Ai Take_036 - Does AI Even Have a Product?

[00:00:00] Is AI real intelligence? And is it actually going to create a massive general intelligence machine that can replace all humans in all jobs that nobody can compete with, and the need for humans just vanishes? And what if that reality is just a few years away? [00:00:23] Or are the economics not panning out? [00:00:28] What if the overwhelming cost of bigger and bigger models isn't actually justified by the lesser improvement in performance? [00:00:37] What if instead we're headed toward a future of small and very specific use models that just become integrated into our software and everything we do and we just expect it to work like that? [00:00:53] What if AI doesn't even have a product? [00:00:58] I am Guy Swan, and this is AI Unchained. [00:01:12] What is up, guys? Welcome back to AI Unchained. I am Gaiswan, your host and this is where you will hear all about the open source world and future of AI. And I want to talk about something that. I know it's been a little bit since we had an episode. I went to Logano and basically just have been drowned in other stuff. But I've been meaning to get back to this because it's a little bit of a follow up to the last episode. [00:01:40] But I wanted to do a show on whether or not AI can actually sustain a product. [00:01:50] Is there anything here? And this is something that we talked about with Alex Svetsky. [00:01:55] And this does not mean that AI isn't useful. And I want to specifically make sure that that is understood. And if anybody's listened to this show, you know that I think AI is insanely useful. Like it's an incredibly useful tool and they're generative. AI specifically. I mean, that's, that's what a lot of this show has been focused on. LLMs, basic machine learning, captioning, translation, image and video generation, all of these things. And some of these things do have, like, smaller products, I would say. But I have a really hard time deciding whether or not they will last. And LLMs in particular seem to have some attributes about them or some elements about them that make them impossible to sell as a product because they aren't a product, they are an interface. And I want to dig into what I mean by that and why. [00:02:56] Why is OpenAI failing? That's really what this episode is about. And I'll recap a little bit. We'll actually start this off by recapping exactly what I mean and by explaining how and why OpenAI is in fact failing and likely does not have a very long Runway left. [00:03:15] So for anybody who didn't listen to that one and or read that Article the Subprime AI Crisis is what it was titled, actually, if you didn't dig into that one, I'll give you a kind of a brief overview here. So essentially he's arguing that the artificial intelligence boom is unsustainable and is basically going to collapse and that it's going to have a cascading effect through, through the industry. And there's actually been a number of things that had me slowly working my way there. And he's got a lot of really, really good points to consider. [00:03:51] But just one of the big things for me has been my experience. We were getting so many. And there are still. I want to make sure to say that there are still actually a lot of very big advancements happening pretty quickly, especially in. [00:04:07] I've seen a couple of really good ones in video generation and like models that do very specific things. [00:04:15] The portrait stuff has done really good. The live portrait and X portrait and these sorts of tools that you can actually get out there and you can find these on GitHub. These are things that let you change the facial expression of an image of literally a static portrait of somebody, and then also record your face or take a movie clip or anything else, and then basically place it on the face. So it's very specific face manipulation and facial animation and emotion and feeling and all of these things, which I think is an incredibly useful tool. [00:04:54] And that specifically, I think the application of a lot of these things to very explicit purposes and very explicit tools, you know, this is something that's not broadly useful, like to just be able to like generate faces or animate a portrait. Like a ton of people would go throughout their day and not need that. But this is gold for storytellers, for animators to not have to manually animate stuff in order to get your story across. [00:05:26] That's a big deal. And a lot of those really specific AI tools I think have an ups, have a margin on them. I think they are sellable as a product. I still use midjourney. I still pay for Mid Journey. I also still pay for an LLM. But I'm getting to a point where it's very, very difficult for me to justify it because of the ones getting better and better the smaller they get. But here's the thing that I really think is being sold. I don't think, you know, we used to sell operating systems. It used to be really expensive to get an operating system or to get a lot of software. And it wasn't long before, after a few generations of it, before it got to the Point where you realized that you want the operating system to be free because you want people to get into the ecosystem so that they use your stuff, so that they use your tools, so that they need the tools that are compatible with yours, your operating system specifically. And so the operating system started to become a loss leader. And then you saw the same thing with search. Search is an incredibly useful tool, but you can't charge for it. Nobody is going to use a search where you have to go pay to open up the search box or you have to subscribe to the search. Why? Because somebody's going to offer it for free and it's going to be too easy to offer a comparative tool. [00:06:51] And I mean that's a little bit less true in Google or something specifically just because you have to do so much in depth indexing. But this is very true when it comes to AI. And I think that's a part of the reason why it's, it's suffering a bit. It's seen just an excruciating amount of investment because it was so flashy, it was so big and it was so in your face and it was so cool and it was so alarming that it got a lot of attention. It was very much like blockchain. But I think it will suffer the same fate as the blockchain technology investments. I think a lot of them. It may not be as spectacular though. It also could be. But the just going back to the article is he, he specifically makes the argument that it is unsustainable and it'll collapse. And one of the big things or the big indicators of this is if you do some basic math on the numbers that have been put out, OpenAI is losing a gargantuan amount of money. [00:07:54] And while they keep putting out, you know, new models are coming out. [00:07:59] They're not, you know, I'm not using chat GPT01 or 1 0, whatever, whatever it is. Their, their new one, the one that was codenamed Strawberry. And I've seen a lot of people say, oh look, it did this really cool thing. And then I've seen a lot of people say, oh, it just completely screwed this up. It hallucinated a bunch of stuff. And, and you can't just naively put a, and I'm not saying that they did this without any thought at all, but you can't just put a model to say, to judge whether or not the model's first answer. This is what they did. They basically took a model that would give an answer and then another model that would basically test or judge that answer as good or bad and then get it to reiterate. So it's doing the manual process of a person who is going to read an answer and be like, this isn't correct. Can you please, you know, this is a misunderstanding, or this appears to be a hallucination, this song doesn't exist, blah, blah, blah. And then it gets it to correct its answer and do stuff. And a lot of the time, probably a majority of the time, if you just keep giving details and you keep trying to eke the answer out from, like, kind of argue with the AI, the AI will actually. The LLM will actually give you the correct answer finally. And. And so they tried to manually do this with another LLM and it kind of has predictable results. [00:09:21] Sometimes it works really well because you're just iterating through things and it actually can be judged, and sometimes it can't be judged and it still just hallucinates stuff. It just takes longer and it costs more. Like, if it does a bad job on the first judgment in some way, it'll just keep running down that thread in the wrong direction for no reason. [00:09:42] Now, the reason why this is important is because this was supposed to be another huge thing that ChatGPT and OpenAI could do, and it was a bit of a nothing burger. Like, again, there was a lot of hype to show for certain things, but it's so hard to just use it for some stuff. [00:10:05] It still just has so a really big margin of error that makes these things a lot of really useful in certain contexts, but completely useless in others. And one of the things to keep in mind is that if you have a product that works 95% of the time, you don't have a product you can't sell, like, that cannot be used in a professional environment. [00:10:31] You know, have you ever heard of graphene? [00:10:34] Ever heard of, like, graphene, the technology and that? You know, it's the thinnest, strongest material because it's, you know, it's literally carbon atoms laid out as flat and as thin as possible. [00:10:46] And it has all of these incredible properties. In fact, it's a. It's like a superconductor and you. It can actually be used as a battery. And there was one thing that I actually saw sometime in the past that you can actually get graphene oxide in liquid form that is dissolved. And you could put it on a cd, on a disc, a writable disc, and then use a lightscribe thing where you can flip it over and it will burn an Image into the lightscribe into the top of the cd. If anybody, you know, back in the days in the 90s and early 2000s, you were burning DVDs or burning CDs and Lightscribe was so cool, is you could use the laser to burn an image into the disc itself so that you could draw pictures. Black and white, but cool nonetheless. Well, you could use that to. After you let the graphite oxide, and this is one of the main ways of creating graphene you use. You let the graphite oxide dry out on the disc, then you flip the disc over and you put it in and you just get it to burn as if it's just drawing black. Like, just get it to laser the entire surface. [00:11:59] And what it does is it takes that dried graphite oxide and hits it with the laser, and the laser flattens it out essentially and turns it into graphene. And this is really the big problem with graphene is how do you make it. It's nothing about the properties we could have. Let me tell you a little bit about the electrical properties of graphene is that if you could make it dense enough, you can make a battery out of this that essentially has no half life. Like, you know how a battery, you charge it, discharge it, charge it, discharge it, dart discharge it, and you can do this like a thousand times, 4,000 times, whatever it is, and then the battery just kind of sucks. You're only getting to like 80% battery, 70%, and then it just goes downhill. And, you know, you have to be kind of careful about like how much electricity you put in it, because you can make it go. If you plug it in too long and just leave it there, it will, you know, it will die quicker. And you have all of these things, the lithium ion batteries and all of these things are. [00:13:00] Have consequences. And they are not strong like long life batteries. [00:13:06] Graphene has incredible electrical properties. Not only is it insanely thin and strong so that you can make it not only tiny, but flexible, but it will literally store a charge without a half life, without degrading the stuff itself. It's not a chemical battery. And it moves electricity, conducts electricity better than copper does, and again, specifically does so at a tiny fraction of the weight that most of our main conductors require. And this would enable a battery that you could charge and discharge hundreds of thousands, millions of times, and it was still just charge and discharge the exact same amount of electricity. [00:13:54] You know, I think I just realized I said superconductor back there, but a superconductor, a room temperature superconductor is a different thing. I meant super capacitor, super capacitor. And that's the thing, is that you can get these little. You have capacitors and devices and stuff all the time, and they're able to. [00:14:12] They don't degrade. They're just. They're storing energy in a very simple and reproducible way. But universally, it's a tiny amount of energy. It's so hard to get a capacitor that can hold a large amount of energy. This is exactly why you. [00:14:31] You know, a lot of times you have to hold down a button for 10 seconds on a computer, or when you boot it down, a lot of the capacitors will literally do final processes and stuff to protect the computer, because you still have power in your computer after it loses power for a short amount of time, and there's like a bunch of emergency services and things that will continue to operate. And when you're holding down a button on a device, like, you know, you're resetting something with a little pin on the back, what you're actually doing is you're. You're specifically, specifically draining all the power out of the capacitors in the device so that any memory being held by those capacitors and is also wiped. So graphene is a super capacitor in that it is able to hold so much electricity as a capacitor that you could literally make, like, smartphone batteries out of it that last way, way longer, have more energy, never die, essentially. And here's the really big thing is that it doesn't really matter how much voltage or amperage you deliver to it or discharge from it in a short amount of time, because there's so little resistance. It doesn't have nearly the potential problem of, like, being overheated or any exploding or anything, because, again, it's a capacitor. It's not like a chemical battery. [00:15:56] Sounds great, right? Sounds awesome. Why don't we have them? I read about this for the first time, like, 15 years ago. [00:16:06] Where are all the super capacitor batteries? I mean, if you put something like this in an electric car, you could get rid of the weight. You could literally fill up. You could, quote, unquote, charge an electric car at the same rate that you could fill it up with gas. You could go to the gas station, plug the thing in for a minute, get a full charge, because they just dump, you know, 20 times the amperage that the car actually uses while you're driving around, and then you can keep driving and the batteries Won't be heavy. The batteries will be super light and they'll hold just as much charge. I mean, it sounds like a dream, right? It's the, it's the, it's the holy grail of battery technology. And we have graphene. We know graphene. What's the problem? [00:16:50] Oh, actually going back to the disk thing, you can actually take those strips. This was a, I kept thinking it's like a Hackaday project or something. You could take the CDs and then cut, cut out all of the, the graphene that you made on top of it and then layer them with a buffer and you could make your own battery. You can make your own super capacitor. And it straight up worked and it held a good bit of electricity. So again, what's the problem though? [00:17:17] Consistency. [00:17:19] Nobody knows how to cheaply make it so that it's actually reliable. If you do the little disk method, it's so inconsistent. You'll get a strip that gets almost none, right? And some that gets like a little bit, some that has like a conductivity air. Like, you know, you don't get good conductivity across it because one area just didn't burn, right? Or the, you know, it's just a dissolved graphene oxide or whatever. So it's not like evenly distributed. It'll just be thin in one area. [00:17:50] Like, this was something that I was super into for a while because I was like, where is this? Why are people not using this? [00:17:57] And it's just because we don't know how to mass manufacture this stuff. It's got like a 90% success rate. [00:18:04] And that 10% just makes it unworkable. [00:18:09] And the cost and not having a reliable way to actually manufacture it just means it doesn't exist. We just don't have it. Nobody uses it. There's nothing we can do. It's just going to take more breakthrough. We just have to wait until there are breakthroughs in the technology of creating, of making graphene that allows us to use it and we have no idea how long that's going to take. That might be an incredibly difficult problem. [00:18:34] So why did I go on this long tangent? [00:18:37] If you have something that a small period of the time produces awful results or causes you a problem that you then have to fix, you can't use, has to work practically every time after. When I was talking with the Keat guys, Math talked about this a lot, is that, you know, we would get the number up to like 90% of the time. It would establish a connection, everything would be great. And then Like a little bit of time it would be really hard or something wouldn't work. And we'd have to try to figure out why is this network working differently, et cetera. Then we get it up to 94%, then 95%, but that was unworkable. Nobody's going to use a piece of software where just like one out of every 20 times it just doesn't work and you don't know why. And so they spent years and years just getting it to 98, 99, and now it's like it's basically 98.99%. [00:19:27] It's as close as you can. There's not really a possibility of getting to 100% because you just don't know what the environment is. But you can get it close enough and they think they're close enough that, I mean, it just works. Every time I boot the thing up, it just works. But they talk about that exact thing, that if it just, if it doesn't just work, you cannot, you almost universally cannot put it in a business environment. [00:19:51] It's not going to replace an employee because you're literally going to have to have an employee watch it to judge what it's doing before it's allowed to make decisions. Because if one out of every 40 times it just completely makes something up that doesn't exist, that isn't real, it starts clicking on things and it goes down this whole process and starts hiring other agents or moving other agents in the company to do something, you're going to have to unravel all sorts of crap. And the whole point of using an LLM in that sort of an environment is, is to remove the need for people to babysit it because otherwise just pay the people. [00:20:30] Hallucinations are still there. And it's a huge problem because it's just a probability machine. [00:20:37] It is where we plug these things in that are important, is how we use them as how you make people able to use them and then able to check to ensure that the answer, the result that is given and the way that it is used, how to test and assess how good it is, and then of course, make the final judgment by someone who can do a legitimate analysis and has good judgment. But an LLM specifically can't do that because it's just a probability system without an example, without something in its model to be trained on, it doesn't actually have judgment. It's just giving you the highest point, probable judgment from the situation, which means every single edge case is just going to have terrible performance. [00:21:30] Now, maybe There's a combination of tools, of databases. I mean, like all of this is just going to continue to go further and further towards how to make, how to make it more useful, how to tie it to real information, how to stop those hallucinations from being significant problems, etc. Etc. [00:21:48] That is the process of technology, obviously, but this is also really fundamental to what LLMs are. [00:21:55] And it has been sold as intelligence as if it's real intelligence and it is not. [00:22:02] It's mimicry and translation. [00:22:05] So the OpenAI vision, the idea that AI is going to take over everything and replace Everybody's jobs and OpenAI is going to be the universal general tool that can do all, all these things and anything that you need it to, has one very significant flaw. [00:22:22] And it's not anything to do. And it's specifically hard to see because it is being sold as if it is genuine intelligence. It literally says when it has two LLMs fighting with each other, that it's thinking and it's not. It is not reasoning things out. It is using two differently trained models to attempt to get a higher quality answer. But it still requires both of those models to not hallucinate and both of those models still inevitably will, because that's how the models work. [00:23:01] But OpenAI is selling, is essentially trying to sell a story that I don't think it can ever deliver on. [00:23:10] And going back to one of the things I've said in one of the recent episodes is that I don't even care about better models now. What I use them for is perfectly acceptable with the quality of models that we have now. I don't even need. I'm not even really waiting for a better LLM. If you know, Llama 3.2 or Llama 4 or whatever comes out and it's like way better and it's a much smaller model, that's cool, but it's only cool because it means that I can use less computer. It just means that I might be able to run it entirely locally with like a 22 billion parameter model or something, rather than using the 405 with Venice AI or using one that I can't run with Unleashed Chat. All it does is shift to where it is that I'm using it. But I don't need it. [00:23:59] LLAMA does the overwhelming majority of what I need it to do and what is really being sold is the actual resource, the actual product is compute. [00:24:11] It's just compute. [00:24:13] They're just selling GPU inference. [00:24:16] And I think another really good example of this is Apple Intelligence. [00:24:21] So we talked about this one on the show when they announced it. And now the newest MacBook Pro has Apple Intelligence. There's a bunch of local models that do a whole bunch of things I talked about, like how I can contextually search in my images, which is so cool. [00:24:37] In fact, just the other day, I mean, excuse me, just this morning, I was looking for a. I would say I had saved a bunch of memes and, you know, data and charts and all of these things. And I was trying to remember a chart about seed oil, seed oils and linoleic acid and you know, what degree of toxicity, like, how terrible are different types of fats and oils and all of these different things. And there's this big list that listed them from worst to best. And I needed it to reference something. And I was like, dang it, where was that? [00:25:12] And then I just remembered that on the page, on the image itself was linoleic acid. It's just that word. So I literally just typed in acid in the search results. Boom. It was a top result. I have like 10,000 images, actually probably like 15,000 images on my phone. I have so many images and I just search something by a word in the meme. [00:25:36] I guess in this case the chart. I do that a lot. That is a useful AI. But here's the thing. I'm not paying extra for that service. [00:25:47] And if Apple tried to charge me $30 extra for that service, I wouldn't pay for it. I would just be like, why do you want. [00:25:56] Why do you want your search to suck? Because somebody else is gonna give this to me with a 2 billion parameter model that I can just download and do myself. Because, hey, I got LLMs that I can just code some of this stuff up myself. And if I get annoyed and have just a little bit of patience, I can probably produce something just like this. Now, of course, Apple's implementation is going to be a lot nicer and a lot more integrated. Cool. [00:26:21] But you know what they're not doing? They're not charging for any of this. [00:26:25] They're just making the product that they sell a little bit more enticing to take, Take market away from competitors. So here's the thing. There's not a new market with AI. I don't think. I don't think this brings in a whole bunch of new people. I think it just becomes the differentiator of who can implement it the best to make their software or products, their. Their devices that much more useful. And that's why I think Apple is selling Apple Intelligence like an Operating system. It's just part of the experience of Apple now. And it's all free. [00:27:04] You just, just getting it with the software. [00:27:08] One of the first ideas, somewhere in the first ideas that I covered on this show, I talked about AI as an interface. [00:27:18] And the interface just becomes the norm. It just becomes invisible. It just becomes the default of. Of course you use this. Of course search is contextual. You know, you don't pay for search. I have Search in my finder, I have Search in my browser, I have Search in my notes and text edit, but I don't pay for that. I don't, you know, get a text edit that doesn't have a search. And then I pay $10 extra for the text edit that does have search. Search is just expected to be in every single one of those tools. And one of the best things about LLMs is just better search, like drastically better search, especially when it comes to local things. But when somebody just integrates that and just makes their app that much better because of how good their search is, everybody else is just gonna have to do it and it's just gonna become the normal thing. But nobody's gonna be paying $100 a month extra to have better search, or at least the market for very specific tools that are only available in that context and that are completely just LLM based, which search isn't, right? Like, it's just that all you need is just like context awareness for an LLM to other information. But I mean like a strict, I'm talking to an LLM and it is giving me something back and it's useful and it is so good and so reliable that I can use it in a business environment. [00:28:48] I'm saying that that degree of product, the, that scope of product in the AI space is just extremely limited. It's just way more limited than it seems. [00:29:01] And when you sell it as artificial intelligence, it just seems like you're going to be able to use it for anything and everything and there's no limit to anything. But it's actually very, very targeted what you can do with some of these things. And that's why I also have a huge interest and focus specifically on video generation and animation and all of these things. Because I think this is a very targeted and very useful tool for a lot of people. It's just a gimmick. [00:29:28] People don't need to, you know, take portraits of a face or a monster or anything and animate it to say stuff. Unless you're making a meme or you're a filmmaker or Maybe if you're a content creator on Instagram and somehow you can integrate it or whatnot. But usually it's just you, you just want to, you just film yourself. Like social is about talking to and meeting other people or interacting with other people. [00:29:53] But I think these are very targeted tools for very specific use cases and very specific industries or creators. [00:30:03] But the product itself is just the software, it's just how you use it. It's the same as it's always been. The models are just the part of the software that let you do more with it. You know, you're still just getting Adobe Firefly, you're still just getting a service that lets you edit faces or animate a tool. It's the same products, it's just a different tool set inside those products. [00:30:29] Just like search is. Sure, you have search and text edit. And if you add an LLM, a 2 billion parameter LLM that you know is super lightweight and can do a much better search where you're like, can you find me everything in this work that has a negative tone, where they kind of say they're getting really doomy, it will literally just pull that up. It'll be like, okay, well this is a specific area that sounds really negative. [00:31:02] That's one of the really cool things that an LLM can do. In fact, I used it for my one small Catch for Mankind video, which is just, it's nothing big or important. I just made a, essentially a compilation video and a little like short from start to finish of the rocket launch of the heavy booster. And then the catch, kind of the video version of what Andrew McCarthy did with the photos had like gorgeous photos. [00:31:31] You should check that one out. That one's a like a James McCarthy, I believe. I'll try to remember to get the, the link so you can see the image. But really cool. The guy's a really cool photographer. Takes pictures of space and planets and the moon and all sorts of stuff. And he took a photo series of the rocket going up and then coming back down and catch being caught at the tower. And it's just really cool. I actually used some of his footage in the video. [00:32:02] But yeah, so I made that video. But one of the things that was interesting about how I used AI was was that I took the entire hour and 40 minute SpaceX video, ripped it, pulled the audio from it, ran it into my Whisper, my transcription app that I built for myself. [00:32:19] Drag and drop, so get to the transcript and then took that text file and dropped it into llama 3.1 and said can you read this and find all of the places where they talk about being unsure as to whether or not this is going to succeed or where things start going really well or there's good news. I basically asked it for a bunch of context for what I wanted to do in the video. So like I wanted to start it off in the video of, you know, they talk about like, not sure if. We don't know if this is gonna work. This is a guess. This is our first try. We don't even know if we're going to try to catch it at the tower because things might not line up. [00:33:00] So I got it to find me a couple of quotes with that and then give me the time codes. And so I searched those on the subtitle and then I could easily cut that right out of the audio. [00:33:08] And for an hour and 40 minute thing, I didn't have to go, listen, this is what I used to do. I've had to do the editing is a tedious job. I have many, many times I have had to go through and listen through something the entire time and then make markers like manually write down or put in a note of markers of where audio clips are that I was like, oh, this is good. Okay, this will be useful. And it's really easy even to miss stuff or to not think about something until later. [00:33:36] And you know, you remember some sort of a line and then when you're making the video itself you're like, oh man, that one would actually been really useful right here. [00:33:46] But I don't remember where it is. You have to go back and you have to listen through all over again or listen through a bunch of it. Just trying to find that spot, scrubbing through, just find that one line that you were looking for. But with a very, very quick transcription AI and then a very quick LLM search, it can actually pull out 20 useful clips, tell me where they are. And then I literally just go through and I clip those things. Then I basically have a set a selection of things to build suspense at the beginning and then make it feel like a payoff at the end of the video. And I can make this tiny little arc that occurs in the video and hopefully make it a little bit better and more interesting. And all that does is let me do the video easier, faster and a little bit better. [00:34:31] And that's just going to be the norm. I'm going to expect to be able to do things like that. And when it doesn't work, I'm just going to be like, I will never buy your product or I will not use your service because that's ridiculous. [00:34:43] This is just going to be something we expect to have. And because of that, the companies that provide it are going to want to be incredibly efficient about providing it. They're going to want the thing that gets the exact tool, get the, it's the exact use case out of it for as little as possible and with as little variance as possible that it doesn't do a bunch of other things like textual search is or contextual. Contextual search is a perfect example of a great place that an LLM is used in which you don't like, like, you know, hallucination doesn't matter because search results are approximate by their design. You're just looking for something specific to use. So you're always, you always just have results that just aren't perfect because you're hunting. [00:35:35] So they're not going to use a 500 billion parameter model to do contextual search. They're going to use a 1 billion parameter model because they don't want to be paying a nickel or a dime every single time you search for something. It goes back to one of the things that we talked about, kind of the fundamental elements that we, or fundamental truths that we've talked about previously on this show, and I think we really talked about this with Dhruv Bansal quite a bit, is that general intelligence is incredibly inefficient. [00:36:11] You do not need it for the overwhelming majority of things. And this is what I think OpenAI is focused on and why the focus might be a really, really bad decision. [00:36:23] Now they also raised a funding round at a valuation of about $150 billion. They raised 1 of the craziest funding rounds. I think it was the highest. It was like $6 billion or something like that. [00:36:39] But the thing that the author talked about in the article, the Subprime AI Crisis. [00:36:46] This is like the weirdest corporate structure ever. [00:36:50] It's not equity like VCs are. Usually you buy equity so that if the company's not profitable but it still has like a huge bunch of customers and stuff like that, you just sell. [00:37:04] You just sell the equity for, you know, partial loss or whatnot. But it's roughly worth what the supposed company is worth. [00:37:14] That is not what they are selling. They're selling something called a ppu. [00:37:21] This is a profit participation unit. [00:37:26] Important caveat. This company doesn't make a profit. This has been the weirdest corporate structure process that they started out as. We're going to be open source and we're going to be open AI for everybody and we're going to bring this all people. And then they went super closed source and proprietary and they were like we're going to, it's for safety, we're not going to share this with anybody. Everybody's going to be terrible with it. So it's really important that we close it and charge for it. But we're a non profit and we're, you know, we're still for open AI and we will release these probably in the future. [00:38:03] But then all of a sudden now we're for profit and but we're not going to sell stock and actual equity. We're just going to completely own the company ourselves. But we're going to create this new like corporate structure equity thing that we're going to call a profit participation unit and it's going to be worth absolutely nothing at all unless we are profitable. And you will get some of that profit based on how many ppus we have or you own. [00:38:32] Because we're so big and we're so important and we're at the apex of all AI, obviously we're going to make a massive amount of profit even though we're not making profit right now. [00:38:43] So if they again they raised like $6 billion selling these and if they do not make a profit, they're worth nothing. [00:38:55] How likely are they to make a profit? [00:38:58] Well apparently they're paying Microsoft an estimated $4 billion in 2024 to power ChatGPT. And importantly this is a highly, highly discounted rate. Microsoft is subsidizing the crap out of the company. [00:39:17] And then he cited another article that basically put together and kind of bragged about the fact that Whoa, they have $2 billion in revenue as if this was a really good thing. [00:39:31] Now understand the 4 billion dollar cost to run ChatGPT is just the electricity, this is just the GPUs and again this is discounted which means the real rate is probably something like six or seven. [00:39:46] And importantly the models are getting bigger and more expensive to run without a seemingly corresponding huge jump in improvement. They are better, at least in my opinion. What I think is happening is that they are, they are better but they are not proportionally better in relation to the cost. [00:40:08] And my example for that, for the use case specifically that I have used a bunch of LLMs for is that when I do use, when I have used 01 and or chatgpt 4 and 4o I don't really care. [00:40:25] Like they all just kind of produce the same stuff. [00:40:28] And sure again one might be a Little bit better than another, but I really can't tell the difference. This is why I've ended up on llama 3.1. And I've used the analogy. I know I've used the analogy previously on this show, but I think it continues to hold up in my experience is that it's like computer graphics in games. At some point you just don't care if it's good enough to do what you need it to do. That's all you need. Then the question is just how do we make it smaller? How do we make it so I can run this. How do we put it in a 70 billion parameter model? What about a 22 billion parameter model to get the best result, to get the same result. And that's what the Nematron 70 billion is actually better at code with then the 405 billion parameter llama 3.1. So they're already like making aggressive new like kind of forks and loras to these models that do some specific task better with a much smaller inference cost. And a lot of these companies are trying to get lean, they're trying to cut down their cost. [00:41:39] This is where I think this is going. And I think it's going to be pushing all away from ChatGPT or OpenAI because OpenAI is just trying to get bigger and bigger and better and better. They're literally running towards artificial general intelligence. And I think what that means is that they're running towards something that doesn't really have a product and cost as much as it could possibly cost with no apparent end in sight for it costing more. [00:42:06] And I think they think that at the end of that tunnel is this super artificial general intelligence thing that can do anything that they want and it's perfect and it's better than every human and there's nothing you can just insert it into everything. And now it's the center of all the corporate and business world and it will just run society for us. And I think that is not possible. [00:42:28] That's not what any of this actually is. [00:42:32] And even after Ashenbrenner's piece, the more and more I think about it, the more I think we're hitting an S curve is that it just gets to a point where it's not worth that much more to keep making them bigger and better. Now I will totally eat my words if I am wrong on this. And it's a pretty short timeline for what Aschenbrenner was suggesting. That we're looking at like 2027, 2028 as possible, you know, super intelligence and AGI. And there's actually logic to a lot of what he talks about. [00:43:05] But I think his, I think he's wrong on the economics of it. [00:43:09] That may be where there is a disconnect in my opinion, at least in my framing what I think he might be wrong about in the context of his argument. But if it's going to be 2027, 2028, well, then this will still happen. This will happen one way or the other while Trump is in office. [00:43:25] So we're right around the corner. So I will still be here talking about AI one way or another when we get to that point. So we'll be able to assess in hindsight exactly how wrong or why. [00:43:37] Why was it that that didn't hold up over the time span or of course we've reached the birth of AGI and everybody's just doing nothing at all. And you know, maybe it destroyed the whole world. Either way, you should subscribe and follow because we'll be talking about it here with Guy Swan, your front row seat commentary on the end of the world. [00:43:56] I know you like that. That's going to be a commercial. I'm going to make a commercial for this. But yeah, so OpenAI, I think has, I think they're moving in the wrong direction, at least they seem to be now. [00:44:08] And importantly, their costs are massive. [00:44:12] We don't know what their costs really are, but it literally could be five, six billion dollars. [00:44:18] And their revenue appears to be evidence suggests that their revenue is 2 billion. [00:44:23] Which means that raising $6 billion in a super weird and questionable funding round that only gets you any sort of money at all. If they make a profit, it literally only buys them a very short Runway. [00:44:44] They still run out of money like super quickly. And Microsoft is probably losing money subsidizing them with the low rate compute. [00:44:56] If OpenAI goes under, there will be a huge bear market in AI. [00:45:04] This will cascade through the entire industry. [00:45:07] There will be layoffs. There will be every company that stuck AI in its name or in their product. Just like every company that stuck blockchain in their products is going to either have to roll it back or just kind of memory hole that whole period and just be like, oh no, we just, it's just like we have a really cool service. This is great. [00:45:28] And then a lot of the big players will pull back heavily on their investment and reframe their strategy. They will. AI won't go anywhere because AI is incredibly useful. [00:45:41] But I think they will shift how they are thinking about it and where they are using it. And we will end up in a place where AI and machine learning are all over the place. But they're all small, they are extremely targeted, they are very, very tool specific and it just kind of becomes a new layer and type of software. That's kind of what I think this is. It's just a new type of software. [00:46:09] So that's why I've been saying for a little bit I think we will probably get the opportunity of really cheap GPUs maybe within the next year. [00:46:18] So I have been. And Nvidia might respond to it as well because Nvidia has been just skyrocketing in value and you know, sell shovels when there's a gold rush. I don't think Nvidia will get hit as bad. Well, unless they are as overvalued, they could be extremely overvalued. So they could just get hit terrible in that situation just because of that. But also when I say that COMPUTE is the thing that is actually being sold, while they will take a big hit, they won't be going anywhere in my opinion. And probably one of the reasons why we'll get really good models to run locally and run smaller also is very likely because we have quote unquote intelligence chips. We have ASICs essentially for inference and for certain types of AI models. So we'll see about that one. That one's, that one's up in the air like it will take a huge hit if the AI industry implodes in the short term, but who knows what happens on a long term timeframe. [00:47:27] But anyway, I think that, I think there's going to be a reckoning in AI. I think he's got a pretty strong argument and I do not think that AI, like obviously we talked about very much on this show about how useful AI is, but I think it's useful in a very different way. And I don't think we're increasingly, I do not think that we are headed to big, giant, all encompassing models because I think the economics of them don't make sense. And we are headed more and more towards small, lean, very task specific models. And I think, and I think that means open source is going to work, is going to benefit very, very greatly from this and that this will largely exist as integration into software and into systems and operating systems and into devices such that with or without an Internet connection, what that LLM does or what that model does is simply available to you. And I don't think this is just Hopium like I would love. Obviously that's like my dream is AI is everywhere and it's all small and lightweight models and it's open source and you can run it on any devices. And obviously it's not going to be all, all that. There will be a huge range and variety depending on what the task is and how much compute is needed. And you will have to offload compute to other people and to other devices. [00:48:56] But nonetheless, I think this is, this at least seems to me now the way that these things are headed. So we'll stay tuned. We will stay tuned. I'll keep an eye out on it. Check out xportrait and Live Portrait, which is what I was talking about earlier. If you want to dig into the animation and like facial expression and feeling translation and stuff, really, really great tools. They seem like for animating, for, for storytelling and obviously that's something that I'm very, very interested in and we will be covering in the future and I think I'll probably do another episode pretty soon. [00:49:35] This will probably be YouTube based because I'll just be going over some tools that I use, some new ones that I found recently and started to integrate and how I use them and I think that would be, I suspect that that would be really helpful and a lot of people like really enjoyed my episode on Pinocchio and talking about, you know, how I get a lot of value out of these tools. So yep, that's what we will do. Don't forget to subscribe. Don't forget to check out my Nostr and Twitter or X whatever you want to call it. [00:50:03] I am the Guy Swan up there. [00:50:06] I have some great affiliates and services and products and stuff that I use and I really, really like with some discount codes right there in the show notes. Check those out. And of course subscribe to Bitcoin Audible and I will catch you on the next episode. Until then, everybody take it easy guys.