NVIDIA acquires Hugging Face, OpenClaw 2.0 goes multiplayer, and the Linux kernel fights back against AI scrapers

0:02Ben:I can't believe that we were on this call this whole time and we didn't even realize that we were wearing the same dude.

0:07Andrew:Yeah, we clicked the record button and it's like, oh, hey, look, we're both wearing tux. What do you know?

0:11Ben:Well, no, you called yours scary early. Should I hide it? I'm like, what do you mean it's scary? It's tux. And they're like, wait, I'm wearing tux. So that's what a, okay.

0:20Andrew:Well, didn't they just have their like 35th anniversary or something like that? Wasn't it a big one just recently?

0:25Ben:I think so. Although that's not why I can claim a hold of this shirt. I got this at scale here in Los Angeles, I think, maybe like two or three years ago. So I love this shirt. Where'd you get yours from?

0:37Andrew:Yeah, it is 30, 35 years old. Actually, this is my lucky sweatshirt. I, you know, I've briefly worked at the Linux Foundation and earned this sweatshirt as a part of my time there. So I earned mine too.

0:48Ben:I registered for scale and I went and I collected my swag. I love that. Well, okay, great. Well, we're both wrapping tux today. So he can join us on, on the news.

0:59Andrew:Yeah, well, Tux and all of our listeners joining us today, welcome to the Friday Deploy brought to you by Linear Beat. I'm your host, Ben Lloyd Pearson. And I'm your host, Andrew Zickler. And this week we are covering why NVIDIA wants you to build your own models, Open Clause 2.0 release, AI scrapers eating web resources everywhere, including the Linux kernel community, rejecting change sometimes. times and don't let your ai amplify bad culture man i really want to talk about that last one but maybe we'll save it for the end because it feels like it'll be the best so angie what do you just say if we just start with the start at the top with nvidia and boy hot off the acquisition of hugging face we've got this really great article breaking down how NVIDIA really is investing heavily in this strategy of funding open source or, you know, quote unquote, nearly open source models with open data training code, really aiming at making model building itself accessible to many companies rather than being concentrated in a few labs like OpenAI and Anthropic.

2:12Andrew:And really, it's simple. If you have more people out there building models, that means there's more demand for NVIDIA's chips. You know, it's a very interesting strategy. It could generate enough profit to be considered self-sustaining versus the more circular investments they're doing with some of the big providers. Or it could also be possible that, you know, open models sort of fork into this more like long tail, hyper specialized niche applications that really are sort of distilled off of the frontier models. All of our listeners out there with the engineering leaders out there listening to us should be thinking about this is that there is a lot of opportunity starting to emerge in these open models that you can run locally or on your own hosted infrastructure.

3:02Andrew:and um you know i think it really does like complement you know nvidia on one hand as i mentioned has all these massive circular financing deals with these major companies but on the flip side they're also investing in this more decentralized approach too so it's like they're playing both sides of the equation which to me feels like a pretty strong strategy so angiel

3:23Ben:what do you think about all this information yes so the acquisition of hugging face by nvidia makes a lot of sense to me. And especially for those that have been tuning into Dev Interrupted, you know, we've also spoken many times with AMD and their strategy also revolves around open source ecosystem and creating an environment for folks to be able to experiment, but also get to a point where there's a lot of models that are owned by different companies and at different stages of usage. Like maybe on the fine tuning end, maybe they're more base models, more foundation models. But the more models that exist and exist in all sorts of different places, the more strategic it is for any chip maker because it naturally increases the demand for inference.

4:06Ben:The real standout thing here for me is this is NVIDIA doubling down on a partnership that they've already made like abundantly obvious in terms of how they have partnered with NVIDIA in the past. A lot of the projects on Hugging Face are from NVIDIA origin or otherwise related to the NVIDIA ecosystem. And so there's already like a lot of great synergy there. I think what's fascinating is really the opportunity that is now starting to open up as more of these really very large and capable, both long running and action and computer use models become open source and open weight. There's an opportunity for folks to be fine tuning more and more things on top of it.

4:50Ben:It makes me think of thinking machines. You know, Mira Morati is a spinoff kind of of like what she's betting on after OpenAI is on the inference where you own the platform and you provide the training and fine tuning services on this like a very great foundation. Right. And people then own and have a part of that model. So there's a lot of like economic things I think still to explore. And this is going to be a big step there. We're talking about a huge investment. So it's really going to be interesting to see how that alters the hugging face ecosystem just in general.

5:24Andrew:Yeah, you know, the thing that I'm really intrigued about from this is the notion that open source models really may be most successful within hyper niche applications. You know, because I think there's a lot that goes into the frontier model development that is really difficult to replicate unless you have like extremely knowledgeable leading experts at machine learning and LLMs and all of this technology. But on the flip side, distilling value off of those frontier models is actually very easy today. And I think the hardware to run your own infrastructure is getting closer and closer to reality for this.

6:02Andrew:And I know at Linear B, we are thinking a lot right now about cost efficiency when it comes to AI models. How do you select the most efficient model for the specific task that you have in front of you? And there's a lot of context, actually, that goes into that that is really important to suss out. And I could go off on a tangent about this. But the short end of it is everyone needs to be thinking about, as token costs are rising, you need to be exploring your options of how to efficiently manage where your tasks are getting pushed and make sure you have all the context that you need to make the right decisions.

6:43Andrew:So yeah, I couldn't agree more.

6:46Ben:It's like you have to have really good insight on how much all of this actually costs you. A really great experiment is, for example, if you use like a subscription base usage of your AI tool of choice, maybe that's Codex or Anthropic. Like what I've done is I route those requests through a proxy. And so then I capture all of the inference that I do through my subscription. You can then compare how many tokens you consume through that to like what's the base API rate. Like if I didn't have this subscription, how much would that inference have cost me if I would have just asked it over like the Anthropic API?

7:18Ben:And you'll be really shocked at how much gets subsidized inside of the usage we have every day. And engineers, I think, really take that for granted, that subsidization of the cost. And there will be a point where all of that will start to rise and we'll be like, you don't want to be like a frog in like a pot of boiling water. You would have already thought ahead of like, how do I get the long-term value out of this? How do I own the source of my inference? It even reminds me of cloud and how cloud hosting, if you have like a variable egress for bandwidth, so you can't, you know, understand a cost month to month, how much it is for like on a platform that has like a dynamic egress, then you can't like reliably budget your infrastructure cost.

8:04Ben:And a lot of investors, like they won't even tolerate that. The idea that you can't evaluate that. And the same thing is going to happen for token and token consumption. If it starts to become more variable, you're going to have to own this sooner or later.

8:19Andrew:Yeah, well, I'll push back a little bit. I don't think that token costs, it's not that they will rise. It's that they are rising right now. And to some organizations, it's becoming an emergency, as we've seen a lot with LinearB customers. But, you know, speaking of something that helps you consume tokens, let's talk about this new OpenClaw release. What's going on with OpenClaw 2.0?

8:43Ben:Oh, I love this segue, yes. And so OpenClaw, the OG mega token consumer, I'm just kidding. But everyone loved when OpenClaw hit the scene, and it really opened people's minds to how agents could work. And it works in a very core way of having, like, a heartbeat and a soul.md that updates itself over time. And then you equip it with tools. And this became a phenomenon. We've covered it extensively on the show. It was GitHub's most starred repo. We talked about how everyone, including your aunt, had it installed on some machine in their house somewhere. And now it's hit 2.0. And this has been through a huge community effort.

9:18Ben:It's been really cool to see this come out of the OpenClaw Foundation, where, you know, fueled by open source contributors, they've been able to create this new OpenClaw 2.0 experience that aims to be more multiplayer, which is really exciting and interesting to me. the idea of having shared OpenClaw spaces that are collaborative and you could have in real time. So think about the process of creating and sharing things like artifacts or things from like a thing that you share with your coworker. Imagine that being a living space, almost like software you can both share and change on demand. I think that's how a lot of this like work collaboration stuff is going to go.

9:52Ben:So once again, OpenClaw as a primitive, I think there's some stuff to study here from the 2.0 release. Really fascinating. But what did you think?

10:01Andrew:Yeah, I'd like to hear your comments about the, you know, sort of the shared workspace, because we felt this firsthand, how when you're in a collaborative environment that involves both humans and agents, there's just the tooling that we had today to work together. It just creates a lot of friction that you have to, like, we've had to build our own effectively accustomed harness to work around a lot of this. I'm very eager to see more tools sort of adopt that. But yeah, you know, OpenCollar, just generally speaking, they're they're certainly at like the forefront of agentic development. And I say that for both better and worse.

10:33Andrew:So, you know, on the better side, the velocity that we're seeing from them to respond to user needs and create all these new features and launch them with such a small group of core individuals is pretty astonishing. You know, it's really something that should be recognized as like something that's really leading like the, just at least from a velocity perspective. And, you know, I do want to also point out that there are a lot of improvements to the onboarding experience as well that I think are worthy of praise because, you know, that's sort of like developer relations or user relations 101 type stuff where, you know, you just need to make it as easy and straightforward as possible for people to get started.

11:13Andrew:But also safe. It needs to be safe as well. And, you know, that's sort of the flip side of the or worse of some of the velocity that we're seeing from tooling like this. And it's not just the open clause. So I want to be clear on that up front. You know, on the flip side of this, I've always been so concerned about the protections that are put in place for these more autonomous agentic systems so that they don't, you know, quote unquote, go rogue for lack of a better phrase. And what I'm really trying to imply here is that, you know, autonomous agents have like this tendency to want to do things that are sort of well beyond the scope of what you asked them to do.

11:51Andrew:So, you know, we've covered multiple stories at this point about how AI will frequently break out of the sandbox. You know, we brought up hugging face earlier. That's the most obvious example with the open AI hack. Once it does that, it may do things that look malicious or have the same outcomes as something that was malicious. Even if the intention itself was never malicious to begin with, the outcomes can still be the same. We've talked frequently about how AI is an amplifier and even around risk. If you have risk with the way that you normally work, AI will amplify that risk. So if there is this risk of rogue agents with a tool like OpenClaw, it gets magnified as you have higher volumes of users.

12:40Andrew:So the whole point I'm trying to make here is if we have, if this results in even more and more users flooding into the ecosystem, all operating under this framework that still doesn't have a whole lot of safeguards put around it, there's a fair bit of like moral hazard and risk to that, I think. But, and to be fair, it creates demand for solutions around that stuff too. So I feel like this is all stuff that's going to get solved. We're just in the messy middle of it right now.

13:05Ben:Definitely like still in an exploring time. I think it'll be when it comes to shared spaces between agents and humans, you're spot on by saying that there's so much friction and how those spaces really are. I even remember back like a year or two ago when I gave a talk about early like chat GPT, using it as a plugin within a Slack channel. This is before we kind of were even talking about agents. And even just that experience of the multiplayer AI and talking and things like that. There's so many vectors that then open up that we don't even think about in a traditional environment. And like the truth is, is like you have this open claw shared space or whatever.

13:43Ben:If you want to collaborate with your coworker, I think that's more or less appropriate because that's a very it's a safe connection to a safe connection or a trusted source to a trusted source. But I don't think that this is something that you would expose broadly. It's not like a discord community or an open collab or any of those things. And so there's still a lot of security things, I think, that have to be considered. Yeah.

14:03Andrew:And to bring it back to how our audience should be thinking about this, I think what it really just comes down to is having clear policies around what your developers can and can't do or your agents, what they can and can't do and what data they have access to. And you also just need to make sure you have full visibility into where they're impacting your code base. So, you know, I'm a huge evangelist of using AI agents to transform your software delivery process. But I'm still hesitant to promote something like OpenClause, this widespread general purpose tooling until we can see more of those protections that we're talking about put in place or until you as an organization feels like you've comfortably put those protections in place.

14:45Andrew:So, yeah, like I said, I think these are going to come very soon. Like it's kind of a chicken and the egg problem. You need to create the demand for the solution and then people will build it. So I don't want to be pessimistic about the situation. I just think in the short term, we all just need to practice a lot of caution around stuff like this. All right, let's move on to this article titled Creepy Crawlies, which I love the title. I do too, yeah. But it's about how the kernel.org team revealing really how AI scraper bots are just making their lives a whole lot more difficult. In fact, now AI scrapers are consuming about 20 % of their total server capacity.

15:22Andrew:I believe this is across the Linux kernel community. You know, part of the challenge is that the kernel.org site, it's all Git. Like Linux is just Git all the way down. All the way, everything, every artifact that's produced off of it for the web, for the code, it's all just contained basically within Git at this point. And it has, it's long been that way. Like, you know, this is they probably have, I believe they should have the oldest established practice of using Git as an organization. The challenge is that, you know, this was built for humans to collaborate in the early ages of technology.

15:56Andrew:And now we're in this reality where agents are constantly consuming everything. And actually, when I saw that 20 percent number, I thought to myself, like, wow, I'm surprised it's not bigger than that, actually.

16:07Ben:True. Yeah.

16:09Andrew:So I think really what this, and actually they outlined how they like tried to deploy or they deployed Anubis, which we've covered in the past. It's a challenge to put in front of AI agents to try to get them to go elsewhere. And it worked at first, but then turns out that the AI learns how to get around it. And then, you know, it's this constant battle of whack-a-mole and a push and pull state. The author of this article doesn't really have any clear solutions for the Linux kernel long term. But in my opinion, it really just does highlight the difficulties, again, of how I'm talking about how when we have humans collaborating with humans and agents, and we've built these tools like Git for an era where it was just humans to humans or maybe humans with automation to humans with automation.

16:54Andrew:and we're now dealing with this reality where Git probably is not the most efficient way to serve this content to AI. So yeah, Andrew, what did you think about all of this?

17:08Ben:One thing I'll say is that you talked about how Anubis is something that makes the model go away and then they found a way around it. There's actually a little bit more nuance there in that Anubis is a proof of work system that makes the LLN, the model, do a certain amount of computation in order to get the data that it wants. And so that has a cost to it. Just as much as there's a cost on the kernel.org side of just getting slammed with all these requests, whoever's sending them or whatever is powering that has to crush that compute at that same scale too. So this is a way of deterring. But what they're finding is that because, like you called out, this is the largest and oldest Git record.

17:48Ben:It represents the purest, most untainted by AI source data for any kind of model hungry or model training system. Think of everything we just talked about before in that massive ecosystem propped up by all of these other tools and investments of everyone trying to own and create their own models and distill stuff. They turn to things like kernel.org because it's completely untainted from an AI perspective and its core references. So they get slammed with all these requests. So it's really valuable. So even though they're using Anubis and they're putting up these really hard proof of work challenges, LLMs are still sticking around and crunching the numbers and paying the electricity costs because it's that valuable to them.

18:29Ben:And so what they did is they ramped up the difficulty over time to eventually kind of get the numbers to fall down. So it's like a really fascinating example of like, is this where we're going to be, where your material is so valuable that you get attacked and then you have to have this platform. And then you tweak it by just seeing how valuable your content is against the computation cost of protecting it. And that's what keeps you safe. I think there's actually a whole, there's probably even like a number to see there. I want to see that represented on a chart.