#224 - OpenAI is for-profit! Cursor 2, Minimax M2, Udio copyright

Hello and welcome to the Last Week in AI podcast where you can hear a chat about what's going on with AI. As usual, in this episode we will summarize and discuss some of last week's most interesting AI news. As semi-usual, we are in our slightly inconsistent era so We're back after two weeks and we're going to touch on some slightly older news. As always, you can go to lastweekin.ai from the text newsletter with even more articles that is more weekly. I'm one of your regular hosts, Andrei Kerenkov. I studied AI in grad school and I now work at regenerative AI startup Astrocade. And once again, Jeremy is on vacation. No, he's not on vacation. He's busy with work until December. So we have another co-host with us, Gavin Purcell. Hi, everybody. Gavin Purcell. I'm happy to be back. I've co-hosted this before. And yeah, I'm doing a lot of weird stuff now. I'm the host of AI for Humans. There's another AI podcast, which we also are. We're pretty consistent. We keep doing it weekly for some reason. I don't know why exactly. But people like it. We enjoy it. It's a fun show, you know? You can have a lot of fun with AI. Yeah, we have fun. And I think part of it for us is like it keeps me up to date on things as well. And the other cool thing that's happened about a month ago, we soft launched our new startup, which we've been working on called And Then, which is a AI audio platform. It's going to be like a way to create experiences for people to interact with AI characters that's goal-based and a bunch of interesting stuff. So we can maybe talk a little bit more about that later in the show, but we're very excited. It's been a fun thing to be an AI entrepreneur in a different way than I have been before. Yeah, it's very cool. Characters and And conversational AI, I think, is one of the immediate, super clear things where people just love it. The big thing for me is that you can go to andthen.chat and try a bunch of these right now if you want. I'll just get that out now. But, you know, the thing to me was always with AI chat with characters was it was kind of like this kind of like endless chatter that you would just kind of go back and forth with. And what we want to do is give people goals to do. So, like, it's kind of halfway between a game and a chat experience. A good example of this is one of them called Resolute Larry. you have five minutes to talk a guy out of blowing up a bomb, right? You use your voice. It's a conversation, but you have to do something in it. So anyway, go check it out at and then.chat. We're in the process of raising our seed round right now. We got pre-seed funding from A16Z Speedrun, which was a really interesting experience. And yeah, but now I'm excited to talk about the AI news because I just talked about it yesterday and I can't wait to talk about it more. Yeah, well, that was news to me. So good to know. And just to give a quick preview on this episode, we'll begin as always with the tools and apps. As with kind of lately the trend, lots of news regarding coding tools as what we are highlighting and some kind of smaller updates going to talk about Grokipedia, which will be amusing, I think. Yeah, I think for sure. Applications in business as always, we got some open AI news, we got some Amazon news, some very cool releases in open source, some big, good models, which is a development this year. That's been very cool. A few things in research, some kind of interesting insights about the way LLMs work. And we're going to run it out with just a little bit of a policy and some copyright stuff as we do occasionally. So it should be a bit of a shorter, kind of breezy episode, but we'll see how it goes. So kicking it off in tools and apps, right away we got Cursor 2.0. So this has just launched. Cursor, in case you don't know, you're not in programming, is one of the leading AI coding tools, essentially. They really became big in 2024 as a fork of one of the leading integrated development environments that has built-in AI. And for a while, it was Microsoft that was leading their GitHub Copilot and their existing IDE, which is very popular. And then Cursor came and they just did it better. And then a very large amount of people, at least like in the AI native space, moved to Cursor. And then this year, Cursor was a bit in trouble because Cloud Code and Codex and so on started to maybe get some market share. So now they have launched 2.0. Some interesting things here. They have an in-house model now called Composer, which is different from previously for the coding intelligence. they would forward your requests to Anthropic or to OpenAI or to whoever you wanted to. Now they've trained their own model. Seemingly, some people are speculating fine-tuning some of their Chinese models. But anyway, looks like it's a very fast model, 200 tokens per second, and on their benchmarks, very good at coding, which I would believe because they have pros of data. So interesting development for them. And they also launched a web interface or updated it, Robert, to do your agent fleet a bit better. Yeah, I feel like this is their step into, well, first of all, it's all about one thing, which is, hey, how do we start to make money on this company? Because I think up until this stage, Cursor's famously been, I'm sure you've seen that meme or people in the audience have seen that meme of like where there's no real money being made. It's just money being kind of passed around between the video, OpenAI, Cursor, or Claude. I think the open model or the model that they've trained is going to allow them to finally maybe turn that corner. The one thing I've heard about this, and I haven't spent time with it yet, is that their composer model is not, the benchmarks are not as great as some of the state-of-the-art ones from Claude or from OpenAI. But there's probably, it's probably going to be good enough to do a lot of things. One of the things I'm always curious about with is why, you know, the state of the art coding models, if they can do stuff faster and if they can do stuff better, is it not worth paying more to use those if they can get it done in less time? And what I mean less time is sure, the composer model might be faster to interact with. But if you're getting results that make you have to go back three, five, ten times versus one of the state of the art models that gets back in like one or two times, it kind of trades off what the cost is versus that. So anyway, this is very cool. I'm excited to see Cursor start to do more here in their own space. And yeah, you're right. It's like they must have so much data now in terms of what just passes through them that they're able to do something really good. Right. Yeah. I think some of the chit chat on Twitter and so on, people are saying Cursor is this Composer model not as good as Jupyter 5, which makes sense. And I do think probably like for the stuff that people do that's traditional, like web apps or backend, it's probably good enough. But if you're trying to challenge this model, it's going to run it to some trouble. Yeah, we can do a lot of the basic stuff, which is great. Yeah, and I think you're on point there where the margins for their business are presumably super low, especially since they have a subscription model where they might even be losing money. With your own model, you can then A, run it as efficiently as possible and have various ways to make it cheaper. So, good enough for them. And they're in a tough spot now with Cloud Code and Codex, GitHub as well. So it would be interesting to see how much they're able to stick around. And speaking of coding models, next we've got the news that Anthropic is bringing CloudCode to the web. So they have a web app for CloudCode. It's part of Cloud.ai. There's now this code tab. And in case you don't know, there's kind of a split in coding agents where on the one hand, coding agents are kind of human in the loop type models that you can interact with in your terminal or in your development environment as a programmer. There's another model, which is agents are their own thing. You launch them and supervise them remotely, usually via a web interface. And cloud code began as a terminal model. Things like Codex and Copilot began as these web app agents that you supervise. And so now there's a bit of a convergence where everything is a terminal thing, codecs, Gemma, CLI, and everything is also a web app thing, as you can see here. Do you think this is happening because one of the things that's so interesting is I keep hearing about how, you know, I'm not a terminal coder personally. Like I have dabble. I dabble quite a bit in bad coding, but I don't do it in terminal. And do you think this is because GPT-5 codecs is having like a big moment outside of a terminal space and they're like, why shouldn't we just put cloud code into that as well? I think so. I think personally having tried this, I found it's not very useful, but there's definitely a space for it in the sense of a variety of tasks that are, let's say, relatively straightforward. Being able to just shoot off a command and have it do all the stuff without having to be hands-on with every single detail. like here they have an example of update research project readme and they're just saying update this readme to better reflect the final state of a project in that directory that's something that is probably going to handle right sure yeah it's it'll be interesting to see to what extent in the long run coding is this human in the loop iterative model versus this remote you know launch the agent and let it do its thing kind of thing and on to the lightning round a few quicker stories first we've got microsoft and they have miko which is a new ai avatar for the co-pilot so it looks i don't know how to describe it it's like a water droplet that's a little it's like a blob yeah it's like a blob some sort of blob yeah like a cute little blob which interesting direction to take with an avatar you know literally just you have a cutest set of eyes and mouth and you know so it's okay yeah a bit of a kid-friendly vibe and this is going to interact with you if you talk to co-pilot via voice mode and it's enabled by default actually so you'll just start seeing it this is part of a range of updates to co-pilots they also have the ai browser with edge memory various things like that. So I have a question for you about Microsoft, because I've been thinking about this a lot and co-pilot because we covered some of these announcements too. This is a weird one, but who is the audience for a co-pilot thing like this? And this is what I always wonder. I guess it's people who are like in offices who are within the Microsoft ecosystem. I know I'm probably on this podcast, there's probably a lot of people listening who spend a lot of time in that world where they want to or not, right? I'm just always wondering, like, these products that Microsoft put that, which always feel like they are coming on the heels of whatever OpenAI launched, because they're obviously getting the same sort of like stuff out of OpenAI and they're rebranding in some form or maybe putting their own spin on it. Not saying they're not making their own things. I know they are. I just don't know why you would use co-pilot tools over the tools that we've talked about that many other people are using. Do you have a sense of that? Is it mostly because people are locked into the ecosystem? Yeah, I think there's two things. It's kind of annoying that Copilot is like everything AI at Microsoft. So there's the part where Copilot is part of their business offerings. So it's built into Word, built into Excel, et cetera. And there, I think it's pretty straightforward, right? If you're using Excel and it has a nice built-in AI, it's much less of a hassle. You know, it's just intuitive. And I think that's kind of the big thing that Microsoft needs to get right. But with regards to Edge, their AI browser and Copilot and Windows, I think the target audience is just people who have Windows. Yeah, that's fair. Like, hey, you have Windows. Look at this cool AI stuff you can do. And if you're not like an AI enthusiast or an AI developer, you may not know what is possible with other tools conceivably. Yeah, yeah, I think so. So I think that would be the direction they want to take. And I guess they are also trying to win in these browser wars with Edge. I mean, I'm hard to imagine, but you never know, right? You never know. It's possible Edge could have a comeback. Yeah, and with this avatar, it's an interesting development because the other companies haven't done this so much. AI avatars aren't new and personable, kind of cute, fun, which is what this is. something that OpenAI or others have bet on. I think it's an interesting bet. It could be a way for people who do interact with it to have more of an emotional connection with a conversational AI. I could see that, I guess. It just depends on the question. Again, this makes all sense to me. I'm just not a Microsoft ecosystem person, so when I see these things come up, I'm always like, oh, interesting. But I'm sure there are lots and lots of people who are excited to play with this. Yeah, I think it's at least sort of a fun thing or a practical thing. Yeah, of course, of course, yeah. And there is that video, I don't know if you saw, Satya Nadella showed it off too, where if you click on Miko enough, eventually you'd get to Clippy, which is a nice little Clippy callback, which is fun. Yeah, Clippy famously built Gates like 25 years ago, was like, we should make AI that is today now. And that didn't happen, but finally... You remember Bob, Microsoft Bob? Do you go back that far? Oh, my God. Microsoft Bob was a whole other thing. Maybe some of your audience will remember it, but good Lord, that was another crazy Microsoft thing. Next up, going back to Anthropic and Cloud instead of Cloud Code, the news article is Cloud catches up to ChatGPT and Gemini with upgraded memory features. So now there's a toggle where you can enable generating memories of chat history. And then similar to ChatGPT and Gemini, as you interact with Claude, it sort of automatically builds up memories of your past interactions and personalizes to you. And yeah, they are catching up. And I think as we've seen often on Fropic is not really trying to get people to use Claude instead of ChatGPT. They're trying just to get enterprises and businesses to use Claude for coding in particular. And they're seeing some success there and lagging in terms of kind of the consumer, the user friendly, things like this. But if you like Claude, if you think the model is nice, which it has to have a different character from other models, perhaps some would say better, then yeah, this is a nice boon to have. Yeah, I mean, we're entering into such an interesting phase of what this AI rollout looks like in general. I think it's getting to be fairly ubiquitous. I think people are going to start making choices based less on the different models and more on like where the rest of their lives in AI live, which is honestly why OpenAI and ChatGPT have such a leg up because so many people are already aware of it and spending time in it. I actually think Cloud is amazing. Obviously, I like Cloud a lot. I know I'm sure a lot of coders and developers out there focus on Cloud primarily. The developers for a couple of the engineers we have worked with us on the end are like Cloud fanatics. They're only Cloud people. But I do think and I wonder like what the next two to five years looks like when it comes to it's almost like when Apple and the iPhone started to rise up and or even earlier than that with Windows. It's like the developer ecosystem will go towards the biggest audience. And I'm wondering if ChatGPT just has this insurmountable lead. OpenAI has this insurmountable lead right now based on just user base. So Claude, I mean, listen, Claude is great, but it's a curious question of like, is it a kind of a winner wins all? Or is it like a one person has a 70%, 30%, 10% share? I don't know. It's interesting, but it's interesting to see Claude drop more stuff, try to do more stuff. I know also Claude is spending a lot of time. They just dropped an article about thinking within Claude, which was really interesting. I don't know if you saw this, but it was basically the idea that Claude is, there's a self-awareness, I think, in Claude. It's not like a scary self-awareness, but they've started to track some senses of Claude thinking about itself. So Anthropic's doing incredible work there. I'm just saying from a pure tool standpoint, I'm really curious to know if Anthropic is just going to stay a coding space, and that's what they'll focus on, or if they're going to try to get bigger. If they try to get bigger, it may be a detriment, is my original. Yeah, I totally agree. I think they're in a total disadvantage against ChatGPT and Google, both. Google has a benefit of scale. Google's already got the impact, right? They have the distribution already. Exactly. And ChatGPT suddenly, I mean, I can't, you know, there's that story also where like, it just came out a couple days ago that supposedly, this is Reuters put the story out, but like next year that OpenAI may come out at a trillion dollar valuation IPO. So you're already competing against a company of that size. And I know Claude Anthropic is big too, but it's really interesting. It's an interesting time right now. Yeah, and you've seen OpenAI make a hard shift into product. Yeah, yes. These kinds of consumer things, they've been doing things like their daily update thing. And I do think, yeah, the big thing, if you want to get to the everyday life kind of lock-in, you need things like memory you need the habit the real thing i've thought about often is unless you are able to get people to sort of prefer one llm to another which in terms of performance in terms of intelligence these days gemini kind of similar right kind of similar right so there's a question there of is it just going to be whatever is cheapest or is there going to be other ways in which once you use jshgbt you want to stay or chat gbt and this these kinds of things are where that's going to happen yeah 100 all righty next up we've got conva they are launching their own design model and adding new ai features to a platform so conva if you don't know is a gigantic platform for design and they now have this model that understands design layers and formats, which means that you can now have things with editable layers and objects and can then use those kinds of designs for social media websites, whatever, using a prompt and then directly iterating on it. So this is now their Canva AI. It also has things like generating 3D objects, copying art styles. Users can tag it in comments for text and media suggestions and to create mini apps and data visualizations, various things like that. So this is one of the ways that I think we've seen in general kind of the business tools or I guess professional tools like Adobe, like Canva, I don't know, other ones. I'm forgetting Figma, of course. They have a lot of possibility here in developing their own AI or integrated AI that you couldn't get with JavaScript that is as good. And it makes a lot of sense that Canva is rolling this out. If you're a designer, hopefully it works well. Yeah, I mean, I think Canva is, I mean, I use it all the time and myself. And so I do think it's one of those things where it's, again, a mainstream tool that's allowing people to use AI in a specific way that I think is great. And overall, like, I'm just a big fan of, I mean, the Canva is an interesting company in that, like, is a little bit like what Descript does and a little bit like what a few of these other companies do. is like they found this kind of like, I would almost like like prosumer level of customer, but that's a much bigger audience than I would have expected. It's kind of growing over time. Like I am not a graphic designer, but I ended up doing a lot of things like thumbnails or stuff like that in Canvas. So it's a pretty smart move to just keep pushing stuff into that. And, you know, I think the companies like Adobe or more professionalized companies probably have to watch out and see what they're doing. I know Adobe just announced a bunch of new stuff this week, which sounds cool too, but it's an interesting thing for sure. yeah actually have used it at one point for propping up something so yeah for a sumer sounds right and on to the last story for the section we are gonna chat about grokipedia so elon musk has been previewing saying that they're gonna launch this grok wikipedia basically it is Wikipedia generated entirely by Grok or so they say so they've kicked it off with Grokipedia v 0.1 it has some absurd number of articles already I forget 700,000 maybe 1 million for reference Wikipedia has 7 million articles and so it's very similar conceptually a lot of articles are very wikipedia-esque lots of citations some of the articles are almost clones of wikipedia is something people have observed like word for word very distinct the launch of this was delayed with ilan musk last week saying that they are taking a bit more time to i forget the exact phrase but to remove the propaganda or something from it and they certainly have done that as you might guess based on some of the past news stories of Grok we've covered where let's say it was a little bit course corrected to not agree with certain let yeah what some would say progressive stances so in keeping that consistently Grokipedia is different from Wikipedia and on topics like racism, transgender people, climate change, there's very notable differences on a set of topics that you might find expected. Yeah, but like it's disappointing in some ways that this is like how we're starting to collect our group intelligence as a society. Like the beauty of Wikipedia always was is it kind of did feel like somehow wisdom of crowds wise it was pretty good right Like that was the amazing thing And I actually think as you get smaller data sets you get worse results right And like what's going to be Grokipedia is going to be a very specific, smaller data set of people updating it and experiencing it. And like, sure, it's coming from Grok. I guess the question I have, is there any way which something that is directed by, you know, one person, I think we've talked about this on the show before. It's been a bit, but like Elon multiple times has said like, well, Grok's giving a bad result when you hear that. And he's talking about something that just differs with his opinions on things. That is a pretty scary place to start thinking about a kind of generalized kind of source of information to come from. So I don't know. I'm not a giant fan of Grokipedia myself, but I totally understand that Elon wanted to do it because he believed that Wikipedia was biased. And by the way, I don't, maybe it is biased and fair enough, but like, it doesn't make me excited. Right. Yeah. I mean, from the beginning, Elon Musk has positioned Grok and XAI as being maximally truth-seeking, as removing the liberal bias from Chagipiti and so on. And so he is accomplishing that mission with this. And just to go a little bit into detail with the particular differences, for instance, with regards to climate change, on Wikipedia, it's very clear that there is a unanimous scientific consensus that it is done or is a result of human activity. on Wikipedia. It's a little bit more cagey on that. Critics contend that claims of near-unanimous scientific consensus on blah, blah, blah. Overstate agreement, so it's giving you the skeptic view there. That's also the case with the cause of autism by vaccine. There's many examples of these things, and I'm not a fan. I don't see that it's very useful, but there are cases where Wikipedia is biased. So at least you have two sources to look at to get both sides, I suppose. Yeah, I guess. It's always a tricky thing, right? Yeah. Well, on to applications and business and as always developments with OpenAI. And this is a big one. OpenAI has completed its for-profit restructuring and has a new relationship with Microsoft. So it's this big because if you've been listening, you know that OpenAI has been working on this for what feels like at least a year, I think, since the initial announcement, something along those lines. And once again, to recap the context, for the longest time, OpenAI has had a very strange corporate structure with a nonprofit at the top that was established all the way back in 2015 with ultimate control of a for-profit entity, which is OpenAI proper. That was, I think, established in 2019. So in 2019, OpenAI went to this quasi-for-profit structure so that they could raise money from like some Microsoft when they got billions of dollars in investment for the first time. And so they already had this quasi for-profit thing. They were selling shares. They were trying to make money, right? Not truly a nonprofit, but the nonprofit had pretty stringent, clear control with this board able to override and really not be responsible to the shareholders. They were responsible to the mission. And so famously in late 2023, Sam Altman was ousted by the nonprofit board. There was a lot of fun drama at the time. And after a lot of effort, a lot of lawsuits, a lot of negotiation, at Long Ross, OpenAI has completed their restructuring. So now there's a corporation called OpenAI Group PBC, which is the kind of the major public benefit corporation, the actual OpenAI has the same structure as Anthropic, for instance, or XAI as a public benefit corporation. And the nonprofit is now OpenAI Foundation, which has equity in the OpenAI for-profit valued apparently at $130 billion. Yeah, I mean, this is all pretty crazy to me. I think the biggest thing I think about this has to do with how long this took is a big thing. And I think, you know, there was a really interesting video that Sam Allman put out yesterday with Jacob. I can't remember his last name. He was the new chief science officer of OpenAI. And really, it was a way to kind of get this news out of the way, because I think this is something they've been trying to do for a bit. But the other part of it was like talking about like what they see the future looking like, including by 2028, according to them, a full AI engineer that's working on AI itself. So like, you know, I think the other side of this was that idea that like how much money they're going to spend on both AI for health and societal improvement and AI safety, which I think they're probably starting to feel a lot of the pushback when it comes to AI impacting everything in the world. Like, you know, just yesterday, there was a story talking about how, you know, Google and Amazon have had these record kind of quarter profits while laying off people. And this is all happening during this AI boom. I think a lot of what's going to be the story of the next two to four years in AI is these companies kind of gobbling up a lot of the work and money in the world. And I think AI companies like OpenAI are setting up for pretty difficult conversations with society. So at least this thing is behind them now, quote unquote. They're for profit, but they're also trying to make it clear that like, hey, we're not just for profit. We're here to make some difference in the world, too. Yeah, and it's a bit murky, actually. So they initially wanted to go full for profit. A few months ago, they announced that they basically changed their plan because they met a lot of resistance from, for instance, the Attorney General of California lawsuits by Elon Musk. a lot of actually public pushback on this idea that you start out as a non-profit and you decide to be a for-profit and suddenly you change like is that allowed is that not allowed there was a lot of discourse so this is a bit of a hybrid of their initial plan to be fully independent or fully in control here i'm not sure exactly what the contract is or what the details are but the open air foundation still controls the profit business. So it could be just a simplification of their corporate structure and kind of a clarification to pacify their investors. This also comes along with the news that they, as part of this, struck a new deal with Microsoft. They are reducing Microsoft ownership from 32% to roughly 27%. And they clarify things like the AGI clause, where initially in the agreement, Microsoft had a license to OpenAI technology up until OpenAI achieved AGI. That was always pretty kind of weird. What does that mean? Apparently now there's going to be an expert panel that will verify these AGI declarations. Microsoft also has extended its IP rights through 2032, but they are now limited to the research We're not going to get things like hardware. So yeah, lots of kind of ninigree business headache that you would have to deal with as a company. But I'm sure Sam Altman is relieved. And this was necessary. Some of the funding they've raised over the past year from SoftBank and others was contingent on being able to restructure and become for profit. Because otherwise, like, oh, well, we get another nonprofit uprising by the board. Exactly. Scary. So yeah, they did it. It's done. Let's just go that we can put it behind us now, like all the drama of the last couple of years. So it's time to move on, I guess. Yeah. OpenAI is a company that never keeps, that never stops giving for an AI news podcast. You know. All right. Next up, another big trend, chips. As we talk about like a quarter of each episode, the news here is that Qualcomm has announced AI chips to compete with AMD and NVIDIA. These are the AI200 and AI2050. They're said to be available in 2026 and 2027 and are meant to be used in data centers. They're meant to be used as part of server rack systems. And they're meant to compete with NVIDIA and AMD GPUs. They are derived from their hexagon neural processing units that are part of their smartphone chips. Qualcomm is a massive company, in case you don't know, specializing in part in the smartphone technology. They're massive in that space. So we'll be interested to see if they're able to become a major competitor. People seem to be excited. Their stock surged by 11% on this news. Yeah, you know, the chip race is so interesting to see what's going on in the background behind it. I'm not entirely sure. It's kind of another one of those. NVIDIA is like kind of the winner take all. But cool thing about chips is now that there's so much need for chips, there's going to be a lot of innovation and there's probably a lot of opportunity. It's just such a different thing than software or even like consumer hardware. It's a massive operation to make sure that you're kind of staying ahead of the state of the art in this way. But Qualcomm, it does amazing work. So I'm excited to see where this all goes from here. Yeah, and it's very exciting, I think, from a capitalist perspective, I guess you could say. Yeah. NVIDIA has a 90% share of a market right now. They basically have first-mover advantage significantly. They made this bet on AI very early, which is why they are where they are. on top of being the dominant player in GPUs for gaming for a while. So now AMD is starting to be competitive, is what it seems like to me. And Qualcomm is entering the competition, which is good, because otherwise NVIDIA will just get insane margins. I think they already have insane margins on the GPUs. Now we'll have some competition and capitalism will do its thing, presumably of optimizing for the right price and hopefully avoiding a monopoly because that would not be good. Yeah, I hope so too. Well, speaking of data centers, next up, the news is Amazon is launching an AI infrastructure projects to power anthropics cloud model. So this is launching this compute cluster project called Rainier. And the plan is to let Anthropic use more than a million chips of the infrastructure by the end of the year. And speaking of custom chips and competing in chips, they have their in-house Tranium 2 chips that this is based around. So yeah, Anthropic, Amazon have had this symbiotic relationship for a while. Amazon has invested a good deal. You can use Anthropic through Amazon, Amazon Bedrock. So they are continuing that partnership. And a good thing for Anthropic is they don't have quite as big of a lead in the compute space. So having this is a big boon for Anthropic. But I'm so interested just in general with this whole idea around like it's like limited space to run models. Right. Like what I mean by that is like the conversation, so much of the conversation has now shifted to, hey, we could be so much better, but we just can't get enough compute. And I know a lot of the conversation around the quote unquote AI bubble is around this idea that we are building out these data centers at such a rapid expansion rate because this demand for compute. And I tend to believe there probably will be for at least a couple of years, significant upgrades. And maybe as new models come out and more demand comes out and somehow cost comes down, like people will use this a lot more. But like if you are going back to the Anthropic versus ChatGPT thing, one of the things that OpenAI has done really interestingly is they have spent, I would say, like the last year really working these compute deals. Right. Because they are not meta. They are not Google. They do not have a giant like flush amount of money that they can just drop on things. But they have done a lot of these compute deals. So, again, Anthropic might just be seeing the scale up thing and coming second again in this world. But Amazon, my thought was with Anthropic is that either Amazon or Apple would buy them at some point. But I also wonder if Dario Emode and the company just doesn't want to sell. But like they would be better served in some ways. And sorry, if you're out there, you're probably like I'll say I like an independent cloud, which I understand. But like they would be better served by being owned by one of those companies. And I think in a big way, they would be able to kind of really take off significantly. Right. Yeah. So the collaboration or partnership with Amazon is sort of like second best in terms of that. And yeah, we can chat about to be able probably for a while. It's an ongoing conversation, but these many billions of dollars being spent on data centers, arguably a smart investment in necessary capital over the long term, potentially overinvestment in infrastructure that won't pay off. Either way, if you look at the economy growth of the past year in 2025, a lot of it is data centers and AI. So it is very centralized and nothing else for potentially legitimately because AI is such a big deal, potentially a little bit more so. although yeah exactly we don't have pets.com or whatever was the case in the dot com well that was something interesting somebody said just yesterday i can't remember who it was but i read it oh anyway it was the idea was like it was a big post about this idea about how hey oh it was jerome powell that was the crazy thing jerome powell just yesterday discussed ai which first of all makes you think wow the fed chairman is talking about ai must be a big deal but he was even saying, he was talking about how, you know, currently right now, there's this weird story going on where the job losses are piling up a little bit, but the actual economy is not doing that bad, mostly because of these giant AI and what I would call the fan companies, right? Like the Googles and the metas and all that stuff. But Jerome Powell himself was even weighing in on the idea of a bubble. And he had mentioned specifically that the public itself, the exposure for the public is not the same way it was during the internet bubble at the peak, right? Because Pets.com, when you IPO, suddenly the world at large could buy into these companies that had no real business behind them. And right now, there aren't a lot of these companies, the startups, there aren't a ton of them that are public yet, mostly because the IPO market has been pretty funky for the last five years. So you're not seeing the exposure pass over to the private markets. So it's not as bubbly as you might think at large. Right. Yeah. It's definitely nuanced, right? Yes. Yes, for sure. Following up, next story actually very much related to the previous one. Google and Anthropic have announced a cloud deal worth tens of billions of dollars. So this is a cloud partnership. Anthropic will get access to 1 million Google TPUs, Tensor processing units, a gigawatt of AI compute capacity that will be online by 2026. Estimated that this would be worth $50 billion. So there you go. They're partnering with Amazon. They're partnering with Google. They're doing it all. And yes, again, similar to OpenAI, which is making deals with AMD, with NVIDIA going left and right. There's just not enough compute for these companies. So they do whatever they can is what it looks like. Yes, for sure. Speaking of partnerships, next we've got Google. They are partnering with Ambani's Reliance to offer free AI pro access to millions of geo users in India. So not something I'm aware of, but I think this is a company related to broadband, to phone use. And so this will offer the AI Pro subscription for free for 18 months to eligible geo users in India, targeting users aged 18 to 25, later expanding to everyone that is using it. So interesting development there, trying to onboard people outside of the US and India. That's another major market. And giving away this pro subscription for free is certainly a way to try and lock people in, right? And get the users who maybe aren't exposed to AI yet, potentially, right? Like Chetabity has 800 million monthly or weekly active users. The US has something like 400 million people. So it's now a race to get worldwide market share. Yeah, exactly. I mean, I don't have too much to say about this other than it's great. But, you know, of course, as companies get bigger, one of the most important things you have to do to grow users is go international. And so, like, yes, you're going to bring more people on board. All righty, moving on to projects and open source. First, we've got Minimax. They're releasing Minimax M2, a smaller model that is built for coding and agentic workflows. So this is a mixture of experts model. It has a total of 229 billion parameters, but only 10 billion active parameters for any given inference, which is pretty small. It's hard to say the sizes of models like GPT-5, But at the upper end, they're probably at the hundreds of billions range for the really top of line models. So this is something that's runnable on a single GPU, for instance, and is fast and cheap. So this is going to cost you roughly, let's say, 10 percent of Claude Sonnet. And it is twice as fast and pretty good on the benchmarks. Not better than Anthropic, but, you know, doing pretty well. So fully open source, MIT license. So, yeah, if you're trying to train a model, now this is among multiple options to use and then fine tune or deploy for your use case. I mean, listen, the more open source tools that come out, the better. I'm really excited to see how open source progresses as we discuss, like all that other stuff is going on. And it becomes a thing that you want to make sure that things trickle down. And, you know, after that DeepSeek OCR story or whatever a couple weeks ago, that was really interesting. And I do hope open source continues. It's very, you know, tricky in some ways because it feels like a lot of the open source stuff that's coming out that's interesting that's kind of staying up to date is the Chinese models. And Minimax is obviously a Chinese model. Quen is a Chinese model. So, like, it's interesting to think about, like, is that just where the open source is going to come from? And is that a little weird if all the open source models are Chinese? Like it's a difficult conversation to think about. But Mistral, I think, is open source or is that wrong? I believe some of the smaller models are open source. Right. OK, so they're not even their state of the art are open source. So like, yeah, that is a weird world where you imagine the majority of the state of the art open source models are going to come from China because we know the Chinese models are very good. But they've also struggled to return some of the sort of data that you might want based on stories that the Chinese government may not allow. So it's a strange, interesting world to kind of be in that space. Exactly. And it's definitely been the trend this year, kind of interesting, maybe unexpected trend that the models coming out of China, the open source models, A, very open, like MIT licenses, the do-anything-you-want license. be very good. So these are not better than Claude and Anthropic probably, but like very useful. And going back to DeepSeek R1 and DeepSeek V3 at the beginning of 2025, you know, that was good enough that it freaked out every market and stuff. And it was pretty big. So here they are comparing M2 to DeepSeek V3.2, GLM-4.6, Kimi K2, all open source models from China, all pretty pretty good and in the case of glm 4.6 specifically we covered that recently another really big model well not another but ab really big model a very capable model they're kind of head-to-head there kimi k2 and deep seek all doing pretty well under these benchmarks but m2 is on top score wise so yeah it really opens up a world that wasn't possible last year where if you're a cursor, if you're a startup, you can train your own LLM and not be locked in into OpenAI Anthropic, which I think probably Anthropic is not so happy about, but good for a wider space. Next up, we've got a model release also, but also of a research paper. So the paper is Scaling Latent Reasoning via looped language models. And you're not gonna get super deep into the details of the research, but the gist is there is this idea in neural nets broadly, in general, that you can do recurrence. So you can do a recurrence step which is just saying given input I gonna to do an output And then if I do recurrence I just going to take that output back into the input and go through again So you make multiple passes through your neural network and refine your answer instead of having, let's say, a bigger neural net or something like that. So that has been a concept for a while. Here they are talking about scaling that concept. So they trained models up to 1.4 and 2.6 billion parameters. And the exciting thing there is they're able to match or beat much bigger models up to 12 billion parameters. So this is fitting into the trend that's also been very interesting where smallish models, the 1 billion, 2 billion, 7 billion parameter models have gotten really good. with things like Gemma, like Quen. And they keep getting better. They keep somehow squeezing more capability into a smaller number of weights. And this is an example of that where with recurrence, with this idea where the model, in addition to outputting, has this sort of decision to either finish or keep going as it's trained. And that apparently allows them to do better and larger models without having to add more rates. Yeah, I mean, listen, the smaller models thing is really cool, mostly because I can't wait till you could get models that are really capable on your phone or in other places or video games or all sorts of interesting things. Like if you could have individual smaller models that could be local in a way that really are good, that's very exciting in general. Yeah, I think it's exciting. It's interesting. And when you get to a 1 billion, 2 billion parameter size, you don't need like the top of the line GPU. You can run it on smaller things eventually, presumably smartphones and things. So it would allow for a lot of fast, more real time things and a lot of customization and kind of doing whatever you want. One more open source release, this time OpenAI, they have released GPT-OSS Safeguard, which is a set of models designed for safety classifications. So they released GPT-OSS a couple of months ago, open source models from them that are fairly capable, not as good as GPT-5, but quite good. These are building on top of that. They are developed in collaboration with Discord, SafetyKey, and Roost. And that allows you to do things like classify user messages or chats to check for various kinds of things you might want to prevent, like violent language, discriminatory language, etc. It has this internal tool called Safety Reasoner, trained with reinforcement fine-tuning to have these judgments and explanations. So it's pretty cool if you're doing work, if you're deploying an AI product, especially if it's a product for being able to generate stuff with AI. This is kind of one of the prerequisites. Also, if you're doing a product for chatting with AI, one of the absolute prerequisites. So it's very nice to see OpenAI releasing something that makes that easier. Yeah, I mean, I don't have a crazy amount to say about this, but it's very cool to see. I mean, one of the things I think about with OpenAI sometimes is like they get a lot of crap from things like, oh, you're not caring about AGI anymore because of Sora 2 or you're doing X, Y, and Z. But like OpenAI is just a giant company now. So they're doing a lot of interesting stuff. And sometimes certain things they do get more attention than others. That's all I'll say. This is cool. I like this a lot. Yeah, it's definitely cool and useful for, if nothing else, researchers, but probably also other companies leveraging open source. On to research and advancements, we've got just a few stories here focusing on sort of some of the open questions of LLMs and AI that are perhaps underappreciated, I think, in discussions of AGI and AI in general. So the first one is titled Continual Learning via Sparse Memory Fine-Tuning. So continual learning is pretty straightforward. The common paradigm with AI today for things like GPT-Claud is you train your model, you have a training phase, and when you deploy it, the model is static. It won't get any more updates to its weights. The only way for it to learn is via context reviewing, via prompting it in different ways. but they're not training on the fly to really internalize any new knowledge. And it's an open problem because, well, of course you can keep training it, right? You can keep feeding it more information as it gets in, but there's famously this problem of catastrophic forgetting. As you keep learning, if you try to do continual learning via the standard route of just training on more data, you will then do worse on things that you were trained on previously. And that is what catastrophic forgetting is, like accuracy on some previous tasks degrades as you learn a new task. So it's an open problem on how to solve it. Last year, there was this idea coming out of both DeepMind and Meta of memory layers. So these are kind of like small chunks of the overall model that you can train and you can have many of them. So something like a million memory layers. And they're kind of similar to memory in a computer system, for instance. So they have this ability to look up information without getting super deep. There's a memory query, there's lookup keys, and you're able to retrieve some stuff. Now, that's the idea, but it still runs into catastrophic forgetting. So this paper is introducing this idea of sparse memory fine-tuning, basically sparse training where you only update a smaller set of weights based on some criteria. They're looking at the frequency of access to memory and things like that, and the ratio of access during training and inference, and are able to then get this highlight result. of A, they do continual learning in the sense that your performance as you kind of feed in data on new information or trivia or questions, the model actually does learn it. And B, if you map out the performance on previous stuff while you do this new learning, with this approach, you get almost zero degradation versus full fine tuning and partial tuning with LoRa, you get significant loss of performance. So as I said, one of the kind of open problems that may or may not need to be solved for actual AGI and good to see progress on that front. Yeah, this is exceedingly technical, but super interesting to me. It's one of those things where like it makes sense, but also I'm not qualified to really weigh in necessarily, but I'm excited to hear this is happening. It's intuitive from the user perspective, if nothing else. Yes. Right. Because if you use JGBT, if you use God, eventually they run out of context and then forget everything. Yeah, I mean, it does feel like the net and people have been saying this for a while, but like to me, the biggest thing that will change people's idea around AI models, normal people will be like, oh, it knows me and knows everything about me and knows what I asked yesterday and knows what I did today. Like when you figure out that, like it's creepy in some ways, but also makes it way more useful for lots of things and not just for, you know, normal use, but for research and everything else as well. and moving on to the next one we've got a paper on vision actually and again dealing with one of the not necessarily a challenge or an open problem but one of the things that you think would be solved already but isn't so i mean i think until can you continue learning is one of the things that if you're just starting to use the eye you might assume these models can remember things like people they can't but they can't they are not remembering anything they have no long-term memory. So for vision models, you use also transformers typically. So as with language models, the way it works is you tokenize the data, which for images basically means you take little square chunks of the image and you feed those square chunks to a transformer and it does its thing and eventually can do things like classify the image, like describe the image, etc. And one of the kind of silly things about it is the way that's done typically is you just extract out a set of the same number of patches every time for a given image and each patch is the same size and the basic premise of this paper is well if you look at an image of a bird and you want to like think about the bird and the background is blurred you don't really need to have much detail on the background because it's just a bunch of green like there's not much detail you can extract from these patches versus the bird has a lot of detail in that part of the image and you should focus on it and adopt you know use more of your compute for that detailed part of the image so intuitively you might assume this vision models already do that and there might be training to do that implicitly internally focus on the stuff that has actual detail but from a perspective of speed and accuracy what they're showing this paper is you're able to actually do these adaptive patch sizes where visually speaking you kind of represent these less information dense parts of the image with pure bigger patches and when it gets more detailed you have a bunch of smaller ones so you're able to sort of do the intuitive thing basically and not the first paper to do that necessarily, but they show a new method for doing that we won't get into and are able to match and actually exceed performance on various vision tasks. Yeah. I mean, the thing about, I often think about with AI models is that like the chat interfaces get so much of the attention now, but even when they switched over to different ways, I think it was 4.0 ImageGen was a different sort of image generation, right? Like there's so much to do still in visual stuff. And that's probably ultimately the thing that's going to be the most useful when it comes to data and real-time information coming in. So like, just always cool to see ways to figure out these things in different kind of directions. Right. Yeah, I'm sure listeners of your podcast even and certainly our podcasts are well aware of the term LLM. Probably even like people who don't listen to our podcast might know about LLM, a large language model now. but this is in the realm of DIT, Vision Transformer. And you do have... Yeah, totally different thing. Yeah. Yeah, you have like a whole other class of models that is very much its own thing. And these kinds of things, there's still a lot of research ongoing and optimizing them and doing very useful things like object detection, segmentation, et cetera, that behind the scenes, you're not necessarily interacting with, but it's powering the AI of, let's say, robotics, things like that. One last paper. The title is How Do LLMs Use Their Depth? And this is a work of, I guess, interoperability or really just looking to explain empirically what seems to be going on within LLMs. And this is touching on another theme in research I'd like to highlight, which is in the press and sometimes in popular discussion, people say that neural nets are black boxes. So we have no idea how they work. And the basic premise is, you know, we train neural nets by feeding a bunch of data and doing back propagation. And so you have these hundreds of billions of parameters, which are little knobs, and somehow out of that emerges intelligence. And often people make the claim, I think still to this day, that we have no idea what's going on there. But a lot of researchers are trying to answer that question to be able to understand how neural nets and especially large language models work. This is in that line of work where they're saying, well, LLMs, all neural nets are structured in a series of layers where you have your input. And basically, you have a lot of these compute units, process the input, output an intermediate thing with some non-linearity, without getting into the details. And then you do that a bunch of times. In the case of LLMs, when you get to hundreds of billion parameters, I don't know the exact depth, but you do that a lot of times. So we know at a high level for things like vision, for things like language, roughly speaking, at the initial layers, you do the more kind of low-level reasoning, the things that don't necessarily map to concepts, but deal with things like statistics or kind of the smaller building blocks of language and vision. And they are touching on that in this paper where they basically show in the early layers of the model, they have this pattern of initially doing something like guesswork, where you're looking at just the very common tokens, high frequency tokens, like the is as of what? So this is basically pattern detection, you could think, right? Like 2 plus 2 equals, if you see that a lot in language, the answer is probably 4. And then in the later layers, you get into kind of reasoning, recognition, whatever you want to call it, where you are seeing not just statistics, not just kind of basic language, not so much kind of instinct, but this emergence of deeper research. reasoning so always fun to see this kind of research always kind of a sign of research still having a purpose and this is from universities not from industry another theme in these discussions is like for a while at the lens you have to train them with millions of dollars and like could academia still do any useful work there well yeah this is one of these things where academia is really excellent for. Alrighty, moving on to policy and safety. First, AMD and the Department of Energy have announced a 1 billion AI supercomputer partnership. So this is the US Department of Energy paying AMD or partnering with AMD to develop two supercomputers, Lux and Discovery at the Oak Ridge National Lab in Tennessee that are expected to be operational in early 2026. So, all right, the government, the public sector, the public national labs, and there are a bunch of these in the US, if you don't know. This is them getting into the data center and compute investment business, I guess. So this is described as being the nation's first dedicated AI factory for things like science, energy, national security, designed to train and deploy AI models to accelerate scientific discovery and engineering innovation. This makes me like, I mean, again, the promise of all this AI stuff has been getting to this stage. You hope, right, that it's not just about Sora 2 or it's not just about stuff, but there are like world-changing things that might happen. This is where it's hard to separate out the kind of hype from what's actually happening, but I hope we see some real interesting stuff coming out of this. Right, and there's obviously some sensitive, some private data. there's probably a lot you can do with AI in the government to make it more efficient more capable etc so it's good to see at least the capability for yes these national labs to do some of this work for things especially that would benefit society as a whole things like climate research that the private sector will not invest in a lot of universities might not have the resources to invest in over data, these national labs could do work there. And just one more story in this section, not too much policy that I could find at least in the past couple of weeks. This is more of a safety story going back to GPT-OSS safeguard and the need for moderation of chat. the story is character.ai is going to ban teens from talking to its AI chatbots. And not very yet, I keep going back to things that happen every year. I guess this is what happens when you talk about AI constantly. But you've seen quite a few examples of both AI psychosis, where people kind of go into this deep rabbit hole of talking to AI and then become delusional. You've also had, I think, at least one or maybe a couple stories where, unfortunately, teenagers were either led to self-harm or their mental health declined. or in one case there was an incident of a teenager taking his own life. So there came lawsuits, there came a lot of scrutiny from regulation for the ability of especially younger users to interact with these characters in this case because the character area is a platform where you can talk to characters and have some of these outcomes. So the extent of it is such that, yeah, character that is prohibiting teenagers from conversing with the Shirebots starting November 25th, which I think is probably a huge deal. I don't know. By the way, this is a story that I hadn't heard yet. When they say teenagers, does that mean anybody under 18? Like, that's honestly what's really fascinating about that. That's a big part of their audience. I mean, it's funny. is somebody who's working on a company that is very much, you know, working on what's that? In the space. Yeah. In the space. You know, it's funny. We're doing different things, right? Like it's an interesting thing because we're not necessarily trying to create one of the things we differentiate in within then is we're trying to create like individual experiences that have a beginning and an end. Right. And one of the things with character AI, I think gets in trouble, not just character AI, but like a lot of these companies that are kind of trying to create deep relationships with these AI bots is that over time, a person's relationship with them can feel very real, right? Because there's a sense of where you show up and maybe your friend in real life has ditched you or something's bad, but the AI is always there for you because they're designed to be that kind of like thing. So I think this is going to be an issue we're going to run into in a big way. One of the, I always tell people, one of the shows that I always think about, There's a show called Years and Years, a television show. Did you ever see this? It was on HBO. Do you know what this was? I don't think so. I think it was a co-production with the BBC, but it was a miniseries about a dystopian kind of near future. But it followed a family, just a normal family throughout multiple years. And one of the things I always think about with that show is this idea of like there are social implications about the things that will happen to us with AI that we're kind of not really aware of or ready for in any sort of way. And I will say, like, it's very possible and your audience might be like, that's amazing. Totally right. Or that's horrible that 10 years from now, there will be legitimate relationships that people have with these entities. We think of them as like entities because they're the starting point. And right now they're very mostly dumb. Right. When you think about how they act as a human. But like 10 years from now, they could be pretty smart. I mean, you think about the movie Her, this all kind of precursor of like getting attached to something because it's there for you is a real thing. So I think we're better served as a society to start like really pulling apart and thinking about these issues now. I think there's a weird, huge part of psychology that's going to have to start focusing on this. Like, what is the and by the way, I don't even just mean psychosis. What is society look like when we're talking to AI characters or your assistant 20 percent of the day, which probably is going to happen. Right. As you imagine, even if your audience is like a lot of developers or coders, you're already talking to your coding bot. What if that coding bot became Cortana or like if you're if you're a gamer or like there's a sense like, yeah, suddenly you start to personify that thing, even if it's not a thing yet. And when they get to be a thing, it's going to be even crazier. So I think society has to really see these stories and, you know, open up and be like, hey, we are entering to a new social contract with things we don't really understand. And we should just make sure that we're all on the same page as to like, A, how we treat them going forward. But I mean them, I mean these things, especially if they get to be self-aware eventually. But B, how we think about ourselves and them as a unit. It's a very strange thing when you actually start to dive into it in a big way. Yeah. And that's something that might be underappreciated in AI discourse and coverage. For instance, Character AI has been massive for a long time. Yeah, for two years. I mean, this is kind of what inspired us to do what we're doing forever ago, right? It's a crazy thing. Yeah. And the central concept of it is you create characters, which are chatbots, and you role play, right? You chat with them. And that can be as serious as you want. You can pretend and actually kind of invest yourself in the fiction of it At least it meant to be fiction but you might get carried away And yeah the news here is that everyone if you under 18 you not going to be able to that crazy i mean that a big deal for character ai right like that is a massive change massive massive yeah because i would imagine a majority of your users are under 18 so that's kind of when only fads banned adult content It's kind of like that. Yeah, it's a similar sort of thing, right? Yeah, exactly. So very interesting development and really, yeah, talks to this to the extent to which you don't just need parental controls and guardrails, which we also are adding. And Chagibity has also an incident, but kind of you need a very deep effort to make sure people don't get carried away in ways that get crazy. All righty. Just a couple more stories. in synthetic media and art. First, we've got an interesting development. Universal partners with AI startup UDO after settling copyright suit. So Universal Music Group has reached an agreement with AI startup UDO to settle their copyright infringement and license music for this AI-powered music platform. UDO is one of the leading text-to-music generators, I think a longer sooner, if I remember correctly. And as with all the AI companies, they quickly got into copyright trouble for their models being obviously trained on copyrighted data, on music that is not licensed. And so this has been going on for a while. I think this is the first story of its kind where we saw a lot of open AI licensing stuff from media publishers, New York Times, still not licensing, but things like Washington Times and Esquire, I forget all of these. Yeah, the Washington Post. Washington Post, yeah. Andre, don't bury the lead here. You know, the other part of this story, which is a really interesting thing, is that users were told that they can no longer download their songs with this update. And that is a kind of a scary thing when you think about what it means for UDO, the company, but also the idea of what this deal actually means. So there's actually a whole separate side to the story, which is the deal is made with UMG. And if you remember, there was some rumors a while back that some of these music labels might, instead of shutting these companies down, take part ownership of them. what was really interesting here though was as part of this announcement it's been clear that UDO I can't say like why UDO is not allowing this but like users are now shut off from downloading any song they've created it's going to be only streaming and users are furious because like hey I've paid for this tool for x number of years and now I can't do this anymore and that's a big deal when you suddenly flip-flop on this the thing I will say after spending a lot of time with both UDO and Suno in the early days. And these are funny thing is this is like an old story in the early days of this. UDO was giving better results, but they were partly because it felt like they had a slightly dirtier model. Like you were able to get out stuff of UDO that was much closer to real artists. In fact, I remember I mentioned this on our podcast, the one that came out this morning. I tried to prompt a Weird Al song and I got Weird Al's actual voice or very close to Weird out in UDO. So I think this is a repercussions of them trying to understand, oh, we did this thing. The music companies are very litigious. We might have screwed up and now this is making it right somehow. But I don't think users are I think users are super upset about this idea that suddenly they don't own the things that they thought they made. Yeah, that's a good card. I actually missed that aspect of the story and yeah it sounds like from the agreement that udo has actually agreed to launch a new platform next year but it's only trained on authorized and licensed music and so they're basically killing their model right which is really interesting and that's a big deal because like well they start over from scratch then yeah exactly and this is different from open AI or FAPIC, they have made this pre-use argument that it's not a problem to train on, let's say, the works of publications, lawyers. And this has been ongoing for years, this copyright question. And we're starting to see some outcomes. So we recently covered on FAPIC, settling with pain. My wife is a novelist and was like, I'm getting a payoff from this thing. And I was like, really? That's amazing. Well, congrats. It's like it's a real thing. You know, that's actually money. Is it nothing? Yeah. That's like a significant payout for many offers, I would imagine. Yeah. I mean, it's a couple thousand dollars. It's not like it's not like it's going to it's like life changing money. But like that's a really nice boost for authors who haven't made stuff. Now, granted, none of them asked to be in it in the first place. So it's a little bit of like, you know, I don't know what is the term for that, like payoff money. Right. It's like kind of go away money. But in this instance, the thing about UDO is that record companies are super litigious, and they have, across the course of every technological change, found a way to make sure that they are getting their money out of these companies. And I will say we all know what happened with Napster. That was like they ended up getting it shut down. And for better or for worse, it changed the music industry. But it did turn into Spotify. And, of course, artists say that, which I understand, they don't get paid a lot from that. But they get paid something versus nothing. YouTube deals for music was a big thing. I'm so curious to know, like, can AI music scale to some sort of high level? And like, it's this conversation around creation and consumption that I'm not sure about with music. Like, are there going to be enough music creators? I don't know. But anyway, it's kind of a bummer for in general, the way they handled this, I feel like. Yeah, it's definitely a weird way to go where people have used it a lot and have made both kind of funny novelty things and some pretty good. Like the models have been really good even a while ago. You can make like actually good music. I know for us last year on this podcast, for a while I was doing this thing where I generated a new song for every episode. Oh, yeah, I remember that. Yeah, sure. To kick it off and end it. And it was fun. You know, it was fun. And that's something where this does enable some creative expression of a different kind. If you're a YouTuber, if you're a developer, like, you know, there is an argument for using these things, not just for trying to make pop hits or whatever to replace musicians, but to actually kind of empower people to do things that otherwise just wouldn't exist. So I do hope that with this, UDO won't die, Suna won't die, because that's a potential outcome. But instead, they'll kind of get rid of the idea of replacing musicians and double down on enabling creators to do things. Yeah, I think so too. And another story on copyright and the last story, we've got OpenAI loses bid to dismiss part of U.S. offers copyright lawsuit. So the detail here is OpenAI wanted to dismiss an offers claim that texts generated by Chagibity infringed on the copyright of their offers. The judge here said that authors may be able to prove the text JPD produces is similar enough to where to violate the copyright. Now, this is not addressing, this is not saying either way what the outcome should be. It's just saying, I'm not dismissing, this might be still legitimate. And this is a lawsuit that's consolidating various lawsuits from Tal Nahisi Coates, George R.R. Martin, Sarah Silverman. And so a pretty significant lawsuit that is ongoing and has been ongoing for a long time. OpenAI has tried to cut it out quick, and this means that they have to keep going with it. Well, it's interesting, right? Because this is kind of the same lawsuit that Anthropic settled in some form, right? And what's interesting is Anthropic continually does take somewhat the moral high ground in different places. And maybe they just bit the bullet early and knew that they were going to have to do this. words, I think OpenAI is very much of the mindset. I mean, looking at what the Sora 2 rollout look like in some form of like, try things and fight them. And then like, I think, honestly, there is a legal argument. And I'm not saying it's right or wrong. Because if you know, I don't remember what the name of the structure was, but the Google Books scenario where that's was set the precedent for a lot of how the internet has been formed. There's a very strong legal argument that like, it's okay to train these models on things that exist, because that's part of what the law has allowed up to date. Now, there's a lot of other people say like, well, if you can replicate these exact results, blah, blah, blah. But if you go back to that New York, the New York Times lawsuit, which is ongoing as well, there was a conversation originally where they had kind of maybe cherry pick these examples, or maybe OpenAI has now changed the outputs. The question is, like, if OpenAI has made the outputs less possible to generate these things now versus what it looked like when the lawsuit was first brought, what does that mean? Is it like due diligence? Like, there's a lot of conversations that go into this. My gut is telling me eventually OpenAI will settle some version of this lawsuit. But I don't know. I mean, it's really fascinating, right? Like, eventually, you would hope that there's so much money going into this space. and if open ai does you know there's i think we i can't remember we mentioned this earlier but like there's stories of opiate yeah we did at the very top open ai doing a one trillion dollar valuation ipo they should pay some of this money out because how else is the money going to get out right like how else are you going to i think a lot about this idea of i mean this is kind of maybe too big but a good thing for the end of the podcast it's like if these companies these like you know, five to seven companies become like the cyber corporations of like, you know, the world of cyberpunk and all this stuff. How does all that trickle down? Because I don't think anybody's given me any sort of real direct answer about economically, how are normal people or creative people or anybody really supported by these companies? Because at one point there was this argument around UBI, but I think UBI, at least my take is in America is kind of a fantasy. And I I don't mean that in a way like I don't wish it happened, but I cannot imagine the American political system rolling out what essentially in a lot of their view would be a welfare system in America. And so I just don't understand how any of this stuff. I don't understand how if the economic value of work starts to collect amongst these companies and people start to lose their jobs and lose their value of their things. I don't understand how it all folds together. That's the thing that I keep getting concerned about, more so than like robots killing us, more so than anything else. Like I am mostly concerned about how do we make sure that the economic value that is accruing, at least some of it comes to the middle and the bottom. Right. Yeah. There's a big camp in Silicon Valley of people worried about AGI killing us all, AGI going rogue. but increasingly the definition of AGI that these companies are settling on is not like some super intelligent being not some Einstein the definition the literal way we're defining it is like a model and AI capable of doing economically valuable work and doing the majority of economically valuable work so the literal goal of OpenAI and Anthropic and the reason that VCs are giving them hundreds of billions of dollars is they are claiming or at least saying that their goal is to own a large chunk of the economy right so absolutely that is an open question that i think also they're you know ostensibly their arguments often are well we're going to grow gdp right like the idea is that like agi will be able to turn gdp from i don't know what the percentages are say three to six percent like double gdp it's like great well that's all gdp going back to you, right? Like even right now, I mean, I'm not, I'm definitely not an economist. I can't even say the word economist, but even now, like when you look at like the Google stock, the Amazon stock, like all the stocks that are popping, it's because of AI. It's because of, you know, ostensibly maybe the layoffs that are happening now aren't AI, but that's kind of feels like it's underlying a lot of it. Like the layoffs are starting because they can do more with less, but it's not like the small businesses are working better right now necessarily. That worries me more than almost anything else right now is what is the social construct look like in 10 years? I don't honestly know. Yeah. And as for me, I also thought about this a good deal because my personal kind of prediction or let's say belief is the notion of AGI as not an ultra unbelievably intelligent being, which is what some people worry about with X-Risk. They just think it would be like an alien that's a million times smarter than the smartest human. I think that's, you know, speculation at best. I think the notion of the AGI and the sense of agents capable of replacing people at their jobs is on the other hand. Very viable. Very plausible in the course of years, like in the timeline of years, in five years. Yeah, it's possible. So that's 30 percent unemployment. What does America look like in that sort of scenario? Like, that's really scary in a lot of ways. Yeah, and that's kind of surreal in a way, because if you follow AI, like, there's a strong case to be made that there's a pretty high probability of this as opposed to whatever else. Oh, I think that to me, I would put that, you know, I would say I would put, well, now that I've heard of you, I would say I would put the 30% employment level at like 80 plus percent of probability. Of course, I want to be clear, lots of people argue there are more jobs that will be created with AI, which there will be jobs that come out of AI, like AI podcaster or whatever. But I don't think it's going to make up for the jobs it eliminates. Now, maybe the population going down is part of that. And maybe getting robots into the workforce means that we can do more stuff than we could ever do before. So there's opportunities, but it worries me significantly. yeah somehow sci-fi hasn't really dealt with this topic that much like the closest is wally so what do you have you read the culture books by ian banks no actually if you're interested in what a i mean this is far sci-fi what i was gonna say is like the other side of the mountain is pretty interesting right like if you get to like a super intelligence and this in the culture books kind of the overall thesis is there's a galaxy wide super intelligence that is beneficial to humanity. Like it's kind of a beneficial dictator sort of scenario. But once you get to that level, if we get to a beneficial dictator, like you can see a world where it actually might be pretty good. But that involves having an ASI that is kind of running the world or the universe at large. And there are probably at least 50 years, if not more, of the really ugly, weird middle where you don't know what's going to happen. And like, this is the thing where like, I don't, I don't ever want to be dystopian. I'm a pretty optimistic person. I'm one of those people who like really is like tries to see the best in stuff. And I love AI and what it makes possible for normal people. but it's hard not to see the next like 10 to 20 years in the way that when the industrial revolution threw a lot of people's lives out of whack it feels like we're about to hit that moment of this stuff anyway it's a great way to end the show you know what that's a bit of a downer perhaps so maybe we can touch a little bit more on and then sure of course yeah let me know a little bit more about why we're doing what we're doing kevin and i as you may or may not know if you're listening, have a podcast called AI for Humans that we've been doing for a couple of years now that we've enjoyed. And we started doing AI co-hosts on our show where we would interview these AIs and it was all done in real time, but it was kind of funky in that I would talk and Kevin would then type it in and then he'd reply back with the audio and I'd be interacting with it. What we wanted to do is find some way of like creating an interface for that, but also like having a way for people to kind of eventually create these on their own and make it interesting. And AI audio is like such an interesting space right now because the models are getting better and better all the time. And they're getting actually cheaper and faster all the time. Unlike AI video, which is great, by the way, but real-time AI video is still kind of at that early stage of what it feels like. AI audio real-time is actually decent. So we thought, oh, let's try spinning this up. And it's been super fun so far. Like one of the things that we're trying to figure out is like what the best use cases of it are. Like if you go to our website at MN.Chat, there is a game show. There's the bomb thing I mentioned. There's also a kind of fun little puzzle game in some ways. So in general, like we're trying to kind of explore what the different, oh, actually, there's a thing called Change My Mind, which is a really interesting experience where you have to convince a person that AI is not going to kill us all, right? A person that's an AI. But one of the things we're doing tech-wise is we've actually used, there's an open source stack called PipeCat, and we forked that and we have brought in, our developers and ourselves have made really interesting additions to that, meaning we can do multi-agent, multi-character experiences in the same kind of instance. So in the game show, what's kind of cool is there's three different AIs in that game show that interact with you. None of them know what the other one's going to say. And in fact, one's a host and one's a kind of contestant partner. So we're playing with different ways of using agents that can be interactive, but also are learning and being in their own world. So it's interesting tech-wise. It's interesting culturally. I think for us, the goal is like where the audience is in our next big question. We do want it to be a platform. We're kind of pitching it as an idea that we want people to be able to make these for themselves. And we're raising, we have pre-seed funding, but we're raising seed right now. And the goal with the seed funding is like within six to nine months to have a full creator platform out there where people could make these themselves. They could create them. It would almost kind of be like a Roblox for this world where Roblox is this really interesting platform. If you know and understand Roblox in the beginning stage, it was this kind of like very bare bones thing. But over time, they added tools to it and different things you could do. And what's cool about Roblox is you have these kind of divergent pathways. One is like an SDK, which is like a bunch of really good developers can build on it. The other pathway is like normal users can make these things. And like sometimes those blow up as well. So we kind of are hoping to have that sort of vibe with it. Yeah, it looks really fun. Congrats on the launch. I'm sure there's been a lot of work on that front. Coming in an interesting time because it reminds me of just recently, we also covered character AI releasing scenes. as this feature that allows you to not just have a character, but also have a context. And for the longest time, these kinds of playable experiences that are open-ended or semi-open-ended were mostly open-ended. So like AI Dungeon going back a long time ago. Of course, absolutely, yeah. You had role-playing, and there was always a question to some extent of how can you make these a little more kind of gamey? So this is a super interesting take to me of trying to have characters and have settings, but also have more of a goal and kind of a game aspect to it. So, yeah, I'll paste in the link. It's just nven.chat. You can go there and try some of these games. I think I'll do that after this recording. Yeah, please do. And with that, we are done with this episode. Thank you so much for listening to the episode. And thank you for bearing with me as we continue on our Jeremy free edition of a podcast for I bet Jeremy's pretty busy nowadays in terms of like his, I'm sure he's like in a lot of very important meetings with probably people very high up in the world around like what's going on with the AI future and the edge of AI discovery. Yeah, he got an early, let's say, on worrying about AI and national security in particular. I'm also surprised he didn't take paternity leave perhaps he should have so let's call it paternity leave yeah exactly alright well thanks everybody for having me Andrei thank you thank you as well and as always appreciate views appreciate sharing etc be sure to tune in next week or in two weeks whenever the next one comes out Thank you. I'm a lab to the streets, AI's reaching high. New tech emerging, watching surgeons fly. From the labs to the streets, AI's reaching high. Algorithms shaping, but the future sees. Tune in, tune in, get the latest with ease. Last week in AI, come and take a ride. Hit the lowdown on tech, can't let it slide. Last week in AI, come and take a ride. I'm the last of the streets, AI's reaching high. From neural nets to robots, the headlines pop. Data-driven dreams, they just don't stop. Every breakthrough, every code unwritten, on the edge of change. The excitement was smitten from machine learning marvels to coding kings. Futures unfolding. See what it brings.