Episode Transcript
[00:00:01] There have been a number of major breakthroughs, announcements and gimmicks and new features of AI in just the last week or two with people finally getting their hands on the R1, the rabbit, with Google IO making tons of new announcements, with Gemini Pro and their new video gen model, their Ask photos and Ask documents and and then OpenAI making a huge announcement with GPT4O, a free tool with GPT4 capabilities where the main mode of interaction is to simply speak to it and it speaks back. But while it's all impressive, how much of it is actually useful, how do we separate the bullshit from the breakthroughs? And that's what we're going to do in today's episode. I am Guy Swan and this is AI Unchained.
[00:01:11] What is up guys? Welcome back to AI Unchained where we separate the signal from the mountain of noise and explore how to use AI as a tool for sovereignty rather than as another tool for centralized platforms to control and surveil us. That is what we are doing here. I am Guy Swan, your host and we are getting into a lot of updates from the past couple of weeks. A quick thank you to Swan Bitcoin and to Coincite, the makers of the cold Card Hardware Wallet for sponsoring this show and making this possible. Swan Bitcoin is the best place if you're just looking for an easy way, no nonsense, to buy Bitcoin or to set it up in your retirement to get it connected to your business to set up a long term savings plan. Swan Bitcoin is the no nonsense place to do it. Please don't go to places like Coinbase or some exchange that's going to encourage you to trade or give you a bunch of shitcoins. Just go to swan bitcoin.com guy and you will not be led astray. They will teach you how to hold your own coin, hold your own keys be and how to think about why you are investing in Bitcoin. Their knowledge base is massive. And then you can get a discount on your cold card which is how you are going to know you are holding your own keys and you are doing so securely. Keep the keys off the Internet. Keep them off of your vulnerable devices and off of malicious software. Use a signing device like a cold card, use a tap signer the many different hardware security devices at CoinKite. The discount code is the title of my other podcast Bitcoin Audible all one word. The links and details will be in the show notes. So in the past week there have been just so you know also I'M digging a bit into quantization which I do want. I had a number of you guys actually reach out after the episode where I brought that up and whether or not you wanted me to go deep into it and said that I was probably going to try to go down that rabbit hole a little bit. Um, but I think we'll probably do an episode on quantization. What it means, what the trade offs are the different methodologies apparently for it and but it will take a little bit for me to kind of wrap my head around it and make sense of it well enough to actually explain it to somebody else without just a horrific amount of ridiculous vagaries and completely inaccurate analogies. So I will do my best. But we will be getting into that in a future episode and I have kind of started my journey down that rabbit hole. But basically we have seen we're finally starting to see a lot of very direct multimodality implementations like how to use the applications of this. And a lot of them seem to be all trying to do the same thing. And there are a couple of areas where it's immensely valuable and a couple of areas where I just feel like it's so gimmicky or I just don't think it really has the application that a lot of these people like Google and their. Their Google I O that just happened.
[00:04:23] When was their recent Google IO? Actually I don't know. I just watched the video the other day but it's very recently so Google has a ton of new updates with Gemini and the rabbit R1 which I actually pre ordered which I wanted to share on this show apparently is out for some people but not for all, which I guess is me.
[00:04:48] And it appears to be a deeply unfinished product. There's a great one by the that guy, the black guy. What's his name? Like he's like king of reviews on YouTube. Oh my God, I can't remember his name. I'll have the link in the show notes. Fantastic review. But he basically says the the title of the of video is the R1 is unreviewable or something like that. It can barely be reviewed because it's just not a finished product.
[00:05:17] And he brings up a lot of really great points. But it's funny is that a lot of what he explores with the R1 which again is its own unique device and attempt to implement AI And I didn't have. It's funny, I didn't have super high expectations for it, but I did want it for one very explicit use case. The idea of, you know, implementing or integrating with a calendar for being able to jot down notes and like have thoughts in an incredibly quick fashion. Because I would do it a lot more with my phone if there was a quick interface for it, but there isn't. Like, it's always like nine steps to jot down a good note. And it does not work very well in voice. And the transcription sort of stuff, the speech to text type stuff is so explicit that like I want an LLM that can then fix all of this. The ums and the me trying to develop my thoughts and not really, you know, even the.
[00:06:19] I mean, this is a great example. There's, there's not a whole lot of cohesion. There's not like good written, written cohesion in the way I'm talking right now. It's very conversational. I'm just talking to you guys. I'm just talking generically to the computer. And none of this translates very well to writing. But the beauty of an LLM, of actually taking that next level and pulling kind of the semantic context out of it, is that it can separate. That it can separate what I mean from what I'm.
[00:06:49] From the way that I'm actually, that it's spilling out of my mouth, so to speak. And so it can take that idea and actually write it down. And hilariously, that is the. One of the main functions that the R1 can't do apparently is you can't integrate with a calendar. You can't put like meeting notes and like a task or something on a to do list. And you can't like write notes. And I'm just like, that is, that's the one thing. That's the one reason I was willing to spend $200, like I didn't give a shit where it. Whether it could look into my refrigerator and give me a recipe for something to eat. And that's one of the things it does in the video, which even when I watched it, it's one of those things where it. You see, you see it do that and they show the refrigerator. And the refrigerator is the most immaculate. Like everything is right at the front edge of the shelf.
[00:07:48] And, and it was, it's very clear. Like you could look in the refrigerator and you could just from 20ft away and just be like, okay, well, yeah, these are the things that I could make. Or, you know, here's all the obvious food items and the, the attempts to get things to interact with the real world and get these LLMs to be able to see, I think have largely been a dud. They've largely been really gimmicky. The R1 is a great example.
[00:08:21] What's his face again? I can't remember his name. I'm not like, I don't hugely follow him. So I know a lot of people are going like, it's this guy. It's good, you idiot. It's this guy. So I've only actually reached, like, I know him and I've watched a couple of reviews in the past, but it's like one of those things that's like so few and far between. I know he's YouTube famous for reviews. I'm sorry for not knowing his name, but I just don't. And I'm not, I'm too busy. I'm not going to look it up right now because I think it's better, I think it's funnier if I just don't know it for this whole episode. But another area where both the R1 and this one, they specifically did in the Google IO, which I do not understand.
[00:08:58] Maybe, maybe I'm crazy, maybe I'm just out of touch. But why on earth does anyone want an AI to plan a vacation for them?
[00:09:08] I just, I don't. I genuinely don't understand that. Like, sure, if you want it to just give you a search and, or help you search for something and find all the options, like if you wanted to go water, right, white water rafting, or you were going to the mountains and you just asked, like, what are some good things to do? I search that sort of thing on Google, on search results and on Apple Maps or on a community website or something all the time. I'll ask it on Reddit or something.
[00:09:38] And if there was context for that sort of thing, that would be fantastic. But that's not asking an AI to plan a vacation for me. It's not putting it on my itinerary, assuming that I'm going to want to go write whitewater rafting. But if I was going to the mountains and I was just like, what the heck is there to do in Boone, North Carolina? And then it had whitewater rafting on a list, I'd be like, oh, snap, I might actually really like to do that.
[00:10:07] And that's also another thing where a lot of people seem to want to target AI as the tool to do this. But this is where I think the social graph is actually way more important. That this is a social element that is untapped rather than an artificial intelligence or machine learning element that is untapped. The LLM doesn't have value, doesn't have judgments, doesn't have preferences and have any idea what is good or bad, that is what is important for making those sorts of decisions or making those sorts of recommendations. And it is also not necessarily in the context of what you like. You can't just look through my history of stuff and know what I might be interested in. But you could look at my friends on Nostr. This is where Satslantis comes in, which we just did a episode very recently on Bitcoin Audible. For anybody who doesn't know it, Bitcoin Audible is my other show. It's my main show.
[00:11:05] And we did a episode called Reviewing Reviews, which kind of seemed like a boring title, but it's actually a fascinating article and discussion, like genuinely fascinating way to kind of remold reshape the structure of how we wait and think of information connections on the Internet and how an open protocol for identity and social graphs fundamentally changes that relationship. And I think there is a massive potential in that. And it's referring specifically to a project called Satslantis, but I highly encourage you to listen to that one and check that episode out. If you if you haven't, I'll have the link to it in the show notes. But it's a really unique idea and I think there's way more depth in finding something that you are not aware of that you would be you would really enjoy through your social graph, through a friend's recommendation, through somebody that you trust, who you have a lot in common with, who you share values with, than in just like training an LLM to kind of predict what you'll like. Because a language model is just going to take your history and if you liked going white water rafting at some point is going to be like you might like white water rafting and Boone, North Carolina or something. And like that's great that it knows that you liked that. But that doesn't mean it's going to know the other things that you might like that you've never done before. But actually reaching out to your social graph and to people who are contextually relevant to you in the context in the sense of reviews or experiences, who are your same demographic, your same age at the same point in their Life, and where LLMs might actually be useful in gathering and searching through that data because, you know, a generic semantic search. But it's not something that the LLM is going to have the data for in the sense that the LLM isn't going to just conjure up those decisions you need to, it's not a Google search problem. Right. What you need is, is like Google Search can obviously give you better search by giving you semantic search through an LLM and embedding all of the information into vectors and all of that stuff. But it doesn't fundamentally change the weights of Google Search. Google search is still going to just pick things based on where you know how many links there are to it and you know, how many times it's mentioned on this website and what its Google Rank is and all of that stuff which isn't relevant to you for a hundred, for 100 different reasons. And one of the examples they gave, Svetsky, the author and everybody who's the team that's going into satslantis or building satslantis is the example they gave is that you know, you go to Amazon to review a product or something or read product reviews and you're probably reading, you might read the top review that says oh this is fantastic. But it's like a, you know, 40 year old soccer mom who, you know, rides a minivan and you know, takes a bunch of kids to and fro like in, in her life, in her context, something might be very valuable or might fit perfectly into her world or her values or even her political preferences or something. Whereas as a 20 year old dude in college or something, it's just not even closely relevant. In fact, might be the polar opposite. Like her great review might be his awful review and how many times, especially as we get more and more politically divisive and we've kind of isolated into our, our pools, our communities or whatever is not being able to leverage that. Like a great reason, a great example is, you know, any sort of politically divisive or politically lecturing sort of movie or media content where because something is too, quote unquote extreme white right wing, it will, they'll come rushing in and give it a whole bunch of down votes and then the same thing to the opposite. This is too woke and a whole bunch of MAGA people will just rush in and give it a million one one star votes and then nobody has any idea whether or not it's good or bad. And where a liberal might love that video and you know, I totally identify with all the woke mess in it. They're looking at something that's got a one star review because they're not being. The reviews aren't relevant to their demographic. The reviews are just some aggregate of everybody's general opinion which includes a whole bunch of people who are politically motivated over There to just give it a bunch of bad reviews and then vice versa. For the Republican, like that's not useful. In fact, what it does is it allows opposing demographics to poison the well for a different group. And an LLM isn't going to be able to sort through that. In fact, the LLMs are largely being trained to be woke or to be anti this or that and you know, tiptoe around, make sure they don't offend anybody. And that is literally not what we need. What we need are relevant reviews. We need relevant and value based and social graphics, weighted reviews, reflections, ratings, experiences. And if the information is actually aggregated through your social graph, every single person that goes to every single product, service or movie or anything will see a completely unique review.
[00:16:56] That is game changing. That is something that will fundamentally change the nature of how we interact with information and how that information responds to or informs our life, our lifestyles, our choices, what we do when we go on vacation, that sort of thing. Whereas an LLM just kind of like arbitrarily searching Google and just being like, yeah, here's all the things you should do in your vacation and just like handing it over to them just seems like I cannot, I just don't know why I would ever do that and I don't know why anybody would do that. This seems like the opposite of the whole reason to go on vacation. Like you're exploring.
[00:17:41] You don't let a computer just guide you when you're exploring. The whole point of exploring is that you're doing it. So anyway, I use that as an example because that was like a prominent thing in Google. The Google IE thing is like look, I can just change the details in my, I mean it's cool that it can actually alter, alter things and change schedules and stuff based on when you want to wake up or based on anything. Like you can move stuff around. Like there is a benefit there but the fact that they're using it, that their example is explicitly just plan me a vacation to Cancun or something and put all this stuff on my calendar and that people are just going to do it. Maybe we are that just lazy or we want to offload our responsibility for everything to an LLM. But I don't know that that's not me. I, I see that as like, I think that's a silly thing. I, I don't, I don't see that as being highly used. Now the, the other big announcement, the bigger announcement I think is chat GPT4.0 and this is something that we're seeing with all language models now when I say that everything's becoming multimodal is everything. All the major LLMs now have vision.
[00:18:59] All of them have integrations directly with data, with embedding and vectorizing data immediately. Like Google Chat or Chat Doc. I can't remember exactly what they called it. They have names for everything. There's Google Gems and all this stuff. Some of the naming conventions around all this stuff drive me crazy. But they, they have basically a chat with documents sort of thing. And you can upload all of these documents, like related to your project and your to do list and all of these things. And they'll have like a sidebar with all of these things that you have. And then you have a team interact like a chat window with your team. And you can also specifically ask Gemini, which is inside all of this, to look through all of that data and give you assessments, give you summaries, you know, how are we doing on this project? Are we on schedule? Where do we need to focus things? Can you pull this information from this PDF or something? And there's a lot of incredibly useful things about that. This is kind of a huge part of the dream, right, is to be able to ask questions to your data, to be able to organize and go through all of this stuff. The problem here is the insane amount of privacy concerns. Like, just unbelievable that all of the things about Google having all of your emails and just being able to open. Like, you realize that every email that you get, Google can just read the NSA and FBI and CIA who just have, basically they probably just have a. A login. They could just read everybody's email. Like, none of that is encrypted. None of that is hidden from Google or anybody that Google partners with. Now imagine someone on the other side of this equation, somebody on the other side of this server asking who is thinking of voting for Trump this year instead of Biden. Can you, can you give me the list of everybody's emails and documents and business arrangements and everything that suggests they might be changing their vote? Who is planning a protest against forced vaccinations? Is there any leaked data or conversations about information that is sensitive to the major pharmaceutical companies or maybe to safety regulations and cutting corners at Boeing, who you know is very important to the government and receives more subsidies and government contracts than literally any other business in the world? Can you find me those people and can you collect all of the emails and discussions in which there may be some hint of this sort of conversation going on?
[00:21:53] The danger, the risk of having all of this, of having your life entirely on Google servers and being vectorized and analyzed by everything on a giant Google supercomputer. And the type of context that people have to start just arbitrarily shutting down bank accounts or preemptively like this, this in no, in absolutely no uncertain terms or vagaries, just 100% will lead to pre crime sorts of enforcement. And it will absolutely be claimed that this type of surveillance and control will be necessary. That otherwise we cannot keep anybody safe, which they can't keep anybody safe. And they don't. They put more people in danger 100%. Their control, their authority. It is tyranny and they are the big risk. But they do not care. They want the control, they want the surveillance. They want to see into everybody's life with a fine tuned comb and fine tooth comb and that this is exactly the sort of thing that will give them that visibility. The cost of convenience will be enormous. Honestly, that is not a future that I want to see. But centralized. These giant centralized institutions remain just, I mean they are barreling ahead as fast as they can to implement these tools in as admittedly beneficial like admittedly trying to make tools that work. But I don't think. But this is a, this is not a do no evil situation. This is a can't be evil. If they can be evil, they will be and they clearly can be. This is created so much centralized and such an enormous source of information and power. There is just an unincalculable amount of evil that can be done. And it is and will continue to be done worse and worse until we get up and do something about it. And that doesn't mean that I have some grand plan or some simple and obvious execution. It literally just means that we need to piece by piece build these tools ourselves and figure out how to use them in a way. Because I think a lot of these, a lot of the basic tools, again like I said, a lot of the things that they mentioned are really kind of gimmicky. There are a handful of things that are enormously useful and then there are a number of things on top of it that they all brag about being able to do that just aren't that. I just don't think even matter that much. I think using the satslantis example, I think the weight, the ability to wait and source and value and understand the relevance of information is actually more powerful than all of this. Like needing a better LLM or needing a huge LLM. A great example actually is one of the examples they gave and a lot of examples that they give, actually.
[00:25:01] And we'll go into it in just a second with chat GPT4O because this is another really big one. And I think ChatGPT 4.0 is free or is going to be free.
[00:25:14] And I think this is an overwhelmingly obvious indication that, that we are the product. They want our information, they want our feedback. They want to collect every piece of data from every single one of us that they can so that they have the best models and they will sell it to, to not only the highest bidder, but everybody who puts a bid on the table. They will sell it to everyone. And as anybody who has not had blinders on for the last decade or so, it is very clear that the surveillance apparatus of the United States government and of governments at large has been outsourced. They just buy the data from private companies.
[00:26:01] Every major tech corporation does business with a government agency. They just, they all do. And there's even a we. It's crazy to me that this is not discussed more, but there is literally a secret court, the United States has a Stasi court that is completely secret that we are not allowed to know what is going on in which they prosecute giant corporations and make judgments and force them to do things which we are. Nobody is publicly available, nobody is able to know. They get none of the public defense for it. And these FISA courts, they're just allowed to happen. This is, this is just kind of like, oh, this is just how things work now. So anyway, we'll come back to the whole chat GPT4O thing, but one of the examples that they gave in talking to the documents and to do lists and team, you know, papers and everything that, you know, is being done with a project is they asked for it to give, you know, a summary and details on how far along they are and reference this, that and the other.
[00:27:07] And all I could think is how many times I've gotten a crappy response from the biggest, most advanced models, right from ChatGPT4, from Turbo, whatever, whatever the heck it is, I don't know. They've all got their super largest context and Gemini Pro now just went from a million token context to 2 million token context. And their mission or their goal is to go to infinite context is that if you want more context, you simply can. And they basically, I mean, in their presentation way, you know, we have a plan, this is going to work and blah, blah, blah. And I'm sure they will, I'm sure we'll figure out infinite context to the point that it will just kind of seek through and find relevance, no matter how deep in the data it is or how much data is being utilized or read from, so to speak. Because this is the one enormous value of filtering all of that information and dealing with information and data overload, which is a lot of our problem today. But the thing is, is if you are making decisions, if we're not checking this, like, it may literally just be like, yeah, we're on schedule for, you know, April 25th release or whatever, and just not true.
[00:28:37] And even ChatGPT4O, there were numerous times where it misunderstood and it could obviously be corrected, but you can't just rely on it. And if you can't rely on the answer, if the answer is only right 80% of the time, 90% of the time, then you basically have to still do the work to check the answer that the answer was supposed to not, was supposed to make it so that you didn't need to do you know what I mean? And maybe it's easier to verify an answer rather than go and figure out the answer yourself. So maybe it is still a huge time saver. But part of me just wonders how many times people will be completely led astray by one error that will cascade through a series of different questions and actions and decisions. And that if you aren't checking, if you are not, if you are trusting and not verifying at every stage of this thing, what are the consequences of more and more of our actions, decisions and things being handed over to just assuming that the LLM is correct about what it is telling us? And how difficult is it, considering the number of times, how big these models are? I mean, I still, obviously I expect a bunch of bad answers or incorrect code or something from the smaller models that I'm running on my machine.
[00:30:08] I fully expect those to be less capable and less accurate than the, you know, 300 billion or trillion parameter models that OpenAI and Google or whatever have. But when you still have those giant models and you still have so many bad answers, what happens when we start relying on it and integrating it in everything? Like if we are able to get, you know, five times as much done, but only 80% of that is actually correct and we're making five times as many errors as we were making?
[00:30:51] You know, what's the cost to have to undo or correct course with faults and bad information that we then act on? And it takes like, how many, how often do we backslide? How often do we actually have to retrace our steps to figure out where there was a problem and then do the work over again because we started from a false premise or a bad answer or bad data from your data set. And I'll use the vision thing as an example because this is one where there are constantly errors and the LLM doesn't know when to say. It doesn't know. It has no idea whether it's right. So it just guesses. The entire thing is guessing. It's literally like a calculated probabilistic guess. And it can't not guess. If you ask it something, the only thing it does is guess. It's just likely right in a lot of easy and obvious scenarios. But the thing is, is most people are generally going to be right on their guess in easy and obvious scenarios. So I want to use a few examples of what I mean by exactly by this and why. A couple of their demonstrations, I think, kind of show why this isn't going to be as useful as it seems. But then also where I think this will be massively useful and where it should be, especially in the context of what we want to use it for and what will be best for just humanity in general, where I think the real, like, unbelievable value of a lot of these tools will really land.
[00:32:43] This episode is brought to you by Swan Bitcoin. So I had a friend who I told to get on Swan Bitcoin who was looking to get in, get into Bitcoin, and they did not. They instead went to Coinbase because they knew Coinbase, I guess, or they saw it on TV or something, and they did not take my advice and they went to Coinbase and they said, yeah, yeah, I saw them four or five months later and they had gotten. They quote, unquote, told me they had gotten into bitcoin. Says, yeah, I got a. I got a number of them. And I was like, what do you mean? And he said, I got some eth and I've got some rally, which is, you know, supposedly, like really cool. I've got some. And he just started naming me and he brought out his Coinbase wallet and he showed he didn't own any bitcoin. He did not have any bitcoin. And even worse, is that some rally token that he was sold that he was apparently promoted to him, promoted to him by Coinbase, had gone up in price and then they had delisted it. So he had no idea what to do with it. He could not sell it for bitcoin. He couldn't do anything with it. He didn't know how to do a Wallet Coinbase doesn't have any good resources for any of that. And now he is stuck as it continues to plummet back down. And I'm trying to help him and I just wish so badly he had just gone to Swan Bitcoin. Understand when I say that Swan Bitcoin is the easy way to do this and the way that is not going to get you in trouble and not going to trap you and will actually explain to you why there is a difference and what you should be doing and how to hold your own keys. Bitcoin is not going to be delisted from Swan Bitcoin. They will never push garbage onto you. They will never encourage you to trade. They will go out of their way to make sure you understand what you are buying, why you are buying it, how to hold it safely, and how to even make sure that Swan is not a source of vulnerability that you can automatically set it up to withdraw to your hardware. Wallet, please. I am sick of hearing this story. I am sick of watching people get wiped out, not know what they are getting themselves into and be led astray by stupid exchanges that just love that you have to buy and dump and watch the charts because every single time they make a fee they always end up as sad stories. They are not happy stories. If you go to Swan Bitcoin and you want to get into Bitcoin, you want to understand why and you want to do it safely and you want to do it for the long term, please go to swan bitcoin.com guy. My links and details are in the show notes. It's a perfect jumping off point to not be led astray. This is why I have partnered with them. This is why I recommend them. Still to this day they're Bitcoin only and they will help you understand why. All right, I'm done with my pitch. But really like that really upset me. He's a super good guy and he's lost thousands of dollars and none of it had to happen. I guess I'm mad at myself a little bit too. Links and details will be in the show notes.
[00:35:51] So going back to the review with the R1 is one of the things he tested it out on. And this is another great thing where there's so much nuance that the idea of an AI being able to do this easily or accurately seems so off the charts unrealistic to me that I can't believe that a lot of these companies are actually boasting this sort of thing. And he like walked around, he just like tried to get it to identify a bunch of plants and it just basically failed. And I could see it from a mile away. My experience with these models and things is especially the vision models. They are good, they are useful in certain contexts, but they are awful. If I had to say a rough hit or miss on a lot of the visual things and how it properly describes an image or how it identifies whether or not something is a meme or something.
[00:36:55] I maybe 50, 50, I don't know, maybe 30, 70. Wrong most of the time. I mean that granted that totally depends on what you're asking it. Like I would say it's probably going to be wrong a majority of the time. If I point it at a plant and ask it what a plant is. Like I could point it at a completely fake plant that's not even like anything explicit. Like it's just a fake thing with green leaves. And it would come up with. It would just conjure something out of thin air. Or it would just tell it. It would just tell me that it's some sort of a plant. And you know, the R1 is a great example because I don't think it has.
[00:37:33] Think about the type of data that you would need. Like the model itself is just weights, right? It's just guessing the amount of finesse that you would need. You would need one of the largest data sets possible with incredibly fine tuned and so many different examples of every type of plant and details about what distinguishes. Like there are so many plants that look exactly like each other and where one is literally toxic, like will kill you and another one is perfectly healthy and something that you could actually eat. And it's like just the veins in the leaves will actually distinguish them because it's literally survived by mimicking this other plant. You'll never be able to get that with a general model and more specifically you'll never be able to get it with a model itself. Only a model that is checking against that database. At least in my opinion. How I've seen these things act, I think the idea of selling that as like a thing that it does outside of just being a ridiculous gimmick that is going to be completely useless in practice. I think it just makes no sense. Like I would never, if I was trying to do any of this, I would focus on something very practical and that has an extremely high success rate because people are going to grab that product and go around and try the things that it showed in the demo and then you're going to actually experience it and you're just going to see it give an error, make something up, produce ridiculous results, like show it something in your refrigerator. It's not gonna be able to see what 80% is actually in there. And it's gonna give you a recipe for God knows what. Like even in the review or in the demo of R1, they, you know, took the picture of the refrigerator and showed all this stuff that I was just, as I was saying before was just like immaculately like presented to the front of the refrigerator. And then it said, you can make an omelet. And it's like, okay. And I think even the review, the review guys said this is that you can make an omelette out of anything. Like you make an omelette, if you have eggs and there are any other ingredients in the refrigerator, you can make an omelette. So it's not exactly like inspired recommendations. And now this brings us to GPT4O, which I don't quite understand the naming convention there, but whatever. But they're releasing this for free.
[00:40:11] It has, it is multimodal in multiple ways. It is visual. You can turn on a camera and show it something. It's, it's, it also has a native desktop app now. And this is another thing that Google and Gemini are doing, is that they're integrating it with desktop apps. And I have to say, just generally, like, this is extremely useful and I am actually kind of excited about the desktop app, but I also have just a massive amount of reservation about doing this. Like this is the integration that I want in the sense of the tools and the, the language models that I want to use.
[00:40:53] And I would love it if those language models or the integration was actually generic to what I wanted to use behind it. If I could just basically swap in and out the models for text, speech, for translation, for the LLM, for the visual model, et cetera, et cetera, that would actually be wonderful because I want that interface, I want what they showed in the desktop app of just this simple way to interact with a bunch of different things on my computer. But I do not want it through GPT servers, through ChatGPT servers. And this is going to be supposedly faster, vastly more capable, available to anyone, and free. Which means they are selling the ever loving crap out of everything that that thing sees, everything that it listens to, every word that you say to it, every PDF and document and important business invoice and contract, and everything that you give it to, every spreadsheet, every chart that you are trying to sort, you know, what is going on with your private finances, all of that is being collected and being sold. There is no other way around it. They are selling the ever loving crap out of all of it. Are they selling it to agent government agencies, three letter agencies? Probably. Are they selling it to other people to pull together data sets to build better LLMs? Probably. We can always say they're trying to use it for the good purposes and I highly suspect they will use it for the positive purposes for the better models for the, you know, more advanced AI, the more convenient tools and you know, assistance and tasks and all of these things. But it will be used for all of the bad things too. In fact a lot of the bad things go hand in hand with the good things. When you start having applied like some the degree of surveillance and funnel of control and information that something like this has. I'm still wondering at this point how they are alive. Like there is no way. I cannot fathom that they are doing this cash flow positive right now. If they are going to release 4o for free, there has got to be something else propping them up. And I don't know what that is. I don't know if it's financing. I know, I don't know if it's just constant VC money like pumping them up. I do not, do not know if they're getting a massive amount of government money in the background or just you know, basically being inflated like costs are being offset somewhere. I don't know what it is but somehow I just cannot fathom that they are cash flow positive like, like in a direct value sense for the amount of people who are probably using their AI API and using that integrating with ChatGPT and the amount of people who will be using it with 4o, if it is free and anybody can just sign up, they are getting something else out of it. I don't know if this is the reason they're succeeding so much is their business model is completely different and it's not what we think it is. But I don't know, I don't know, maybe I'm crazy. Maybe you know, you have a 200 million, 500 million users, whatever it is and you know you've got everybody to pay for premium. But if they've got 4.0, if 4.0 is intended to be free, it seems like they're shooting themselves in the foot. If that is actually their business model, if they've got that many people and they are actually cash flow positive with that, I don't know just seems like a really massive what they are trying to have for free. Is a bit crazy. So let me go a little bit more into detail in it since I haven't covered it in length here. They've got a desktop app and the main interface now or the main interaction with ChatGPT or with GPT4O is through speaking is explicitly that you speak to it. Then it generates a response and it speaks back to you in real time. And they demonstrated this in their release or in their, you know, announcement of this, this shift. And what's funny is that their demonstration of this really kind of overshadowed because Google I O happened at the same time.
[00:45:35] And while a lot of what Google was releasing like their video gen stuff is actually really cool, I'm. I'm very interested in Google's video gen stuff. I'm a big fan of the generative media anyway or generative AI as I know a lot of, you know, if you've listened to the show at all, I think there's some really incredible storytelling potential there.
[00:45:58] And there is specifically a story and a project that I want to tackle when I've got a couple of other things under my belt and when those tools kind of reach the place where I think I can tell that story pretty easily or at least with very low funding and a skeleton crew, so to speak. But I did. I haven't dug enough into the Gemini or I guess it's not Gemini. They have their own name for it but the, the, the video generation that Google has recently announced to compete with Sora, which is the one from OpenAI which still is not accessible like you know, they announcement announced it months ago now and had a closed beta or closed private access to and I signed up for it with three different emails and still haven't gotten anything. I have not gotten access to it. And the video gen capabilities of Google that they showed at IO, they're. They're very impressive. They. I think it does compete with Sora, but I really want to see, I want to see people kind of like randomly test stuff. I'm questioning a couple of things that I saw. There was. There was one that I saw that was specifically about a.
[00:47:20] It was like a panda or a bear. So I can't remember exactly. That was playing the guitar and something about the way it was done made me think that they didn't have a strict video or not necessarily. Obviously they have a strict video generative video gen model. Right. They have a video AI model that will generate video based on the prompt. But I think they also might be using a combination of videogen technologies to create the Result in the sense that it's not just a straight like this is a like stable diffusion video is a single model. And I'm inclined to think and maybe it's just the single video that I saw that led me to this. And again I haven't dug much into this. This is mostly kind of my first thoughts on kind of the overview of a bunch of things that have happened very recently.
[00:48:22] But I saw the Panda playing the guitar video made me think that it was actually the combination of a videogen model but also those. There's a handful of different models and different methodologies that allow you to take 2D images and turn it into video. And it almost gave me the feeling the way I saw some of those work like there was a new one from Microsoft called. Was it Vasa Man? What was, what was that one? I just read this paper very recently but it was a fantastic next step sort of upgrade from Microsoft on the idea of taking 2D images and being able to turn it into 3D. And this was a talking head sort of video generated generation from 2D images. And it was really good. You could still tell with the warping at the very edges and the warping when the head turned a little bit too much. But it was extremely emotive. It did a fantastic job of filling in the teeth where you couldn't even see the teeth for the person of showing facial expressions and movements and everything was extremely, was very natural. I won't say extremely, extremely is a little bit more, a little bit much. But it seemed very natural and it was only when there was a lot of motion and a lot of things. And of course obviously the, the, the clear basically boxed in movement like there's limitation to how far it can move because it's coming from a 2D image and it's having to fill in the background and that sort of thing. But it was increasing. It was really impressive in, in comparison to a lot of the other 2D 2D to video like image to video talking head sort of models.
[00:50:33] But the thing that I saw that was supposedly a demonstration of the Google video model seemed eerily reminiscent of that in the sense that it might be a combination of things almost like there might be a videogen mixture of experts is that oh well this is clearly like a, a person or this is a, this is an animal or something in this posture. Well, let's use the, you know, animate anything model in combination with a certain weight of our generic video gen model in order to create consistency and in doing so like I'VE been wondering if something might kind of go this route in the combination of different models that specialize in different things. Because you could get really good consistency in certain objects and you could essentially map that or weight that. On top of all of the inconsistency and frustrations and weird morphs and all of the things that you get with the typical videogen model when you're talking about a generic model. I've been kind of waiting for that mixture of experts moment of video generation. And because these aren't open source, we don't really know, you know, I don't know what methodology they're doing, but something about the way it looked just kind of felt like that to me. And it makes me wonder in that one example if that might be part of what's happening here. And that's really interesting if that is. And it makes me rethink what SORA has done in order to create that consistency.
[00:52:15] Because, you know, whether or not they did a mixture of experts where they just have like the biggest model, they have like you know, 500 billion trillion parameter model or something ridiculous on video generation and because they have such a long context length, I don't know if you would call it context length in that example when we're talking about video generation, but the fact that you can generate up to 60 second videos, that is a, you know, that's a huge difference. And the ability to get consistency across that. Like a lot of things that are like really odd is when you have somebody walking or an animal walking and like legs cross over each other and then like morph out of the shadow of another leg is suddenly the leg of the the next leg. Just really weird things, especially the longer the video goes or like they're walking but they're not like properly moving across the landscape at the same pace, you know, like a video game, like you'll like run into the end of the level and you're just walking but you're not moving anywhere. Like you're just moonwalking in place. There's a lot of that going on in the video generation models.
[00:53:31] And Sora, that was one thing that SORA just kind of like blew my mind is how well it seemed to cross that chasm.
[00:53:40] Even, even crazier is some of the videos that they showed as examples in which it did clearly defy physics or defy the fact that like there was, there was an image of a dog moving from like one window seal to the next and the shutter for the window. Again, this is OpenAI's Sora model, not the Google one.
[00:54:05] The shutter on the window was like sticking out where obviously the dog could not cross in front of it. But when the video tried to get the dog from one window seal to the next, it just kind of had the dog cross in front of it, even though, you know, obviously you can do it in a 2D image, but in a 3D world that he would have been passing through the shutter. Well, the crazy thing was is that a lot of times when the video generation models will hit those oddities or the, the kind of contradictions or whatever, it will not do so elegantly. But it was kind of fascinating to see it where it clearly was doing something that the dog shouldn't have been able to do. But there was nothing unnatural about it. When you saw it happen, that happened. The dog still looked consistent. The window seal, the, the building and everything appeared consistent. It was almost as if the, the shutter that looked 3D like it was sticking out was suddenly, you know, painted on the wall for a second. You know, he just like passed in front of it. And it was just interesting to see that because showed what happened. It showed how incredibly consistent and how it tried to create the sense of reality even when things didn't quite line up, even when it did make a mistake, so to speak. And you know, the normal human could look at it and be like, well, that's not supposed to actually work that way. But all that is to say that I was just still am incredibly impressed by Sora in particular and it seemed like such a huge leap when it came to all of the other video gen models. And I'm very curious because Sora didn't give me that feeling at all.
[00:56:04] Their consistency seemed genuinely in like the entire model just was producing consistency and there was a whole thread of or just kind of a series of people who were getting access and who were asking, you know, giving prompts to. On. On Twitter that were then, you know, producing the results of those prompts. And I never got that sensation, I never got that idea or inclination that this was something where they were using Animate anyone or Vasa. I might be wrong on what that was. I just remember like asa something specific. I'll have the link to it in the show notes if I remember to. I've got it saved somewhere for the. The talking head one but I never got the idea that it was using a combination of different things. And then I did read a paper.
[00:57:03] I think it might have just been the release from OpenAI actually, I'm not sure it was A research paper. But, but I do remember reading that they to, to produce their model that and this actually allowed them to do stuff in greater resolution as well is that they actually broke up all of their videos into like 128 by 128 pixel squares and then broke down and described what was in all of those squares versus comparing it to what was in the overall video.
[00:57:42] So I don't know how you would even accomplish like the degree of finesse I guess and precision that you would need in describing what a small square of the video was, you know, that like just a dog's leg was going across it and this was the, you know, the window seal or something and then having not only the model weights about just that square, but then also the model weights of that square in relation to all of the other squares in the video and then the model weights for the entire video and all of the pixels like that just seem, I mean from a mathematical perspective that just seems like a monumental amount of computation, which I'm sure it is. And especially if you're generating 60 seconds of stuff but just huge leaps in that and they immediately, immediately like the bar skyrocketed for oh well, if you're not producing, you know, at least 30 second videos then you're not even competing anymore.
[00:58:52] And it makes me wonder and I guess that's all just to say like this is all total speculation on my part about how Google is doing it. I haven't read, I mean maybe they released a paper in like, you know, explained all of their, you know, details. But I would love to and intend to dig more into that especially as I maybe start to unpack the project that I want to tackle.
[00:59:19] I want to have a lot of video on exactly how I do it and how I pull it off.
[00:59:24] So that'll be something to watch out for the YouTube and or rumble channels.
[00:59:29] But yeah, I'm really curious if Google has done something a little bit different to kind of shortcut that and gone the mixture of experts route from the context of a videogen. You know, if you can combine a generic video generation model with one that is excellent at pure object consistency or simply the animation of people like that's all the video generation model does. And then when you have a, when you're generating a video where you know that you are generating a person walking, well then you, you know, in your quote unquote mixture of experts, you grab your generic video gen and then you grab your person animation videogen. So I'm deeply curious about that and I Want to dig into it more, but this isn't really kind of the focus. That was just one of the things that Google I O is sharing because they have a much better Videogen model now than they did. And GPT4O, there's another part of how they demonstrated. First off, the demonstration was really impressive and it will absolutely be very, very useful. But in my trying to separate the signal from the noise, there were also a lot of things, especially the integration with vision of being able to point it at something or do something and you know, watch like a math problem on a piece of paper that again just felt gimmicky like, and I don't mean it in the sense that like I'm not impressed, like it's profound, just from a general concept of technology that this is even possible.
[01:01:33] But it doesn't seem extremely useful. So I use the example in the video and I'll have the link to this in the show notes if you want to watch it.
[01:01:43] But they're talking to this thing, they're talking to GPT4.0 and one of the things that you'll notice with because it's being very emotive, it's like, oh, hi, yes, it's fantastic to talk to you. How are you doing? How can I help you? And then he's like, oh, I want to show can you help me with this math problem?
[01:02:04] And then he points it turns on the video and points it at a piece of paper. And it's funny because one of the things that ChatGPT does, which frustrates me and was clearly a problem in this demonstration is the fact that Chad GPT will just keep talking like it will just, it will give you it. I can't tell you how many times I've looked for just a yes or no answer or like a one sentence, just like explain this or break down the clarify something of like a paragraph that I gave it and then it literally wrote me like a three page term paper. And it was really funny because in the demonstration chatgpt just kind of kept talking and multiple times they just like immediately were like stopping and talking over it. And it's funny because you know you're trying to get this sensation that you are talking to a human and it's doing a fantastic job of mimicking a person. Like I can easily see how people would want to be polite and like get all of the inclination that they are speaking to a human and having that social interaction. But what's funny is that if it's just constantly, just over talking like it's just wanting to give a lengthy answer and essentially to get it to do what you want it to do if you're constantly interrupting it. I can see that as creating a very, very bad habit in a lot of people. Like so easily. I could see myself talking to gpt4 oh, for a whole day working on a project or something and have to interrupt it like a thousand times and then turn around and have a conversation with someone and just endlessly interrupt them to the point they're like, guy is an asshole, like, would you please stop interrupting me? But the example that they used, the example that he used on the piece of paper and this is why the whole vision thing just, it has suffered every time that I have seen it done it. Despite the, the amount of like, I'm literally impressed that this is possible that they can point a camera at this thing and it will read the, the equation that they did on a piece of paper. And he did three x plus one equals four. And it helped him, quote, unquote, solve it. It gave him hints without telling him how to do the answer, but just kind of, you know, asking the right questions to lead him to get to that question or to get to the answer.
[01:04:30] But this math was shockingly simple and he had to write it in massive super blocky text where 3x +1 equals 4 literally took up the entire like from side to side of a, you know, 8 by 11 sheet of paper. So if you're doing a, you know, a 3 variable, like 10 character math problem, it is probably not going to be of any use to you. And there were also multiple times, even though you could easily correct it and it would adjust, there were multiple times where it was taking context from the last time you had the video on. And this is something that it does I'll notice as well, even when I'm using it. So like, one of the things that I do with this show over all of my episodes now is I have a built in prompt in my notes that I have slowly adjusted to get the best answer and I've talked about this a couple of times, is that I will get it to describe the show, describe the episode, and also pull out any of the links that I've mentioned. Like I've already said multiple times I'm going to link to this thing or this other content or the Google I O video or the OpenAI video and it will specifically link those out or it won't actually get me the links, but it will tell me that don't forget that you Mentioned this in the episode just by generating the transcript and then putting it in ChatGPT and then asking my prompt. However, one thing that I've noticed is I'm constantly. Now I used to work in the same conversations a lot where I would just keep asking. Like, I had One thread with ChatGPT that was my asking of describing this episode and get me my links and all of that stuff.
[01:06:30] And then I noticed this was after doing this like 30 times, right? I noticed that sometimes I would ask the question.
[01:06:38] And again, I've never actually, I think once I used its description, like, almost entirely. And it was really good. And I've gotten good with the prompt to do better, but I mostly just don't like the way it talks. It's just. It sounds so markety or just there's something like a little bit fakey about it that bothers me. But it's great to just remind me of, like, oh, I talked about all this stuff. That's right. This was part of the thing. One thing I noticed is I would ask it sometimes and it would give me the answer from, like, two documents ago.
[01:07:19] So I asked it one time about an episode for AI Unchained, gave it the transcript of the podcast, and everything was right there. And it properly pulled the links from it. But then it described a bitcoin chat interview that I had had that was like the previous document that I had uploaded. Like, the context link was just suddenly, it was still just asking about all of this stuff previously. And that's not the first. That was one of the first times that I think I noticed it. But I think I had gotten a hint of that happening a couple of times, but it wasn't so explicit to make me catch that. It was clearly pulling from a different document.
[01:08:05] But this time it was obvious it wasn't even about AI the description had nothing to do at all with the transcript that I had just handed to it. And this was the same thing that I saw in the demo with the video with the vision. In fact, one of the things that he was talking about, he was asking it to describe. He's like, hey, hey, GPT4O. Can you tell me what, what my facial expression, or excuse me, can you look at my facial expression? Can you tell me what emotion I'm probably feeling right now? And you know what can you tell me about, like, my mood? And it was like, well, it seems like you have a wooden surface here. And started describing, like, something that didn't make any sense, had nothing to do with the video that was being shown and he was like, oh, no, no, no. That's something that I showed you earlier. Ignore that. He's like, oh, okay, I apologize. That was my mistake. Oh, yes. It looks like you're smiling and you're excited about something. And what it was is the previous video that he had did, which had been cut off and, you know, taken out, it's still grabbing from old context, was him putting the phone down on the table so that the video was showing that the wood surface of the table as he put the phone down and then cut off the video earlier on in the conversation. Now, again, I do want to note that I'm not unimpressed by this technology. Like, I'm.
[01:09:48] It's unbelievable that this is possible, but I'm never gonna want. I'm never gonna ask GPT4O what my mood is. Like, what it. What does it think my facial expression like. The fact that it knows facial expressions and can make an assessment of that is cool, but I don't know of a practical application of that. And then the same thing with 3x plus 1 equals 4. Like, the vision aspect of that. Like, I would just type it in. Like, if I needed it to help me with the math problem, it would be infinitely easier to just type it into the text box and get it to. I could ask it to write me a math script. Like, write me a Python script to solve the problem. So despite the fact that the idea that it can read this on a piece of paper if I write it in huge letters and numbers, you know, across a whole sheet of paper, despite the fact that that is impressive, I don't see any way, any reason that I would really use that. And that would be the easiest and fastest way to actually accomplish that task. It's just something that like, wow, AI can do this. And then like, I'm never going to use it again. Especially when we talk about the amount of errors that the. That all of the Vision AI, which I'm sure OpenAI is probably leading the pack with actually being useful or actually giving consistent results.
[01:11:17] But again, you know, it's not going to be able to guess a bunch of plants. Like, I think there's just two.
[01:11:26] There's just enormous amounts of complexity and specificity in the vision AI stuff. And I suspect for it to be really, really useful, like, it might be in very specific context.
[01:11:42] I would bet without even the slight. If I was looking for a practical application for this, without even the slightest doubt, this would be a gold mine in like a warehouse setting with a Robot that is trying to sort boxes, read labels with addresses and stuff. Like, there is use for this 100%, but I don't see it as like generically useful that I'm going to pull out my phone and get it to tell me what kind of plant this is, then listen to some lengthy diatribe about the plant and describing it, which I'm then gonna have to type that name into Google to make sure A the plant even exists and B, that yeah, it actually does look anything like the plant. However, there is one thing that I have seen from a couple of different companies now and from There was a TED talk recently about, I think they called it the audio computer, if I'm not mistaken, that was running like four or five different AI models at once in real time, basically. And then this is also something that OpenAI did with GPT4O. I cannot remember. I don't think I saw anything specific with Google I O on this.
[01:13:07] But live translation, this is a massively valuable. This is one thing where AI is just going to obliterate social barriers, which is going to be crazy. You know, we've been thinking that we. For anybody who's been, you know, really trying to assess or think about the consequences of social media and basically everybody publicly broadcasting what they think and their arbitrary, like their random opinions about all of these different things and events, we have been going through a global identity crisis where all of these ideas that have both basically been private, quiet and, you know, reserved are now out there. They are public, they are loud and they are enraging everyone. We have basically had a clash where all the ideas that were niche, that were edge, have basically been piled onto the public table all at once in a massive way. And I think as we get live translation as the borders, the, not the values and cultural borders, because those will remain like we will still have very different values and have very different perspectives as to what any of this means and how to judge what people are saying. But as the language barriers fall away, oh man, this identity crisis is going to get a lot, lot worse because there is still a massive divide between, you know, just the normal community, the normal like cultural norms of Chinese culture and American culture where they simply do not cross because there's just not a bridge. The, the, the bandwidth of information exchange between Chinese language media and culture and conversation and dialogue and English speaking of all those same things is relatively tiny. Like, tiny. Like I don't, I don't, I don't speak with, you know, when I go on Twitter for all intents and purposes. To me it looks like it's just the US and just, you know, Europe that's awake at the time. People in the US are. But it's extremely and clearly Western culture dominant. And every once in a while I'll find myself in like Chinese or Eastern Asia, Twitter or Nostr in particular. Like there's a ton of like Chinese language, like relays and stuff, which is kind of crazy that I, you know, will come across it every once in a while. And you know, part of me is just always associated it with spam because that's the only, like I get DMs that are in Chinese or something like this. Like, oh, well, this is a scammer. But when the, the ability to translate, the ability to live have a non delayed conversation. Well, I mean, delayed to the point that it has to, you know, take it in and then spit it back out in a different language, but still the bandwidth to live, translate and have this conversation with someone who speaks an entirely different language. And specifically where those cultures don't have a bridge, like that's where those culture differences are going to be greatest. Where the people who only speak English and the people who only speak Chinese and they don't have any exposure at all or experience with that other culture, they are suddenly going to be able to interact. They're going to witness each other's perspectives. And I only use Chinese just because Chinese is such a different language, you know, and, and also just the Eastern Asian culture is very different from Western culture. But I mean it generically. Like the whole world is about to open up in this way. And again, if you haven't seen the OpenAI demonstration of this, you should. It is very impressive. And I also have the TED talk of.
[01:17:43] I really am inclined to think the demonstration of the audio computer is completely scripted like that it was fake because everything that they showed is 100% possible to the point that I can do almost all of it with various models or tools on my own computer. I know how to find a model or whatever that will isolate certain things and that will do the studio sound and all of this stuff, remove background noise. But to do it real time and to do it in a small device that is literally in your ear and to do it in a restaurant, which is the example that they used.
[01:18:25] And they, they showed it as a video and he's holding this on his ear and they're watching the video go back and forth, but he's saying this thing and then, you know, they're changing the audio in the video as if this is happening real time while they're having a conversation in a restaurant. But obviously he is not in a restaurant. And I am very skeptical, especially with all of the demonstrations that I've seen and the amount of times that is vastly oversold the capabilities of a lot of these things. I'm highly, highly suspicious that that audio computer cannot actually do that to that degree. And that's what they want it to do. And they think their final product will do that, but maybe not yet.
[01:19:18] However, they are showing us a glimpse of the Future and GPT4O can, I mean it can't do all of the voice isolation, all this stuff, but it can actively end live translate between two people and bridge a conversation. You know, I tell you a good experience or a good example of this and experience I personally had is as a technician when I was doing Internet stuff. Google Translate, you know, was getting really good. I guess this was.
[01:19:55] Damn. Was this like 10 years ago now? I don't know. It was, it was early on when I was a technician, so.
[01:20:01] But I, I had a customer which was really fun actually. We kind of had a, it was a comical little adventure of us trying to communicate because she only spoke Spanish and I, I do not speak Spanish. And also certain technical terms and things don't translate very well. You know, when I'm talking about modems and routers and you know, certain types of chords and stuff and trying to ask and get those sorts of things, you'll get the, you'll get more generic term terminology sort of translation and it just won't come across the way I'm trying to describe it. But regardless, we try to have a very slow and very complicated text translated conversation over Google Translate on a phone. And I'll tell you, just from watching the demonstration from Google 4.0God, not Google GPT4O, I can see that having been an incredibly easy endeavor, just like orders of magnitude easier than the, the situation that I had been in at the time. And then the ability to easily do this in new voice, to, to easily emote this like it's a big step. It's a big step and it's one of the more direct and clearly practical, like highly, highly valuable use cases that, that I, that I really think are going to shine with this technology, but at the same time the consequences of it are going to be crazy. Like I said, we are going to have.
[01:21:50] Our global identity crisis is probably going to get an order of magnitude worse as all of our language barriers fall away.
[01:21:59] Anyway, I will try to hopefully the AI will pull all of the links that I said I wanted to add into the description for this so that you guys have it and you can explore all the stuff I was talking about in today's episode. There's still a number of different things that I didn't actually get to talk discuss, but we will cover them in a later episode. In fact, there's. I might try to reach out to these guys actually because there's a really cool project called Nova that is having a lot of growth and kind of interest in their project and how they're doing it. And they're working specifically with a lot of businesses. And the reason is, is because they are focused on open source self hosting and privacy. But businesses that have very sensitive data is how do they get the benefit of all of that, all of these AI tools without this huge privacy concern. And this is one of those things that I mentioned, I think in the last episode or the one before, I'm not sure very recent episode about how I think the.
[01:23:08] When AI gives the illusion of human of having a social interaction, I think the awareness or visibility of the privacy problem hits a little closer to home.
[01:23:30] When you're anthropomorphizing the computer that you're talking to, suddenly you kind of feel like you can be judged. The social weight and the social concern of like, oh, I need my privacy, I don't want you to see me naked sort of thing starts to creep back up to the surface. And because of that, I think and hope that the concern and the focus on privacy is going to increase a lot as computers behave and start to act more and more like us. Like humans and Nova, I think, and the fact that they are. They appear to be succeeding and they have a business model that is specifically around this I think is a really good indication of it. And so it's something that's very worth unpacking. But that's the reason I didn't really bring it up because I felt like I needed to unpack it in maybe an entire episode. But I'll try to reach out to them and see if maybe we can do a show. Otherwise I'll just kind of explore and see what's. What's worth unpacking on the episode on the AI Unchained here. So stay tuned, stay subscribed. Lots of really fun stuff to cover. And I will be breaking down and trying to make sense of quantization pretty soon as well. Hopefully sooner rather than later. Depends on how complex it is basically.
[01:25:00] And that will do it. Don't forget to check out the links. Don't forget to check out our amazing sponsors, Coin Kite, makers of the Cold Card Hardware wallet to keep your Bitcoin safe and Swan Bitcoin to buy and stack automatically, easily and without being led into a bunch of bullshit. Thank you all for supporting this show, for following me and my work, and I will catch you guys on the next episode of AI Unchained. I am Guy Swan. And until then, everybody, take it easy. Guys, the limits of my language means the limits of my world.
[01:25:52] Ludwig Wittgenstein.
