Are We Seeing The End Of OSS Models?

0:07Aaron: Good morning, good evening, and welcome back to the Enterprise AI Show. This is your host, Aaron. And as I record this, it is the beginning of August, and I'm actually in Utah getting some much-needed R &R before we head into the fall. So today, we have the first in a two-part series talking about OSS models. both a potentially positive outlook and a negative outlook as well. Brandon is back for this episode and we explore the potential decline of open source and open weight models in AI. We discuss market trends, financial challenges, and the future of OSS in the AI landscape. And again, this is the first of a two-part series and all of that is coming up right after this break.

0:59Today's show is sponsored by Nasuni. There's a growing gap in AI right now between what's possible in theory and what successfully works at scale inside an enterprise. The difference comes down to unstructured file data. Many AI initiatives struggle because the file data they depend on is scattered, unstructured, and disconnected from where and how work actually happens. Nasuni changes that. It brings your unstructured file data into a single, secure foundation so AI, both generative and agentic, can access it with the context, governance, and performance it needs in production. Bring AI to where your unstructured data lives.

1:33See what it takes to activate your data for AI and request a demo at nasuni.com slash AI.

1:41Aaron: And we're back. And we are continuing our Models and Markets series. and again, continuing his creep to take over the podcast slowly over time. We have Brandon back as well. I'm not even going to introduce you this time because if they don't know you by now. That's right. They should know. If not, you need to retrain your model. I forgot who Brandon is. Go on. All right. So we're going to continue our series on this. And our topic for today is, are we seeing the end of OSS models? And oh, by the way, I'm going to tell everyone the next show is going to be ROSS models taking over the world. So this is one half of a show.

2:22Aaron: But we're going to start with the end. We're going to start with the doom and gloom side. We're going to be the half empty side of this argument to start all of this. And so first of all, why now? So I'm seeing a trend in the industry, and Brandon, I would love your take on this, where we're seeing fewer and fewer Apache models, especially on the high end. I feel like there was a decent amount of Apache 2L models, and that number is shrinking. And then there's a lot of present things we can talk about. But I feel like it's probably coming down to as the market is maturing, I mean, hey, making the models, especially increasingly bigger models, increasingly bigger costs, these things aren't cheap.

3:10Aaron: A lot of these companies are burning through cash, but they got to show some recoup at some point on a lot of this. And I'm going to go back to the past for some of this, like the OSS rug pull joke. I'm going to bring it back out, right? This comes to mind. I was like, I was just trying to make up a list of rug pulls that have happened in the industry, you know, not even AI related. I came up with HashiCorp, MongoDB, Redis, Elastic. And that was just off the top of my head. Yeah. And I mean, there's numerous others. And so like, what's your thoughts, first of all, on like OSS and the rug pull philosophy and the origins of it?

3:58Aaron: Like in my mind, I feel like a lot of them happened because especially in the database space, it seems like a ton of them are in the database space. we had all these oss databases and then they had very good licenses and then quite frankly the hyperscalers cough aws cough took them made services out of them and made more money than the original companies did and then you you know the the original company didn't recoup the costs absolutely of that and you're seeing that same thing with with the models right like hey you want to put that thing out there you want to get lots of adoption and then oh by the way once you get lots of adoption, you're like, oh shit, how do I make money on this?

4:39Yeah. You know, I kind of, are we seeing it all over again? Well, I see it slightly different. I mean, I think, you know, all those open source projects basically were able to build out, build big communities at a relatively low cost in comparison to what we're doing today. Right. And as those customer bases got bigger, right, they were able to sell some stuff. And I think that attracted competition. Now, currently it kind of feels like the opposite. Like the models that we're all talking about are the proprietary frontier models, like Anthropic, OpenAI, and a few others actually outbuilding that.

5:08And the cost, though, is just so high to build a frontier model, right? I mean, you have to have so much money and access to such big compute that it isn't like it almost doesn't follow that classic open source. Like, you know, three guys in a room or three people in a room can go out and just build a new database. It's like, you may be a very smart researcher on your own, right? And we even see some of the best and brightest, what Ilya Skutsuver went out on his own. But even he had to raise several billion dollars to actually do that. And now that some of his ideas are coming to fruition, he's got to go raise even more money.

5:45So this idea, and he's like a superstar in this world, right? He's like one of one. So if you think about it, I kind of just think of it, it's just financial. It's one of these things where it's like, as it currently stands today, there's just a handful of companies in the world that have the money and the infrastructure to actually train a model. And really, if we get down to like the scientists that are coming up with the new ways to do it, it's like, I don't know, you probably know better than that. It's like, are we talking like 100, 200? I mean, it's a very small group of people that are truly like, you know, into the deep math and figuring out this is how we want to, you know, we're all prompt engineers, but there are only a few of us that are actually like, what's the way that we should, you know, do the math and the matrix math and the tensors that that model is better, right?

6:34There's a very, very small skill set. And most of those people, rightly so, are paid extraordinarily well by companies. And I don't think any of them think to themselves, I would like to go home on the weekend and just create my own model and release it. And I have to somehow magically come up with billions of dollars training. I think, as you You're around these people a little bit more than I am. I think they enjoy, I have a really good office. I've got a nice salary. People leave me alone. I write my papers. And my papers become the basis of these incredible models that do things. So I think that's just like, I don't know.

7:08It's just kind of like business logic. Now, the only thing that I would say that to counter myself a little bit would just be like, if there were a place, and maybe the government is the only thing that jumps to mind, or other governments who are like, yeah, we're going to actually give all the money, and we're going to make it free. right kind of like nist in the united states and other places like the super collider kind of thing right we're gonna do all this research like okay maybe that's the only other group that has the money to even do something like this but other than that i don't know i just don't know how you

7:35Aaron: how you do it going forward what are your thoughts yeah i well so here's a couple you know data points if you will here's something else in the market i don't even know that the market i mean probably saw coming but so inkling you know just released a model here a while ago and for those that don't know inkling was miramirati and and some of the urban ai folks that left they they came out with their model well it's a poorly kept secret that it's actually distilled from the chinese models and so you have if you believe the rumors the chinese models are distilled from the frontier models and then now you have vc companies paying to distill models from the chinese models like we talk about circular logic and circular funding and all these other things now we're even down to circular you know model development but and so like i just find some of those like that part of it pretty comical at times but also too like will the end come because of a couple different things And what I mean by that is, will governments jump in?

8:47Aaron: You've already seen the U.S. government, and not necessarily for OSS models, but U.S. government stepping in on Mythos and the latest version of GPT from OpenAI. And everyone was like, oh my gosh, go use OSS models because the government can't restrict you. But at the same time, there's rumors China is going to start restricting the high end of their models because their models are getting so good. and you you're the companies themselves you know k3 is an example it's it's it's limited if it's above a hundred million users or 20 million in revenue and you're starting to see others with restrictive licenses modified mit in some way and and some of them just aren't releasing the weights or they're like oh we're releasing the model today but the weights will come out in a week or two weeks and it'll be on their api for a while so they can catch some revenue before you can download it.

9:39Aaron: And so I feel like what's going to happen here is there's going to be a fine balance of this, of there is the high end, I think is going to be restricted more and more because there's just too much money at stake now. But I feel like at scale, the only way this is going to work is with oss models longer term and so like i don't know is this like is the world going the way the databases kind of did of like hey it's modified licenses like a modified mit is like maybe the future and and is the revenue models become more of like hey the the model is free the same way like maybe the you know the oss software was free but then you charge for services or you charge for, you know, other things wrapped around it, you know, it's, you know, charging for expertise.

10:39Aaron: And is this where like the Palantir's of the world make their money going forward? I don't know that I have an answer here, but I find it really interesting. I feel like the end of the, Hey, you can get a frontier model completely Apache 2.0. Those days are numbered. Well, what about this though? I mean, I just wondering if it's more like point in time it's like will you be able to get like the best frontier model at Apache Tupo probably not but as these models age right and you know age could be six months could be a year it does feel like at some point I don't know I'd have to go look at them like you know obviously Facebook or Meta right they open source the model and you know as these these models I get gold or I don't know I don't know what anthropics policy I probably should know off time and I don't know if they what they do when they retire a model and they deprecate it if they make it public or anything, probably not.

11:29But there is something along the lines of like, as these things, it's just like, yeah, well, the model that two years ago is available, right? And that maybe that can be used for academics or for learning, but maybe we'll talk about it on a future show. It's like, there's other places where maybe running a small local model actually is more than enough. And that kind of gets into what you're saying. It's like, there are going to be companies that do exactly that. They take a open source model and they help you embed that in an application where it makes sense. And that's a whole business unto itself.

12:00And it's always a little better. Just like we think about chips, right? We all, quote unquote, want the best chips. But if we get in our car, our car has all these chips that cost a few cents. They're tiny little chips. And they just do one or two things to start the car, turn on the radio, or whatever. And maybe there's just a bunch of that kind of stuff. And I don't know. Do we all want an LLM running in our car? Maybe, maybe not. I don't know. Maybe that's the whole self-driving thing. Maybe that's a different, that's a whole other episode.

12:26Aaron: The non-determinizing nature scares me. Yeah, maybe that's a whole other episode I'm talking about. But I think maybe that's what we just see. It's just more like a lifestyle. And the fact that these things just like the half-life of a model, I mean, I think it's like six months, right? I feel like as soon as it comes out, you basically get six months to be the big dog. You're like the number one model for six months. And then you're quickly banished. And then two years go by and you're just like, I think you're just a toy model. People just don't look at you as like anything. No one even cares anymore at that point.

12:54Aaron: Yeah, agreed. So I'm going to come up with, you know, I actually just thought of this when you were talking about, oh, what do they do with the older models? And like, do they release them later on and all that? And as far as I know, most of them haven't, right? Like it's either kind of like it's open at the start or it's not. I don't know many that have like changed their minds lately. But I mean, for all the gamers out there, id Software. And I don't know if you remember Doom and Quake and all of this, right? Right. And, and id Software had, had one of the, the coolest things ever was they had an idea of when they released a game, the game two generations back, they fully open sourced it.

13:34Yeah.

13:35Aaron: And that's how like there's running jokes of like, oh, does it run on Doom or does Doom run on it? Like Doom runs on every platform on the planet. And you look at, you know, hey, I'm going to compare Doom to actual useful software. But open sourcing Doom, however many years ago, it literally runs on everything because it got open sourced. And so, hey, maybe at some point these models are, to your point, once it's a year old, two years old, hey, fully open source the model, fully open source the weights. And who knows what would happen. Yeah. Well, and also too, I guess we kind of talked about who has the money.

14:10But one thing we know, computing, it's always faster and better. So we're not really great at thinking of in 10 years, but like in 10 years from now, you know, the GPUs and the technology available, like it's very conceivable that like people could make their own model. Right. I know that seems crazy. Right. But it's like, that's just what we see. Right. It's just like it happens. So, so at some point, like there probably will be when it gets, if you will, the computing cost goes down in such a way that people could replicate these smaller models that are, do have some utility. Then maybe, you know, maybe we do see, maybe we see an uptick in open source.

14:44And when it's more, if you will, it doesn't require the, you know, It's cost prohibitive, I guess. So maybe that's what will happen. And at the very beginning, it seems crazy, but time goes by. And maybe as I sit here on my Mac, it's really just some version of Unix. It's been around for like 50 years at this point. And it's just like, yeah, at some point, I'm sure when they thought about it and they released the Unix, they were like, oh, this is so advanced. It'll never be commoditized. And now it's like we have all that source code for free forever in a million different ways.

15:14Aaron: Yeah, yeah. And I don't know here, but here's the other flip side to all of this. I was just thinking, thinking through all of this. I mean, we all know, and we've kind of hinted at it on, on this series of shows here. What if they, what if there was a reason not to fully open source the models and the weights? Because, oh, by the way, not all of the data and everything under the covers may have been gotten in the best ways. Yeah. And maybe there's legal reasons why some companies never could. Yeah, no, I mean, very possible. But it feels like at some point, like it'll get cheap enough that it'll just be replicated.

15:58Like it's just like nothing stays. I think you used the phrase, you know, gene in the bottle. Like none of this stuff ever stays. I mean, say whatever you want. People are like, we're never going to like release XYZ or like no one is ever going to have your social security number. And it's like my social security number is all over the internet. All this stuff is everywhere. And it's like, sure. I mean, maybe there's an intent. We shouldn't release it. Maybe in a perfect world, we wouldn't do it. But it's like, eventually, all this stuff eventually gets out. Just, I don't know. It's the nature of humanity.

16:26Aaron: No, I agree. I agree. I feel like that's a good place to stop. I know, by the way, as I mentioned in the beginning, this is first of two-part series here. This is the doom and gloom, half negative, the Ed Zitron side of OSS models. But the next episode, we're actually going to do the glass half full. And so certainly look forward to that and stay tuned for that as well. So for everyone out there, thank you very much for listening. We certainly appreciate it. And as always, feel free to reach out to us. And if it is possible, leave a review wherever you get your podcasts as well. And that is it for this week.

17:00Aaron: And we will talk to everyone next week. Thanks for listening. Check us out at theenterpriseaishow.com for past shows, newsletters, and all things Enterprise AI. you