. You have a very interesting setup there. Like a nice, I think about this as the island in Thailand where I, where I like to go. It's like this small economic zone on itself where like you could try anything. And I feel like Malta where you are is similar. Or you have this test ground to see what works and you can play with the ads and you can play with whatever business, because it's just a small place in general to test stuff out. Right. You mentioned something interesting last week as well. I don't know how it came up, but you said pixel them all, Facebook pixel them all and then retarget. This was related to something AI as well. Do you recall? Can you recall this? You want to share more about this? Yeah. I'm just, to be honest, I've been meeting a lot of people who are building software with AI and this sort of thing. And they have never heard of a Facebook pixel. So that was a bit shocking to me because I thought that was something very basic to understand. And I mean, when you're creating a pixel from Meta, it's something that is basically capturing, is adding cookies to all of your users who are visiting your website. And then Meta is cross-checking that cookie with their IP. So it will actually know which profile visited your website. You won't know which profile, only Meta will know. But when you have about, you know, at least a thousand of these people on your pixel and you have so-called seasoned your pixel, you are then able to retarget that audience directly through very direct Meta ads to those people and then to create lookalike audiences as well, where Meta will analyze who that is in that audience and find people who maybe like similar things or have a similar psychological profile. One thing that we didn't mention as well with UGC is it's important to also season new accounts. And Ben and I were also talking about this, even where you open that account, which country you open that account in will affect where it gets distributed to organically. So I'm going to be trying some of these experiments as well with a new Instagram account. But before I post my first reel, I'm waiting a week or two, liking some stuff, giving it some kind of warmer process. I think that's something important for these kind of strategies as well. We had a coffee and connect two days ago, I think in the garden here. And the topic was media buying. And there was like brought up by two people that they actually get better results by using the Facebook or the Meta AI suggestions versus lookalike audience. Have you noticed anything yourself there? Have you been running ads with the AI built in versus lookalikes? Yeah, but I mean, I've been using these new features from Meta, like AI audience expansion, AI, sometimes even allowing them to rewrite my ad, you know, and choose which words should come first or last, which you need to be a bit careful of. But the way I've been using it is I would let it rewrite my ads. Then I will see which ones are getting kind of some traction. And then I will look at that ad and be like, well, maybe I need to change this word because it's not accurate or so describing my product correctly. But I have been trusting it to at least go and test different media and different kind of combinations and then reviewing the result after a few hours or a day to make sure it's accurate. We have this post of A16Z, the World Lab CEO, Dr. Fei-Fei Li, with a quote, the world is not made of words. Going deeper into language models and what's behind them, Ben, can you introduce one? Yeah, sure. So Dr. Fei-Fei Li, she published a functional taxonomy of world models on her substack on June 3rd. A16Z, they signal boosted it. Basically, the frame was that language models, they learn the statistical structure of text. World models, on the other hand, they learn the statistical structure of space and time. So basically, she breaks world models into three buckets. You've got the renderers that output pixels for human eyes. You have the simulators that output state like geometry and physics. And then planners that output actions. Her thesis, though, is that the simulation is like the underrated linchpin because it's the substrate for which both pixels and actions derive. So what does all this mean, essentially? Large language models is only one aspect of this kind of what comprises, what is needed to achieve AGI. And I think A16Z is kind of also doubling down on that as well. They're basically saying that, you know, in order to have this AGI, you need to combine, yes, the text and reasoning, which is LLMs. But you also need the space, time, light, force, physics aspect of understanding the world to really be able to get us to this, you know, land of abundance that Elon talks about. And so, you know, it's a very complicated thing that I think, like, you know, there are many ways to break down how we look at the world and how we make sense of it and be able to get to this, like, you know, utilitarian, okay, we've made it land. But I go back to the figure robot example that we looked at a few weeks ago, right? And thinking through, well, how is that built and how universal of a solution was that? I don't know how far off we are, but I think it is way further off than maybe some folks think, and maybe how Elon thinks about it. When it comes to moving this AI from, like, you know, an LLM world, sure, yes, we can, the information age is changing. We all know that. If you are a measurer, as some people call it, your job of moving things within Excel is going to change when it comes to LLMs because they've gotten to this point where they're just so damn good, right? Opus 4.8 can do 70 to 80% of the tasks that most people who are sitting at a desk can do. But when it comes to this unified world model where you have someone in your house that's a robot that's, like, assisting with things or on a manufacturing line, it seems like we're kind of maybe further off, according to this woman, Fei-Fei, and A16Z is, like, doubling down on that as well. Mm-hmm. Yeah, it's interesting. A new, duh. I saw a bit, well, first of all, really well explained, Ben, because I was listening to this on the paper's talk of YC last week. I mean, like, bits of it was coming through, but I think it's a very complicated subject to kind of explain simply. So I think that was a good explanation. But it does make sense, hey, when you think about it, that there's the LLM part, which is the knowledge that you have in your brain, that is what you study, what you read about, your experiences. But then when you wake up in the morning, there's always news, there's always the weather, there's these different inputs that will affect your perception for that day and how your decisions based on those inputs. So, yeah, this is where I think that AI is going to expand a lot more into, is having more kind of context on environments rather than just being kind of these one-off knowledge bases. Yeah. Because I guess this is what leads to having awareness and consciousness. Totally. And another way to think about it is, you know, we just talked about the icons and the AI UGC thing, right? And that's in a world where you're generating video. Video is, you know, still not necessarily three-dimensional, right? It's just being able to even output, output something that looks like, you know, like a two-dimensional kind of video. Right? Getting to the 3D world, this person, you know, assumes, and I think a lot of people are maybe thinking that as well, is exponentially harder than just even being able to generate AI video. And, you know, we haven't really seen AI UGC or any type of AI video really be able to be produced at a mass scale that can emulate like, you know, how, for instance, if we were generating a podcast like this, it would be pretty wild to see a podcast in real time that could live up to something like this, where just three people are talking. Well, I'll be honest. I've listened to a podcast and a couple of them where they get two philosophers and they make them debate each other. There was one that was super interesting was Machiavelli versus Marcus Aurelius. Both of them titanic titans in leadership, titans in how to have a good philosophy for running an empire, a country, a company, but very, very different ways of doing it. One much more empathetic, one more understanding, leading through empathy with Marcus Aurelius, you could say, and one being cutthroat and, you know, almost evil, which is Machiavelli. And having two AIs, one of them trained exclusively on Marcus Aurelius' text, and one of them trained exclusively on Machiavelli's text. Having a debate with each other was super interesting. I listened to it for two hours. But this was, was this text or video based? Audio. Audio. Audio. Yeah. Because audio is basically just text turned into speech, right? But if you don't want to turn it into a video, that's going to cost a lot. And we're using large language models, predicting each pixel of the frame to then add a lot of frames together, which becomes a video. And one second of video needs 25 frames or 30 frames, the frames per second that you select when you make a video, right? And what's the price of generating now? I think like you pay almost a dollar, like a dollar per minute. That's the price-ish. And then you're not, you don't even know if it's going to be good. We're using a technology, large language models to generate visual. And I think if unified model, unified world model might be the next step after LLMs, then that's going to enable, then suddenly the video production can be, the cost can be cut significantly as well. Or we get, this is actually the next step in, yeah, the next step in AI where we've been using LLMs, large language models for all the use cases that we've been seeing, which have been great. But we're also like, because there was no other technology available for doing this. Yeah. Very, very cool. Yeah. And that might be a good segue into the other one that I put on the docket, the open-sourced emotion one. And why don't we play that video to kick it off? Well, so this one's cool. I mean, for one, I think it's open-sourced. Yeah. So we've open-sourced the model weights and then they'll have API access. But what I'm really intrigued by here is the 110 milliseconds of latency, because when it comes to conversations, you basically need to get around 200 milliseconds. Like that threshold is where it feels acceptable for me to answer something and you to be able to like pick it up. And they're below that threshold. And I thought what was amazing about this thing as well, and I've been digging into this a lot the last week, I've been going down a rabbit hole of trying to figure out how to build automated Instagram reels with a narration that's going on. And I've noticed that 11 Labs and a lot of these providers so far, they sound too good. They sound too professional, right? You almost need a filter because like the context of a YouTuber speaking versus the context of a friend speaking in a room or in a podcast, the therapist, all of that. There's so much context built into the context in which that conversation is being had. And this demo to me made me really excited about like seeing how you can use AI to emulate voices in very different environments along with 110 milliseconds of latency, which is amazing, right? Because that's one tenth of a second. And I think you and I can agree that if we're having a conversation with someone like, let's say, a therapist, you really do want someone who's able to respond in that time frame. Now, the big thing is like how good are the responses? How AI generated are they? Are you feeding that then to a model? If you're feeding that to Claude, it's probably going to take a decent amount of time for Claude to be able to come up with like, you know, significant response that a therapist might say in that moment. But there could be some pretty clever things where, you know, maybe a therapist starts by saying, yeah, I totally understand, Rick, that you're having that issue. And then in that time frame, it's able to generate something that's really a thoughtful and meaningful back and forth conversation to be had. Yeah. Yeah. There's also, like I see, like they open sourced it. And we've discussed this previously where you see more and more businesses open source their creations. And then is this a business strategy that's been the new way to onboard people into your ecosystem and then somehow offer extra services or extra things on top to then like post this? Like what was it? Clicky and some others. We discussed all of these, right? It's interesting to see that. Open source is the new business model with open sourced is basically the freemium model, right? So you have open source be the first version. That's the one where you turn on the hose and you get to try it out and all of that. But that's not to say that the second, the third version of the application or product won't be open sourced or at least some component of that if you want the premium aspect, you know, whether the API that gives you the podcast format or the teacher format. Well, that's where you have to use their actual closed sourced version of the model. Right? So I think of open source as a way for people to be able to try things out, demo it. And with software, everyone wants the latest model. And we've seen this like Opus 4.8 comes out and everyone's like, oh my God, that's amazing. So why not give them the Opus 4.6? They try it out. And then they're like, by the way, we have Opus 4.8 and that's not open sourced. You got to pay us for that one. Yeah. Suno, the AI music generator does this since the beginning where you can use their older model. You get two, like you generate, it gives you four outputs. Two of them are the older models and you can directly download them. And two of them are the better sounding songs, but you got to pay to use them commercially or to actually be able to download them. So that's another way. That's not even open source, but that's the freemium done well, in my opinion. Did you see, like talking about AI voice, did you see this retweet of Jason? I'm going to play this video as well. It's quite funny. Your business runs on business. So I'm going to subscribe. Sounds really good. I don't know what's the VO script. What's a VO script? Voice over. Voice over. Okay. This is so good. Business AI is here for your business. Oh, well. Shall we jump to the next one? This is a reflection. This is basically a reflection of how Uber, I think they almost catalyzed this whole questioning of the return on AI, right? And a lot of people are saying, we've spent so much money on AI. Is it really working? And, you know, I still think we're headed in the right direction. But as costs get out of hand, people are kind of starting to question and ask those things and saying, like, yeah, is this just the next, you know, hype wave? Yeah. And I feel like this is going to contribute to maybe the bubble popping. If you remember the All In episode with the Salesforce CEO, I think two, three weeks ago, where he was saying that they're spending not a lot yet in terms of their $300 million on AI, which is not a lot compared to their total market cap. But they're not looking to grow that significantly in their spend. But all the valuations of those big companies of Entropic, for example, they're forward looking in like, okay, these revenues are going to continue growing. But now we see the businesses like, like you say, Uber, like you say, Microsoft, cutting down on their AI spend. So then what's going to support those big valuations that are suddenly becoming public companies? Is that then actually the bubble popping? Because then suddenly revenues will not have the same growth as they've been having because all the clients, all the partners are actually lowering or getting aware of their huge spends that don't have this return on token spend. Yeah, this happened in fiber. This happened in fiber. There was an overinvestment in delaying of fiber cables all over the US. And it took much longer than anticipated for that capacity to actually be used up. Eventually, it has been used up. And they thought that investment did well for the infrastructure and for the society. But from a pure return on investment, it took much longer than expected. And maybe we'll be in a similar situation with compute and with AI and this sort of thing where every country is racing to stand up their own national competence in this area. And we might end up in an area of overabundance. But for the consumer, that can only mean good things. It can only mean lower prices. As an investor, it's a scary thing to think about indeed. But then if you're a company like Google, like any of these companies, they have a lot of cash on the balance sheet that they need to invest. And this is the only way that they cross the chasm and do not get left behind and have some other company become the tech leader in the space. So I would argue it's still the right decision for these companies to do. Yeah. I know last week we're also talking about IPOs and SpaceX and there's the Antropic one and the OpenAI IPO. Yeah. Maybe it's not the smartest to go into those at this point in time. And that's the last kind of squeeze that's going to happen at the top of this valuation craziness that's going on. Yeah. I've seen some other tweets saying that ahead of big IPOs, many people sell off the risk in their portfolios to then allocate to the hot IPOs. And then that followed by potentially what we've just been describing, the consumers of AI cutting their spendings. It's a very interesting dynamic that's going to happen over the next six months with a lot of moving pieces in terms of the capital that's involved in all of these. Well, time will tell. We will see. Next on the docket is agents. Let's talk agents. We have Now's News Research announcing their Hermes Agent desktop app. So we don't have to... Let's play it without music. We don't have to use our Hermes agent anymore from the CLI or the terminal. There is now a desktop app just like Claw Desktop, just like Codex, just like ChatGPT, any of those. This is great for Luca. He can offboard from Replet and can go onto Now's Research to start using a ChatGPT-like interface to then have a self-learning agent that you can give a soul, that you can give everything that you could give OpenClaw, but now all within one desktop app. Guys, most importantly, you can choose whichever model you like. Like, that has to be the best thing. And I have been using Codex and Cloud Code over the last week. Definitely been missing out on some advantages over there, especially when I'm working with large files, especially the DNA files. I've found that the best way to work is to have your offline agent. Again, I was disappointed because I don't want to be locked into any company at this point in the coming year. So seeing Hermes basically create the exact same UI, but I know I'm going to have the power to choose what's working in the backend. It's really cool. It also means that, you know, you can have your closed subscription, you can have your OpenAI subscription on the lower tier and swap them out whenever you want. That's super good. No, it's amazing. Indeed, I downloaded it. I signed up. It took me three minutes. I linked my ChatGPT subscription in there and boom, you could just get started. And like you say, you can choose whatever model you use. So I think this is great. It's kind of moving away from, oh yeah, you have to do it through Telegram, which was funny because that was the original unique selling point of like, you can use agents now from the apps that you actually use most. You can use agents in Discord and Telegram. And now we're back to, you can use agents in the desktop app. So circle is round again. Yes, they're still without a mobile app, I imagine. But would you, would you then, so you can't use this remotely. So you like, you'd have to use it on your laptop. This is literally, yeah, I think this is sort of, I saw this as a comment in Twitter. I think this is sort of a research preview or like the first, literally the first version. I'm pretty sure the mobile app is coming very soon. I'm pretty sure you can load this up on your VPS as well. You can still have your Hermes agent running in terminal on a VPS and talk to it on Telegram. That's not gone yet. It's just making the light, like making it much easier for anyone who never used a terminal before to now actually get started with a agent that's self-improving and has basically the same functionalities built in as any other IDE or ADE. How do we want to call it? Ben, did you play around 20 agents yet? I haven't, but I'm really trying to lean into figuring out how to make Claude Code with a virtual private server act as my open claw. And when we're in Malta in the next week, like that's like top priority for me is to figure out how to get Telegram back, right? We could get Nick back or whatever, and I could somehow use my Claude Code OAuth subscription, but running it on the virtual private server. And if it can do everything that my open claw agent was able to do, I don't see why I wouldn't be able to. And I do wonder why Anthropic and OpenAI haven't really been able to match this level of agency that like Hermes is proposing here. What could be the reason? Is it like, do they think that giving this much autonomy away will lead to a lot of issues? Is it risky for them? Because OpenAI and Anthropic, they've really still leaned into, no, use our application, use our application, but they're not letting you use it, you know, on the fly, in the cloud, in a more accessible way. And it just seems like such an amazing unlock. Yeah, it's interesting why they have not done this because every release of new models, there is progress or there's improvements that allow users to spend more tokens and how to make people spend more tokens. It's by enabling using your OAuth for OpenClaw or Hermes. And you can do this for a chat GPT, but your OpenAI subscription, right? You can use that for logging into Hermes. So it's kind of just Anthropic. But why isn't OpenAI doing this themselves, right? And I know Microsoft Build came out this week and it seems like they're trying to build OpenClaw into the Windows application itself, which is, you know, that's cool. But I think we all agree Microsoft probably is not going to do the best job at implementing that functionality. It's not, but so many companies are on Microsoft governments and they're kind of stuck in that closed system, right? So what I saw from Microsoft is they're going to put this Clopilot available for people who are using OneDrive, SharePoint, Teams, Outlook. So you can model, we'll have context on what's happening in your company, what's in your email inbox, what's in your shared drives. And I could not like to try it out. I would be more excited if Google released something like that because I know I'm native on Drive and Google Docs and that would be more interesting for me to see. But yeah, it's crazy to think that then you have Antropic not implementing any of that sort of functionality. I would assume it's the security side that they are afraid of since that seems to be their biggest championing point is they're trying to save the world from AI, right? I still don't believe them. Let's play a game. I've put together all the five IDEs I have on my computer. The first one you're probably going to guess, but I'm going to make each of them bigger. Here, don't look at the name on top. Don't look at the name in the top right, left corner. Okay. Just say by the interface itself, which one is what? Okay. What, which one is this? The one that's big now. That's, that's cloud code. Cloud. Yeah. You see it in my history, I guess. Not the most used. But that's the only one that looks a little bit different. Then if we go next. Okay. Here, here we go. Next one. What is this? Is this cursor? Wrong. They all look the same dude. They, they literally. This is anti-gravity. This is anti-gravity now. Right. Okay. Then we have next. Here we go. New agent automations, customize and home taps with different projects. What is that? I'm pretty sure this is codex, but it looks really similar. I know one's light, one's dark. Like that's kind of the vibe. This is cursor. This is cursor. That's cursor. Wow. This is cursor. Yeah. Cursor has changed a lot. Right. Crazy. Then we're going to the next one. This is codex. Yeah. Correct. There's literally barely any difference. It's like, let's see what codex says. New chat search plugins, automations, codex mobile. Cursor says new agents versus new chats. Automations. Same. Customize. I think customize is exactly the same as Claude has. Claude has new session routines and customize. So it's like just copy paste. And then the last one here is, yeah, the one that we just discussed, right? This is the Hermes. Hermes. Yeah. Okay. Yeah. The Hermes agents or the desktop app. Same. New session, skills and tooling, messaging and artifacts. Like it's also the same. So then it maybe is indeed like maybe the Hermes one is a great unlock. Same as Cursor where you can use the different models into one. Yeah. I'm wondering how the harness is because we've been talking about harnesses a lot as well. Right. Well, like is how would Cursor perform versus Hermes agents on different harness benchmarks? And that's something I would love to see over the next week to come out. Would be nice or interesting. So yeah, guys, I thought you would score better on this game. But hey, final topic. Go ahead. If you want to say something, otherwise we move. Final topic. We have AI slop design. And I'm actually wondering on your opinions here. Do you guys think this has been a topic on discussions on Twitter? The four horsemen of the apocalypse. And then sharing four design principles, which has actually been used by AI a lot. And how people can easily see your app is vibe coded with the purple gradient, with the live, with the letter spacing and with these side things. And I actually noticed that I have these on my apps as well. But what's your thoughts here? What's more important? A good product or a good design? First, the good product. Then the design. But there's some basic things that you obviously you fixed before. It's not even basic design. It's just stuff you don't you didn't want there to be that the AI put there for you. Like the live thing is so ridiculous. It gets placed in the weirdest places, you know. So there's some things which is just like, you know, proofreading your essay. If you write something out for the as a dirty draft, that's your dirty draft. You just clean it up a bit and then you have something that you you can ship. But but what you ship doesn't need to be, you know, build design or you don't need to have some kind of Claude looking launch video. I think that's overkill. Yeah, I mean, there's some basic stuff which you should look out for. I think people get way too pissed off about the gradients or the purple color. Those are in my opinion. I like purple. I like gradients that are so like, like, yeah. I think most websites designed by AI and even by humans have like some form of a gradient. So, yeah, I don't think it's going to be the key to your success. So if you have a good product that you ship it out there and fix things along the way, these are just some basic things you can look out for. Don't treat them as as I wouldn't treat this advice as never design with AI because it can't design. It's just, you know, hey, if you're designing software with AI, these are some common things to look out for that can make your product look a bit more professional. That's all it is. Agreed. I still feel like UX is just so much more important than UI. And at the end of the day, most people like we may recognize whether or not it's AI generated, but so many people don't notice it. And I still look at people who have embraced lovable and they went from their Wix or their Squarespace before and the app still looks so much better, like whether it's a personal website or something like that. I spent the last couple of weeks going through there's a Wikipedia article I posted in the chat that's called signs of AI writing. And I think this is actually a really good exercise for people to do. Probably my suspicion is that Google and other, you know, Bing and other search providers are probably going to start identifying like what's AI slop. And they will start deprioritizing apps if they haven't already that have these like telltale signs that they just used AI directly. So I had all of my apps basically be rewritten the verbiage, the copy on the site and said like, hey, look at all of the signs of AI writing and let's remove that from my website so that we can kind of almost remove the paper trail that, you know, AI was used in it. You know, that being said, obviously, I'm also like prompting, right. And as the models have gotten better, it's my conviction and being able to build good products has almost been amplified by, OK, if you're a great architect, you are going to be an even better architect if you can use AI. And same with a developer, same with a product manager or a real estate agent and all of that, because the AI is just going to create the same thing. Then it's tweaking it beyond that. And, you know, having all these this knowledge of, you know, let's look at the Wikipedia. Let's remove all those at AI slop. Like it's basically giving you that V zero. But then there's so much rework in the UX to get it to a place where it's beyond just AI slop. Yeah. Was it did you hear this, too, that people that there is going to be someone predicted the shift happening from now people hate it when they see that something is that something is AI written to people might starting to appreciate AI written things. Was this the podcast that we had as content as the week last week? I wasn't been mentioned there as well as a contrarian take. I mean, I don't disagree with that. At the end of the day, I think AI is improving my ability to communicate with other people. Why would someone be unhappy that I'm communicating more clearly? Obviously, it doesn't mean you should use it recklessly and not proofread, right? But if you're improving the way you communicate, both from text or visually, that's only better for us to function, right? Like if you put a little bit of soul in your creation of your write-up and... And on the topic of writing, I'm actually writing two books at the moment. One of them I started over a year ago. It's been being titled Ambitiously Insecure. It's a bit of a self-reflection journey on my own evolution as an individual and what motivates me, what scares me and all of this stuff. And then I came across... I was going through some old journals. And in 2016, I wrote the plot to a movie or to a book that I never wrote. And it was called The Diversity Isles. And the whole plot is, you know, there's one million people put onto a social experiment with carefully balanced parts of society, different genders, ideologies, race, all of these things. And then the whole plot is a bit of like a sci-fi twist that we, even in a perfectly balanced society, we still find a way to discriminate against each other. And so I'm now experimenting, getting that storyline that is obviously my own invention, my own intellectual property. And then using AI-assisted writing to accelerate writing this book. And then using AI to create a trailer for it, to create some social media content for it. So try and get this book to market it, you know, and maybe even to tell the story. And maybe turn it into its own series, you know. And I didn't just write a book and I am actually going to be able to produce a TV series. Why not? So, yeah, if you're a creative person with ADHD, there's a lot you can do in today's world. So it just takes a bit of fun. You enjoy it, I guess. It's AI-assisted creation. You get your ideas into something and you let AI structure it better. You could just draft it all or talk it all out. Like talking is the new typing, right? I was wondering, so there's actually three books in the making, Luca. The one, there's this one with chapter 7, the quick ghost. And chapter 8, the Instagram ads. And chapter 12, the professional rejection. Well, yeah. That one, I've written a couple of dating stories. I don't know why I published that one. That's an Anon sub stack that's coming out. So if you see one of these titles, you know it's been Lucas Anon sub stack. Great. All right, guys. So this was a great episode. I'll see you in two days in Malta. And yeah, next episode, we do a live one. See you later. Bye-bye-bye. Bye-bye. Adios. Ciao. Coding and App is the new starting of podcast. We are three dudes. We just started the podcast. It's incredibly reckless in this world to not learn AI. Everything is code. I recently realized that everything is code. Luca, what are you predicting? Underwater drones run by people. Okay. Wow. Bombshell. Use an AI to write your promise, bro. Come on. I am the greatest moderator in the world. And we see you guys next week. Love you, besties.