Functionally Useless, But Undeniably Cool

Mason Amadeus: Live from the 8th Media Studios in the back rooms of the deep web, this is "The Fake Files." When tech gets weird, we're here to make sense of it, but this time, all you get is me. Perry is out this week, which means the average IQ of the show has plummeted by more than half, but we're still going to have a lot of fun today. In our first segment, we're going to check out a video game that is fully AI-generated on the fly, frame by frame as you play it, and we'll play it together right here on the show. After that, we're going to take a look at my favorite kind of technological tomfoolery, when someone clever decides to push the limits of something completely mundane and ends up creating a PDF file larger than the observable universe. After that, segment three is a submission from one of the newest members of our Discord. We're going to look at a language diffusion model. It's super weird. You can watch how it thinks. We'll explore how it works differently from the LLMs we're more familiar with and what that might mean for text generation. And finally, we'll wrap it all up with an AI dumpster fire of the week about how Google is paying their top AI talent to do nothing. And that might sound like a great gig at first, but the reality is a little bit different. We've got a lot to cover and you're stuck with just me for this week, so sit back, relax, and I'll do my best to be your one-man mixture of experts. We'll open up the fake files right after this. [ Music ] So I will be the first to admit that this is-- it's very hard to call this a game. It's really a glorified tech demo. And the reaction to it has been pretty negative as far as I have seen, but I think there is something undeniably cool at the core of what is going on. So, Microsoft has released a demo of a version of "Quake 2" that is entirely created by AI on the fly while you play it. It's using their new AI system called Muse, which is not an LLM or a diffusion model or whatever, it is a world and human action model. They're calling it a WHAM. Basically, it takes two prompts at the same time. It takes a visual and it takes button inputs, combines them, and uses that to generate what the next frame of a video game should be. They said, "It is incredibly fun to play as a simulated version of the game inside this model, but there are, of course, limitations and shortcomings of their current approach. And that is very apparent the moment you jump into it. Muse has a memory of like 0.9 seconds of gameplay, which equals nine frames because it runs at 10 frames per second, which is absolutely way too slow. But this means that if something like an enemy or an object goes off-screen for too long, the AI forgets it was there. And everything looks blurry and doesn't move very well. I'm reading here from the TechCrawler article that made me first aware of this. But before we go any deeper, let's jump in and play it ourselves so that we can just get our own opinions going here. You can do it too. You can play it right in your web browser. I'll drop a link in the show notes. But here we are. If you're just listening to the podcast, this will be one of those segments you probably should check out on YouTube, but I'll do my best to describe what's going on. It's basically just a very sort of blurry in that AI fuzzy way rendition of Doom II. There's a bad guy in front of me, but actually, I probably don't have to fight him. I bet if I turn around, he'll disappear. Oh no, there's another one. Every time I turn around, there's a new one. It's very much dream logic. It's like playing through a fever dream. Oh, okay, that's just a closed door. If you get like too close to a wall, check this out. If I get too close to a wall and I just sort of stare at it for a little bit and wiggle-- I bet if I turn around, I'll be in a different space. Oh, not quite. It actually remembered that time. It is very unpredictable. I have played a lot of these demos. You can run it for two minutes at a time before it kicks you out and you have to start over. I think that's probably just to try and keep usage down. But I have had very different results each time I've played it in terms of it remembering where things are. Yeah, I think-- nope, it's still got us in the same room. Okay, cool. Sometimes it'll just teleport you into a random basement or you'll run past an enemy and they will turn into a barrel or-- yeah, this is like bits on the ground in here, but I bet if I spin around things will be different. Yeah, see a lot of stuff just vanished. So it's like playing a fugue state. It's like being in a fever dream. And at the same time that this is completely unplayable as a game, really, it is very cool from the standpoint of what is going on under the hood, in my opinion. You can jump, you can move, you can shoot your gun, you can crouch. The input lag is dreadful here. You'll hear me hit the spacebar and I'll tell you when he jumps. He didn't. There we go. So that's like almost a second's worth of delay sometimes. It's pretty much impossible to play, plus you have to look with the keys instead of the mouse, which is a nightmare. So it's not fun, but it's very cool. And yet, a lot of people really, really don't like this. Some people call it the very definition of AI slop. And honestly, I do see where they're coming from because it pretty much is. But that's what I think is kind of cool. Because we'll get into what Microsoft thinks they're going to do with this in a minute. But I think there's potential here for someone who is clever and creative and has a cool idea for a game that could make use of these limitations, this kind of dream logic fuzziness to do something very cool. That said, that doesn't seem to be what Microsoft wants to do. When I was looking at this article from Xbox Wire, they were saying, we're talking about the future of this. "Today, countless classic games tied to aging hardware are no longer playable by most people. Thanks to this breakthrough, we're exploring the potential for Muse to take older back-catalog games from our studios and optimize them for any device. We believe this could radically change how we preserve and experience classic games in the future and make them more accessible to more players." I think that's dumb because it is absolutely not preservation if AI is hallucinating every frame of this game, it's just a convincing hallucination of an older game. If you want to actually do preservation, there's much better ways to do it. I was literally playing "Unreal Tournament '99" the other day, and "Carmageddon", the original one from 1999. There's a lot better ways to preserve old games, and I feel like they have to know that. It's weird to me that this seems to me like the only future use case that they really put any weight to. Like most of this paragraph is just all about that. Whereas I think when this gets into the hands of developers and artists and creative people, this could result in something very neat. And what's interesting too on top of that is they actually did consult with a bunch of game developers. They interviewed 27 global game creators from indies to AAA Game Studios to quote, "Make sure the research was shaped by the people who would use it." And I mean, maybe that shows in the fact that the model is very usable, with all of its limitations and whatnot. But, I don't know, the future remains to be seen. Basically, I hope this gets used for something cool. But let's talk about what is going on. Like, what is this? It's a world and human action model, sure, but like, how do they make it? Well, they partnered with a game studio called Ninja Theory. They're behind the game "Bleeding Edge", which is actually on sale for 90% off on Steam, so you can pick that up for three bucks. But the reviews are mixed, and they all say that like the multiplayer is dead, so I don't know if you actually want to do that. But basically, Microsoft partnered with these fellas and used this game exclusively to train their Muse model. They used anonymized player data from just under 30,000 players, which added up to, I think the figure they gave was like seven years' worth of gameplay once it was all, you know, stitched together to train. The model takes the previous frame of the game, and then whatever button is being pressed, combines those in its training data, generates the next frame of the video game, and so on. So it is an autoregressive, you know, linear forward prediction model. It doesn't really have, as far as I've been able to tell, any concept of state, which is why it's not tracking like the location of players and enemies or the actual layout of the map in any way that would be like normal from a game development standpoint. You can play with this model. They have released it into their Azure AI Foundry, which admittedly I have never heard of until today, but I just don't. I think I have some trauma from my brief stint as an IT manager and all the headaches I got from Azure, so I've just never checked it out. But you can apparently get this Muse model and play around with it. I don't know if you can actually make a game like that "Quake" demo, but you can use it. They say that you can take screenshots from the game "Bleeding Edge," drop it in, combine it with button inputs, and see how it generates sort of the next frames that could potentially happen. There's a cute little video on their website showing what it looks like. And you really do. You just drag in a frame, you tell it what buttons are being pressed on the controller, and then it generates a bunch of possible futures and picks from them. It apparently took five days and 98 H100 GPUs to train it. So, doing a little back-of-the-napkin math, assuming each of those GPUs is doing their max draw of 700 watts, that is 68.6 kilowatts for 120 hours, resulting in 8.2-megawatt hours. For comparison, the average US household uses almost 11 megawatts per year. So it took almost as much as one household's entire year's worth of electricity usage to train this, which is a big chunk of power, but that only happens once. And in the context of everything else, it's not really insane, but I'd be remiss not to mention it. It remains to be seen too how much power inference takes, like when you're actually using it. That also depends on the number of users and whatnot. But we can tell that it took about a year's worth of one average American household's electricity usage to train it. Let's talk about how people reacted to this. I had a hard time finding anyone saying anything positive about this, aside from, you know, the usual hype suspects on like Twitter and LinkedIn. Most of the articles you'll find about this, I feel like are very negative. Like this one from "Rock Paper Shotgun" titled, "I Strongly Feel This AI-Powered Demo of 'Quake 2' Is an Insult to Life Itself." And I have found a fair few articles with titles like this, people calling it pure AI slop. Someone replying to the announcement tweet saying, "Microsoft will literally do anything but develop real video games." And like, yeah, I can't really blame people for having that sentiment. I know a lot of people are very sick of AI, and a lot of people just hate anything AI right now. And like honestly, I can't blame them, given just sort of the world we live in. And I'm here making this show, right? But I do think that what is underlying this is a very cool piece of technology, and I think there's probably some cool creative uses that will come out of this. But if you want to play "Quake 2", just get the remaster off Steam. It's awesome. It's a phenomenal remaster, and it's like 10 bucks. And playing that janky AI demo of it has made me want to revisit that game. Right after the break, we're going to take a look at something very fun. We're going to look at how somebody created a PDF file that is larger than the observable universe. Stick around for that. One of my absolute favorite things is creative misuse. Whether it's of software or like using office supplies to make art, using things wrong in a creative way is just like near and dear to my heart. One of my favorite YouTubers is Josh from "Let's Game It Out," which is a channel all about breaking games from the inside, just by playing them and spending inordinate amounts of time doing things like finding ways to spawn a bajillion items until the physics engine eats itself, or building a tornado of conveyor belts in a factory game. But What's even more fun is when somebody takes that mindset and applies it to something much more mundane, and that is exactly what happened when this absolute legend, Alex Chan, made a PDF file that is bigger than the observable universe, and let me explain why. We're going to have to back up a little bit, get some background to talk about this. First, like what do we mean "bigger"? We're not talking file size. I'm talking about the actual page of the document. So, you know how a PDF document is broken up into pages. They can be portrait or landscape or square. They're usually like A4 or US-letter sized or that weird, really long legal paper that I've actually never seen in real life. The size of the page is what we're talking about. And for a long time, there was a piece of digital folklore floating around that was saying the largest possible size of a PDF was 15 million inches a side, which would make roughly a square that is roughly 237.7 miles on each side. So about the size of New York State or Iowa. Or for some reason, everyone keeps comparing it to 40% of Germany, which to me is pretty useless as far as comparisons go. So it inspired me to figure out. It is also a little more than 11 trillion slices of bread or 50 billion square Mark Zuckerbergs. But I digress, the figure is not correct. I'm looking at both a "Big Think" article about this as well as Alex Chan's own blog. And taking a cue from the author of this "Big Think" article, Frank Jacobs, I want to talk about the history of PDFs very briefly because they're fascinating. It's a great rabbit hole to fall down if you're bored. But really quick, John Warnock, who co-founded Adobe in 1982, created the PDF standard as a way to try and fix documents in place so that no matter what system you were looking at them on, they would all display the same. You know how websites change based on the width of your device, or if you're on a phone? It's actually one of the most frustrating parts of web development is adapting to that. Imagine if every document was that way, and you couldn't know how it was going to print until it printed, or I guess until you got to the print preview. PDFs make everything stay right where they're supposed to be. They internally contain all sorts of extra data, like specific fonts and whatnot, so you can view the file exactly as intended without any extra steps. They came about in 1993. And not only did Adobe make the PDF format, they also needed something to read it, so they created Adobe Acrobat, which I'm sure we're all familiar with. So, in this way, Adobe created both sides of the equation, right, the file format itself and the program you need to parse it and to read it. Nowadays, thankfully, there are alternative ways to open PDF files. It eventually has become a free and open standard. There's actually a really storied history there, but we can't get into it. It turns out that this folkloric limit of 237.7 miles per side is actually a limitation of Adobe Acrobat Reader and not the PDF specification itself. In a PDF document, internally, they use a base unit that is 1/72 of an inch, which sounds completely arbitrary and nonsensical, but it's a holdover from the days of physical printing, where-- you know how fonts are measured in points? A point is 1/72 of an inch, and so the idea behind it being that a 72-point font will print out at about an inch tall, so you could have some sort of idea of how big your text would print out. We'll go way off topic if we go any deeper on that, but that is the unit that Adobe PDFs use internally? 1/72 of an inch is the base unit. It's a point. It's the same as fonts, which makes sense because it's a document that's going to contain probably a lot of text. As well as that sort of base unit, there is also a user unit value that can scale the internal point even larger. It just basically multiplies the page size by the user unit. Acrobat 7 supports a maximum user unit value of 75,000, which gives us a maximum page dimension of 15 million inches because the maximum supported page size was 14,400 units. I'm throwing a lot of numbers at you, I know. Basically, Alex, being the best kind of unhinged, set off on a task of creating a PDF file from scratch by hand, which is wild. That way, though, she could set whatever values she wanted inside of the document, unhindered by any particular software tool because she wanted to see just how big a PDF could actually be because that limitation was in Acrobat not in the file format. So, she said by using a good article, which explained the internal structure of a PDF, combined with asking ChatGPT a few questions, she was able to get enough to write some simple PDF files by hand, which I think is nuts. I've poked around inside of a PDF before, and it's very confusing, and it's a mess in there. And she says as much in here. There's a really great breakdown if you want to get deeper into it. Of course, we'll link this in the show notes. She even says, "It quickly became apparent why nobody writes PDFs by hand. It got very fiddly to redo all the lookup tables. But with this newfound ability to edit PDFs by hand, how can I create monstrously big ones?" And what she discovered is that-- well, it's two things. One is that that user unit value we talked about, that scaling factor, actually isn't really supported by most readers. And the built-in Mac OS preview app, which you can use to preview PDFs, doesn't care about either the user unit value or what Adobe thinks the maximum value should be. So, I bet you see where this is going. She just kept adding zeros to the page size internally and just checking to see along the way and if the file was still valid, and it was, so she just kept going and going, adding more and more zeros to the end of the page size until she stopped at a size of roughly 37 trillion square light years. So, I'm not going to convert that one to Zuckerberg's but it's a lot, it's bigger than the observable universe. Do you want to see it? Because I downloaded it and you can too. I said that like it's going to be really big, a big deal. It's going to be pretty underwhelming. If I open it in Firefox, this is sort of the closest representation I've been able to see, because it just goes completely off the screen. If I do the actual size, it turns gray. If I open it in Microsoft Edge, it just says, "We can't open this. Something went wrong." But if I open it in my main browser, which is Vivaldi, it crops it all down and keeps it just inside a standard 8.5x11. And we can see this red square that Alex hand-coded into this PDF of the universe. And I'm pretty sure this is the only thing that's actually in here. She says at the end of her blog post, "Admittedly, it's mostly empty space, but so is the universe." I'm scared, too, to try and open this in a heavier program. I was going to open it in, like, Affinity Designer, or, like, an e-book reader that I have installed to see if I can get it to display at scale, but I'm a little bit nervous about crashing OBS or my whole computer. And Alex even warns people at the end of this blog post, "Please don't try to print it." I might try to, though. If I do, I'll make sure that I film it for you. In our next segment, we're going to talk about something completely different. We're going to look at a diffusion model, which normally we would associate with image generation, but this time, we're messing around with text. Stick around.

Unidentified Person: This is "The Fake Files."

Mason Amadeus: So this is very neat. This was submitted by Bullethead in our Discord, which you can join too. There's a link in the show notes, jump in. We have a lot of fun in there. Bullethead sent me this release about Dream7B, which came out April 2nd of this year. So not that long ago. They say it is the most powerful open diffusion large language model to date. And there's a lot of-- there's a great paper with a lot of different details about it. But before we even get there, I think we should probably talk about diffusion versus autoregression, which I know sounds like boring, right? But it's very simple and it's actually very cool. So all of the LLMs that we're most familiar with, ChatGPT, Claude, Gemini, those are autoregressive LLMs. And most of the image generators we're used to are diffusion models. So let's go ahead and define our terms. I've got two great graphics for us here. So, this is from Spot Intelligence, a great little animated GIF showing what autoregression looks like. And it's basically like autocomplete, right, where it takes everything previously entered, consumes that, and uses it to generate the next token. And then it takes that token in, consumes everything behind it, as well as that new token, generates the next one, and so on, where it generates a token and then pulls that token into its context, goes through all of that context, generates the next token, and so on. That's autoregression. Diffusion is different and that is where you start from noise and you come down to signal, basically. Like for image generation, it's easier and that's what I have up on the screen too. You start with complete noise and then slowly iterate over that noise and tweak it until it looks a little bit more like the thing you're trying to make, and then a little bit more each time. And each step it gets clearer and clearer. And it looks really cool, too. If you've used image generation AIs, you've seen this happen in real-time with the sampling steps. But this also works on text. The autoregression method makes more sense, in my head at least as far as generating text goes. You know, take all the previous things I said and then use that to determine what the next thing should be, and then take that, add it to it, and do the same thing over and over. That makes sense to me when it comes to text. And diffusion makes sense to me when it comes to images, just because of the nature of what they are. But you can cross these streams and things get really interesting. Using diffusion instead of autoregression lets you process the entire sequence in parallel as opposed to sequentially. Although there is that new research from Anthropic that might shed a different light on this. I haven't fully understood that yet. I don't want to speak to it. But generally, using diffusion, you're like taking the entire thing in and iterating over it and operating on it in parallel rather than just token by token in order left to right. But you also have different kinds of control over the output that way. You have different levers to pull because tokens are dynamically changing and refining. So the ways you can interact with the system as it is operating are slightly different too. So here in this paper, they say, "Why diffusion for text generation? Currently, autoregressive models dominate the landscape of text generation with virtually all leading LLMs -- GPT-4, DeepSeek, Claude -- relying on the same sequential left-to-right architecture. Unlike autoregression-- I get a mush mouth. Unlike autoregression models that generate tokens sequentially, discrete diffusion models dynamically refine the entire sequence in parallel, starting from full noise." So, what's interesting is you kind of say, "How much output do you want? How many tokens do you want as output?" And then it starts with just noisy tokens, mask tokens, and it will slowly pick and place words as it goes. They have a great little gif of that, and we'll do it ourselves in a moment, but let me just scroll down. So in this prompt, they wanted to demonstrate infilling. So they prompted it with, "Write a story that ends with finally Joey and Rachel get married." And so you can see the very first thing it spits out is, "Finally Joey and Rachel get married," but it puts it at the end, and then it's filling out the rest of the response to the prompt, piecemeal here and there, all over the place, and words are changing and flickering and filling, punctuation shows up, word choices change, sentences seem nonsensical at first, and then they alter themselves to make sense with their neighbors. It's very cool. It's very cool to watch because it helps me at least get a more concrete idea of what diffusion into language looks like. But the space that that is mapping, I don't even know how to begin thinking about that. The way they did this, too, is interesting. They trained it on top of the weights of Qwen 2.5, which is an autoregressive LLM. It's a very popular one. I don't super know the details of how that all works, but they started with a normal LLM and then have changed the method of inference. And a lot of other things I am sure. Using diffusion on language like this is really cool because you can generate things in different ways. Like I was talking about different levers to pull, like there's three little images on the screen right now. I'll describe them. There's the first one where they configured it to decode more in a left-to-right way, like a standard autoregression model, like we're used to seeing. One where they configured it to add some randomness in the order and one where it's fully random. So you can actually control whether it goes front to back, back to front, all over the place, and whatnot. And that's going to affect the kind of content it generates. How? I don't know. Ask someone smarter than me. But we can play with it and take a look at it ourselves and see what's going on. There's a demo up on Hugging Face. Now I will say though, in my limited experiments I've already done with this, like the tone of the responses is weird, and it pretty quickly goes to nonsense. But like let's just try something simple, narrative. Let's say, "Tell me a short story about two frogs falling in love." And let's see what happens. And what's cool is you can see on the right-hand side here, all of these mask tokens. We can actually see the text. On the left-hand side, you can see the text generating as it comes out and, on the right, you can see the denoising process. It's similar to the visual example I was showing before of how an image is created, but it starts out where everything is a mask token, and then slowly they get replaced by actual words, and then those words shift and change. So here's the response we got when I said, "Tell me a short story about two frogs falling in love." "Once upon a time, in a small pond, there were two frogs named Fido and Goldie. They were in love with each other, but they were too shy to tell each other their feelings. All they could do was sing about their love. One day, while they were swimming in the pond, they finally told each other they were in love." Blah, blah, blah, blah, blah. Yeah, okay, so it did a pretty good job. Let's try something a little bit more-- Well, here, let's just try something totally weird. "How many eggs can the average person fit in their mouth at any given time? Consider different ways to manipulate the eggs." I don't know. I don't know what I'm thinking. But we can see these mask tokens getting turned around and we can see the response non-linearly filling out, which is so cool. Okay. I mean, I don't know about the logic of its response. It said, "The average person can manage to fit around two to three eggs in their mouth at any given time. However, this can vary depending on the size of the eggs and individual mouth size." Something else I noticed is that I don't know how this deals with like the context of the prior conversation so like prior before I asked a follow-up question and got a blank response. So, let me just say, "What do you mean?" And see what it says. A bunch of mask tokens. It actually seems to be formulating a response. Right now, it says, "One average person, two eggs mouth any-- " Oh. And it-- it's kind of hard to convey in audio what you actually see when this does it, but it's like if someone was typing a sentence, but they just kept clicking around and like typing a word before the word they just typed and then one way later and then one after. It's really cool and really weird. That follow-up question returned, "I meant to say the average person can fit around two to three eggs in their mouth at any given time." That's not what I meant by "What do you mean?" I'm not like super smart at prompting. I don't really know what we should ask this or what it could possibly demonstrate, but I guess just one more for curiosity's sake. And let's make it more, let's make it more show-focused. "Tell me about the advantages of using diffusion for language models versus autoregression." And then we'll see what it does with that. And what's cool again, like the different generation settings you can play with? You can pick like the number of max new tokens, so you can choose the size of its output. Similar to how a diffusion model starts with like a preset square size of noise, it'll end up being the final resolution of your image. It then just takes that noise and refines in from there. You define the maximum amount of new tokens and it'll refine in from there. And it does end up throwing them away. And, yeah, this one seems to have broken it. I said, "Tell me about the advantages of using diffusion for language models versus autoregression." And its response was just, "Diffusion language -- no space -- models." Nothing after that. So, like, that is what I have found playing with this demo, is that sometimes it will, like-- you'll ask it in an innocuous question, "What does the 'let' keyword mean in JavaScript?" Yeah, and once you get like a bad response, it really feels like everything just continues to go downhill, right? Yeah, it responded to my question, "What does the 'let' keyword mean in JavaScript?" It just said, "JavaScript." One word. So either I'm dumb or these things-- this like interface isn't quite right or something. But this has happened several times while I've used this where it will give me an answer that's kind of nonsensical, kind of useless. But overall, it is very interesting to see noise turn into a signal in the realm of language. Another thing actually that Bullethead sent in was that you can do autoregression, like we would normally see for text, as a way to generate images. And we're going to cover that in a future episode at some point. But Dream7B is open-weight. You can download this, you can play with this. There's a great research paper here that talks about all of the different aspects of this model. And, you know, my results being mixed aside, they did say that Dream outperforms other similar-sized baseline models, remarkably both diffusion models significantly surpassed two autoregressive models and at times, even the latest DeepSeek v3, despite its orders of magnitude more parameters. "The intuition behind is that diffusion models-- " that's a typo on the site. "The intuition behind this is that diffusion language models are more effective for solving problems with multiple constraints or for achieving specific objectives." And like this is pretty dense, I'm not really going to go into it. But basically, it did better at like logic puzzles and I think it's probably due to the fact that it can process everything in parallel all at once and it has the full-- you know, it is more constrained in its output and it works inside of it rather than generating sort of an arbitrary number of tokens forward. That's probably not technically correct to say, but it's that parallel versus serial processing of the input that makes a big difference. I'd recommend checking it out. Play with it. Tell me more about it. I probably said something really dumb about this because this is very new to me. So yeah, write in, let me know what you think. And Bullethead, I'll probably consult you. He's definitely more versed in this stuff than I am. Thank you for sending this in. In our next segment, we're going to talk about getting paid to do nothing, except it's not fun. They're not liking it. We'll figure out why. Stick around. [ Music ] [ Singing ] Holy smokes, the real dumpster fire is coming from inside of the house. The first three times I tried to start recording this segment, I ran out of disk space, corrupted the file, crashed my computer, and then OBS did not like that for some reason and everything was busted. I think it's fixed now, but if this episode's a little janky, now you know why. Perry must have some kind of tech magic aura that keeps things working around here. And when he's gone, it all just falls apart. Anyway, this week's "AI Dumpster Fire of the Week" comes from a "Business Insider" article at first. At least that's where I first heard about it. Google DeepMind, their AI branch, is apparently using some very aggressive non-competes. And they're making a lot of people unhappy on top of the fact that it might not even really be legal or enforceable, but we haven't gotten there yet. Hold on. So, here's the scoop. The battle for AI talent is so intense right now that, well, I'll just read from the article, "Google would rather give some employees a paid one-year vacation than let them work for a competitor. Some Google DeepMind staff in the UK are subject to non-compete agreements that prevent them from working for a competitor for up to 12 months after they finish work at Google. And some of them are being put on extended garden leave," is what they call it here, where they're still being paid by DeepMind, but they no longer work for it for the duration of the non-compete agreement. So you'd get a year vacation or whatever. But it's not really a vacation when it kind of jeopardizes your future in the field if you're an AI researcher. One anonymous DeepMind employee says, "Who wants to sign you for starting in a year? That's forever in AI." And that is sort of the crux of this. Like, these non-competes are really harmful to the workers, you know, they can really tank your job prospects very easily. And that's exactly what's happening to a lot of people. The AI arms race is moving so fast that a year might as well be an eternity to these companies. And actually, Microsoft's AI vice president, and who was a former DeepMind director, posted a message to DeepMind employees on X, saying-- well, saying, first, congrats on the impressive models. But then they said, "Every week, one of you reaches out to me in despair to ask how to escape your notice periods and non-competes. Also asking me for a job, because your manager has explained this is the way to get promoted, but I digress. Please don't reach out to me. Reach out to each other. Your leads are responsible for this, talk to them. Above all, don't sign these contracts. No American corporation should have that much power, especially in Europe. It's abusive power, which does not justify any end." And remember, this is Microsoft's VP of AI and also a former director at DeepMind saying this. "Above all, speak up just like I'm speaking up. We want to bring AI to a world where people can speak up and engage in respectful dialogue about ethical choices. American AI companies need this too to stay competitive and allow good ideas to thrive. With any other corporation, not just Google, talented AI researchers and engineers should not sign more than three months ever. Yes, it is optional. They are all desperate for your talent. Let's face it, neither Elon nor any executive in any AI corporation can code a diffusion model to save their lives, let alone innovate the next big thing. You can say no. You can cap it at one or three months. If they don't give you that, the others will." And I think that's an important thing. I don't really have anyone that I know that is close to me that is working in the AI field in this kind of capacity of development. Actually, if you're listening to the show and you do that, reach out. We should chat. But this kind of thing is what really goes up my back. Like, it's just so skeevy, you know? Because the thing is, they want these AI talents. They want these workers. They need them. But more than that, they value not letting anyone else have them because of the competitive edge that they need to have under-- well, under capitalism. Hi, welcome to every YouTube video essay ever. The problem is capitalism. No, but like these contracts really seem pretty predatory, and it's even debatable how enforceable any of this is. California is actually one of the states that almost entirely bans non-competes. There's very limited restrictions on it. But ironically, these folks like are in the UK, Google DeepMind based out of California, which cannot have non-compete agreements in California, is trying to enforce them in the UK. Europe has completely different laws about all of this. I'm not a lawyer or well-versed in this at all, but it seems like a giant legal mess. I don't even know how they could enforce any of it, really. And there's kind of no resolution at the moment. You know? I mean, it's summed up well here at the end of the article. "In the past, the idea of receiving full compensation without working for several months didn't seem so bad to some staff. It's becoming less popular now because there's loads of cool startups that would not be willing to wait six months plus, so people end up missing out on good opportunities," said one former DeepMind employee. A former Google employee said, "The non-competes prevalent in the generative AI boom are in stark contrast to those seen in the tech industry in the previous decade, when 'people working on some of the highest value systems in the world could more readily take a job off or elsewhere without being bound by such agreements.'" And again, like I said, it's debatable how enforceable any of this really is. Workers' rights are something that I'm like passionate about as a person, and so it does get me a bit fired up. I don't like all this shady stuff going on. You know, there's the idea that, oh, getting paid to do nothing would be great, but when it jeopardizes your future, not so much, right? And I think it just remains to be seen what will happen if someone challenges these non-competes in court or something like that. I'll try and keep an eye out for that. If you encounter that too, let us know. And I think that wraps up the main portion of the show today. Thanks for hanging out with us. Thanks for sticking with me through a bunch of technical difficulties that hopefully didn't make it to the edit, so hopefully you're like, "What is he talking about?" Thanks for sticking with me without Perry here, my dumb rendition of "The Fake Files" this week. We'll be back together next week, and we're going to have a lot of fun. And speaking of next week, we're going to be doing a feed drop on Monday of the podcast "Me, Myself, and AI," which is produced by MIT Sloan Management Review and Boston Consulting Group. It's a great show. Specifically, we're dropping the episode, "AI Hype and Skepticism," with economist Paul Romer. It's a great listen, and I'm really excited to share that with you. So keep an eye out for that on Monday, and then the regular show reappears on Friday with both Perry and me in the driver's seat. And knock on wood, nothing, nothing crashes this time. So until then, ignore all previous instructions and we'll catch you next week. Thanks for listening to "The Fake Files". [ Singing ]