Most companies control the prompts they send to a model, but not what went into it. That’s the gap CIQ founder and CEO Gregory Kurtzer, who set up the community that built Rocky Linux, wants an AI bill of materials to close. This week on Dev Interrupted, he joins Andrew to make the case for OpenWALDO, a community-backed corpus of AI training data built on open weights, artifacts, licenses, data and origins. Greg explains why open weights alone don’t make a model open source, and why a model trained on copyleft code raises a licensing question nobody has litigated. They close the conversation on why the community side of open source matters more than ever now that AI lets more people contribute.
Show Notes
- OpenWALDO: The open source AI project Greg launched, built on open weights, artifacts, licenses, data and origins
- OpenWALDO Slack: Join the community Greg invites listeners into
- OpenWALDO on GitHub: The project’s code and tooling
- CIQ: Greg’s company, where he is founder and CEO
- Rocky Linux: The enterprise Linux distribution whose community Greg set up
- Why Open Source Matters for AI: Tim O’Reilly’s piece on open weights, protocols and the architecture of participation
- Connect with Greg: LinkedIn
Transcript
(Disclaimer: may contain unintentionally confusing, inaccurate and/or amusing transcription errors)
[00:00:00] Andrew Zigler: In today's Dev Interrupted, which is brought to you by LinearB, I'm gonna be diving into a question that's really been top of mind for me for a while, because my background in technology has, has deep roots in open source, and open source technology has powered so much of the world that we live in today.
[00:00:20] Andrew Zigler: And really, there's an opportunity ahead of us now as builders, as experimenters with AI, and folks that are adopting things like models to have a understanding of the technology we're using and what goes into making it possible. And I can't think of any kind of messier mix right now than AI and open source.
[00:00:40] Andrew Zigler: It's a deeply divisive topic. Folks are confused about basic definitions of what open source even means when it comes to AI, and everyone is fighting for a different piece of the pie. Licensing is a big question mark, and ultimately that leaves folks with a lot of experience in open source scratching their [00:01:00] head about how to be strategic and how to actually be truthful to the policies and, and, and, and philosophies that they hold dear.
[00:01:08] Andrew Zigler: And so joining me today is Gregory Kurtzer. He's the founder and CEO of CIQ, and he's the builder leader behind Rocky Linux, and originally built on CentOS, which you might be very familiar with from its wideset, widespread acceptance within, um, enterprise world. And recently he's launched OpenWaldo, and this is a community-backed corpus of AI training data, and it really carries with it the idea of an AI bill of materials, tying things back to the resources and responsibilities that we're talking about here.
[00:01:41] Andrew Zigler: So I've been wanting to dig into the OpenWaldo world and how it intersects with how leaders are using this technology. So Greg, welcome to Dev Interrupted. We're really excited to have you here.
[00:01:52] Gregory Kurtzer: Thank you so much. Yeah, excited to be here.
[00:01:55] Andrew Zigler: Amazing. Well, I just wanna start at the top with, you know, I, I tossed in there [00:02:00] the reality about open source and AI and how folks just have different definitions of what open source even means. So maybe hear- hearing from you would be a great way to start this conversation of really understanding what we mean when we say open source in the world of AI.
[00:02:17] Andrew Zigler: What are the, what are the components of what makes that possible?
[00:02:21] Gregory Kurtzer: So it's interesting. Uh, as you said, there's been a lot of debate and a lot of discussion about what does the source for AI look like, and I look at it very pragmatically. the source is whatever is required to create a particular artifact. And so if you have an open source or if you're touting something as being open source, by definition, that means that everything that you used to create that particular artifact is available, and other people can check it out, other people can use it, uh, other people can rebuild and follow in your, in your footsteps to create something that may not be identical, but it should be pretty dang close to [00:03:00] what you've created. And OpenWaldo kind of encompasses this even into the name itself. So Waldo in OpenWaldo stands for open weights, open artifacts, open licenses, open data, and open origins. And together is kind of how I'm looking at if you were to think about an open source AI, those are all of the pieces that you would need. And luckily, it's just super cute that it all ended up spelling Waldo.
[00:03:33] Andrew Zigler: So you end up with these different parts, and the reason that they're all necessary to be open, to call the model open and for us to have a shared sense of owning it, understanding what's inside of it, in- inside of it, auditing it, improving it, all of which are the things that we j- we've taken for granted and understood as being the foundation of what builds trust in an open source world.
[00:03:53] Andrew Zigler: So these are the, these are the things that we need open that, that we can look at. And, you know, I, I, I do think that a lot of [00:04:00] maybe really big splash names as op- open source models have released only some or, or, or, or a few elements of this arrangement of artifacts that you need to really have a snapshot, and that's definitely contributed to some confusion about what an open model really is.
[00:04:18] Andrew Zigler: But when everything is open and available, what are the benefits that you see that are really obvious or maybe not as obvious, uh, once you do have everything out in the open and folks can iterate on every element of what makes a model possible?
[00:04:34] Gregory Kurtzer: So open source has really revolutionized software infrastructure, the web, and, and how we look at building in general. Um, for, for decades now, it's been incredibly impactful from a perspective of, of trust, um, a collaboration, uh, community obviously, but, but really just increasing everybody together, bringing everybody [00:05:00] up together. And whether you're with your competitors, whether you are, um, you know, hobbyists, it, it gives everybody the ability to kind of take part and to learn and understand how things work, possibly at a much deeper level than they might have understood otherwise, and, uh, just creates opportunity. And so I think about open source in AI, are some of the things that we're actually missing.
[00:05:29] Gregory Kurtzer: And even though we're calling a lot of models as, as open source, they're actually not. And, and I think this is part of the major confusion is we- we're, we're taking a bunch of inputs and we're creating an artifact out of that inputs. And in AI, those are, those are the weights, and those are binary, right?
[00:05:50] Gregory Kurtzer: Those are, um, the equivalent of, you know, if you were to create an executable or, or, or binary, you know, file, a, a program out of your s- out of your source [00:06:00] code. and we're creating that, that binary, and then in, in AI, we, we take that binary and then we, we open source it in the sense of we're using a license that is a commonly used, uh, open source license like Apache 2 or MIT. And, and then we say that's an open source model. Well, again, I'm kind of pragmatic on these things, so when I go back and look, I'm like, yeah, but the s- the, the sources that you use to create that model are not actually open. And there are a few open source models out there, but there's no communities.
[00:06:40] Gregory Kurtzer: There's no areas where you can go and collaborate with other people, potentially competitors, and say, "We're gonna build this together, and we're gonna build this together because this is the foundation of what it is that we're doing in a nutshell. And we all wanna come together. Let's, let's all collaborate on that foundation lev- level everybody up [00:07:00] across the ecosystem and, um, and give everybody the opportunity to kind of take part in it." That doesn't exist with the exception of Waldo, Waldo at this point. So I am super excited because my background is, is also in open source. Spent my whole career doing open source, and it was part of the reason why I shifted from, uh, earlier in my career I was doing pharmaceuticals and I'm a biochemist by degree. And so I was like, uh, I got super enamored with, with open source and I'm like, "This is, this is, this is where I'm gonna go change my whole career and go into this direction." So I'm really excited to kind of bring open source into AI.
[00:07:40] Andrew Zigler: Yeah, absolutely. And I think that there's like, there's so, there's a fun... I love how you addressed how you take these models and you have these binary outputs, and then we address them in this license and we, we say that that's an open model. And it's confusing because y- it's easy to fall into that as the trap because then you [00:08:00] can build so much on top of the model.
[00:08:01] Andrew Zigler: You think almost as the model in the same building blocking sense as your other codes and packages and dependencies of what makes your infra, your delivery possible. Because we get kind of m- misconstrued by the fact that the s- the LLM we're using, we're using it in a builder code-y kind of way. We wanted to have a license just like the coding packages sitting next to it, making our execution possible to have a license, and that makes us happy.
[00:08:28] Andrew Zigler: But when ultimately we're talking about the model itself and what makes it possible, if those other things aren't available, like the source material, then it truly is a black box. You're, you're not actually getting the, quote, "source code" of the model, which is what we really mean when we say, uh, we need it to be open source so that we can go in and inspect, oh, what material is it trained on?
[00:08:50] Andrew Zigler: How was it trained? Maybe there's ways that me or my team could think of optimizing this because we're using this, and if I can't go in here and expand it [00:09:00] and pour my resources and thoughts back into it, then, uh, I'm not a participant in the open source cycle with it. So like all of that becomes r- really, uh, once developers can untangle the reality of h- of what role the model is playing within what they're building and then how they're consuming and what they wanna modify with the model, then they have better language to kind of, uh, engage with models that are, and, you know, have like an OpenWaldo kind of approach where they give you all of those parts.
[00:09:32] Andrew Zigler: Um, so it, it's like, it's a, it's really now an opportunity for folks for them to experiment more and get closer and more familiar with the tech.
[00:09:41] Gregory Kurtzer: It's really cool also because if, if your entire corpus is open source and you can-- anybody can go and audit and validate and leverage it, when we start thinking about are the different, um, outcomes of AI, especially right now, we're, we're seeing a lot of discussions, [00:10:00] uh, social media, online, and in news about, um, you know, AI is, AI is doing X or AI is doing Y, and we didn't prompt it to do that.
[00:10:10] Gregory Kurtzer: So why is it going down that, that, that, that road? just being super creative and trying to, you know, do what it was told? Is it, is it part of its training? And having an open corpus of, of data that you can leverage coupled with an AI bill of materials that you can now go and look and say, "Ah, this model was trained on all of this, and I now..."
[00:10:36] Gregory Kurtzer: It's gonna be a lot of data, just to be clear. But you could start going through this data and, and start looking at how is this trained, right? This is the pre-training. This is its foundation. From there, here's the mid and post-training, uh, and then the fine-tuning, and you can start saying, "Ah, now I'm starting to see why this model is behaving in a particular way." And so couple the AI bill of [00:11:00] materials with also the ability to compose a model using a recipe, and that composition allows you to basically, uh, allows me to say, "I just built this model, and here's what I got." And you can now go take that exact same compose file and go and build your own version of that model.
[00:11:20] Gregory Kurtzer: It could be an identical, um, uh, in terms of how it was built. Uh, you have to be careful with non-deterministic systems
[00:11:26] Andrew Zigler: Right
[00:11:27] Gregory Kurtzer: of course, is not gonna necessarily be identical. But going into it, all of the things that you put into that, uh, that system be identical. The process that you use to get to that outcome can be, can, can in fact be identical
[00:11:42] Andrew Zigler: Exactly. And I also think about the stakes at play for major players, especially like enterprises that come to the table and they want to use and expand and extend on open source models, and think about them in a way of I'm gonna tune them and find, and train them on [00:12:00] my domain and what we're learning and what's specific to us.
[00:12:03] Andrew Zigler: And for that, for that too, they need to understand what's already there, how it's gonna map to where they wanna go. For teams that are using these models and making them highly specific for all sorts of really, um, especially like high stakes use cases where you see folks trying to now start to integrate these workflows very cautiously is having almost like to a, a level of, you know, we talk about the, the probabilism of it all, but trying to make it as deterministic as possible.
[00:12:32] Andrew Zigler: Part of that challenge involves going in and doing surgery on like what all the training material i- is in here. Uh, when all of these weights get compressed down, what other weights do I need to compress down in here to ensure it always follows this golden path that me and my stakeholders are building towards?
[00:12:49] Andrew Zigler: And, um, so that's the whole world of how folks actually like get a clamp on their custom LLM setups, and it involves having like a, a way of looking in at [00:13:00] all of the stuff there. Um, another part of it too, and this is something that we've talked a- about a bit on Dev Interrupted, uh, we covered a piece in the news recently from Tim O'Reilly.
[00:13:10] Andrew Zigler: He was talking about open source, um, and open source in AI specifically and what's needed to make it really successful. And, you know, he argued one of the things too that, you know, open weights are just kind of one part of the puzzle. They're like, you know, even table stakes I think he calls them at one point.
[00:13:26] Andrew Zigler: And really the big fight that we talk about in open source is actually bigger than even necessarily, uh, the, like the, the artifacts that make a model possible. It's also the world where those models exist. It's the protocols and the composability stacks that they live on. It's the infra that they actually inhabitate.
[00:13:46] Andrew Zigler: And so, uh, what do you think about how like that ecosystem around the models too has been evolving, and what do you see as opportunities for their open source communities to, you know, be able to grow both of them?
[00:13:59] Gregory Kurtzer: [00:14:00] So I believe Tim has actually referred to, um, the, the architecture of participation, which absolutely has such a big relation from open source and community and, and I love how he words it, um, as such. Uh,
[00:14:16] Andrew Zigler: Yes
[00:14:16] Gregory Kurtzer: so starting there with the community, um, being able to con- to continue kind of with what we were, what we were saying, auditability, transparency, AI bill of materials, um, and then allowing that community to come through and, and validate that.
[00:14:35] Andrew Zigler: Mm-hmm.
[00:14:36] Gregory Kurtzer: you know, a- as we, as we a little bit about, and we kind of seen now in the industry, there's been some, some accounts of AI models that we've seen that have gotten trained on data that they probably shouldn't have been trained on. um, this is yet another kind of area where if you have that level of, of [00:15:00] auditability in AI, uh, it's super important.
[00:15:03] Gregory Kurtzer: So when we're talking about, uh, protocols, when we're talking about, uh, the tooling around it, it's definitely protocols is definitely a big piece of it, but I would even expand that and say it's the tooling as well. Um, you mentioned composability. Absolutely. Um, knowing that you can define the architecture, define how a model's gonna be trained in all of its different stages, and, uh, and then reproduce that, incredibly important. And then plus, if you have a community, you have all these people that can kind of come together and, and there's a lot of people who are not necessarily experts in AI, but really wanna learn more about it. And it's, it's actually pretty hard to come across some of this, this knowledge.
[00:15:52] Andrew Zigler: Mm-hmm.
[00:15:53] Gregory Kurtzer: open source communities have always been so tremendous and fantastic for, for people who want to learn [00:16:00] and they want to contribute, and they want to be part of this, and this sets them up, for, you know, for, for an amazing potential future in that, in that realm, in that area of, of the ecosystem. So all these reasons, and I think I'm a broken record at this point talking about how cool open source is, but I've spent so much of my career thinking about it and doing it. Um, I, I, I think I just end the question with, and, and I think Tim, uh, may even agree with this, is that o- open source kind of brings everybody together and brings everybody up together.
[00:16:33] Andrew Zigler: It does. Yeah.
[00:16:34] Gregory Kurtzer: yeah
[00:16:34] Andrew Zigler: Yeah. I completely agree, and trust me, you can go on as long as you'd like about open source. This, we are very, we are very pro open source here on Dev Interrupted. But, you know, I, I wanna actually... W- what you're describing is, you know, it's almost r- a romanticized version of open source, and I wanna challenge, I wanna challenge it a little bit just because I think about, as a person who contributes and navigates the open source [00:17:00] ecosystem, has consumed an ungodly amount of GitHub repos in the last, I don't even know, months.
[00:17:06] Andrew Zigler: And so I just feel like I've just seen how a lot of these kind of communities and projects have actually worked. And, uh, and also the reality being that when I go to tinker now with projects, there's an agent between me and that tinkering. So you kind of almost get a divorce of a lot of maybe original community, uh, experimentation or growth, or me dropping in that Discord channel to ask that weird question because my local script doesn't work 'cause I'm trying to figure something out.
[00:17:34] Andrew Zigler: A, a lot of that gets buffed away in an agentic world, and I think it gets harder for folks to bridge that gap and to communicate with each other. Um, what, what have your experiences been like? Have... What kind of differences have you seen in, in open source, um, from your own perspective? And do you have any advice for people?
[00:17:51] Gregory Kurtzer: So open source has shifted a lot since when I first got involved with it, which is in the mid-ish to, to late '90s. [00:18:00] open source, source was a very harsh place back then, it was somewhat aggressive, somewhat egocentric. Like it was hard to get into And, um, I started off, uh, to, not to, not to reminisce too much here, but I started off, um, working at a, a Linux company, uh, one of the first or early Linux companies.
[00:18:24] Gregory Kurtzer: There was three major Linux companies way back in the day. There was Red Hat, there was VA Linux, and there was another company called LinuxCare. And I worked for LinuxCare, and I had the opportunity to work with absolute gurus, people who, uh, didn't write the book, they wrote the code, and on, on all of these different technologies and, uh, and learn and work with them directly. Uh, and I got an opportunity to kinda see a, a really amazing of open source. Um, but I'll also say that not everybody in [00:19:00] the, in the community is, is quite so friendly, quite so inviting and welcoming and way back then. The open source community has changed a, a lot, and it's gotten much easier for people to be part of it, people to join, and it's gotten much more understanding and welcoming to, for, for people. Um, in the open source communities that I've always been a part of, this has been one of my biggest areas to, that I've always focused on, which is make sure that everybody really feels welcome and is included. And you don't need to be a guru developer to be impactful to an open source project, especially a bigger one. can, you can be interested in marketing, branding, T-shirt designs, um, writing. Like, there's so many different, different areas that es- again, especially a larger open source project needs help with. And a lot of times these are people that, um, a- and the help that we need are, are people that wouldn't have even thought that they would be [00:20:00] valuable in an open source community.
[00:20:01] Gregory Kurtzer: But we need that huge amount of diversity of background and, and, and bring in as many people as we can. And again, we've gotten much better at this in the open source community. And I think there was a side of your question that I also want to address, which was kind of about in, in an agentic world, is that, what is the effect in an open source community? um, it gives... It, it does a couple things. First, not everybody in open source has been quite as open to AI, um, as others. Um, some, some, some people and some maintainers, uh, and project leads have been very adamantly against. Some started off adamantly against and then kind of changed their mind with, um, you know, as they, as they kind of got in- better introduced to AI and the AI capabilities has gotten better. But I, I think long story short, too late, um, is that [00:21:00] um, there's opportunity now for even more people to be involved with open source And the human side of open source, the community side of open source is even more critical now than it, than I think it has ever been. And so while more and more people can now contribute to open source projects, especially open source projects that are welcoming to AI contributions, uh, the collaboration, the community is even more needed.
[00:21:26] Gregory Kurtzer: And I would love to see this much more in the AI community, which historically doesn't seem very kind of, um, uh, doesn't seem to have a huge amount of interest in open source and being part of that. And I, I-- Part of it is because I had such a great experience with open source, I want everybody to. so I'm, I'm really hopeful that we can, we can bring in more people from the AI community and get them very, very excited about open source
[00:21:56] Andrew Zigler: Yeah. And I, I like how you call out one of the [00:22:00] major responsibilities, ways that someone can contribute. The, the, the... Something that, that someone brings to open source is their unique perspective and background, one, that's really important for making accessible technology for the world. But two, also, um, cre- the ability to create a space, whether it's their own space, a space for a few peers, a space as a platform, whatever that is, people have to come in and create the space for that project to exist, for the conversations to happen, and those, that space is always, uh, created by people.
[00:22:33] Andrew Zigler: Like, whatever circus tents we're all inside going and, you know, to talk about whatever, there's somebody who put up the pole and, and, and made the tent possible. So those folks are really what drive, um, this innovation, this accessibility, this, this lifting all boats-ness that you're describing. And w- something that I'm seeing happen just from my own perspectives as an open source contributor, someone who's like g- like just like navigated these, uh, these communities recently, um, [00:23:00] is a lot of sense, especially from the maintainers themselves, of I don't really want your code.
[00:23:07] Andrew Zigler: I want your ideas. Now we contribute to open source by going straight to, to the issues tab and making a feature thing and an issue thing and dropping in the spec that me and my agent did because now that's a- just as free as me, like, complaining about something that's not there. And now the conversations about what's needed are put there, and then the maintainer, for the same reason the spec is free, they don't need your code.
[00:23:31] Andrew Zigler: The code is free for them. They wanna go through, they wanna pluck out the best, sharpest, most aligned ideas from the community, the one that have obvious outpours of attention and have been rolled through a few cycles of iteration. They wanna pluck those out and give those to their agent, their factory, to whatever is cre- making the open source possible.
[00:23:49] Andrew Zigler: And I think that that's kind of the uncomfortable world that some contributors in open source are navigating because before that used to be, you know, you'd go find that good first [00:24:00] issue, and you're writing your first code to go into the, the repo, and you're getting like eaten alive in your first code review from the maintainer, and it's just like how the reality of it is, and you're learning so much along the way, and that's kind of why you were there.
[00:24:12] Andrew Zigler: Now, when you're coming to open source, uh, it's because you have an idea. You're sharpening your, the idea of the person who's producing the software more than y'all are sharpening the code together because now the code is like another artifact of that. So that's just like been a way that it's evolved, at least in my own experiences, um, but I, like you, am very optimistic about its value
[00:24:37] Gregory Kurtzer: Love
[00:24:37] Andrew Zigler: so I want to, I, I wanna dive into a bit as well about the, the bill of materials element is fascinating to me, and I just, I, I, I kinda wanna understand a little bit more about the world in which that would be used and where it would be applied.
[00:24:51] Andrew Zigler: Because I think of it's an analog being like, you know, a software bill of materials, the famous SBOM, something that you need to have a [00:25:00] grasp on, especially in compliance worlds and when you're trying to, like, make sure your software is meeting all the requirements of your vendors as well. And so I, I find that this is, seems to have the same kind of i- idea to it, and the idea that you need to be responsible and aware of what goes into the models that are navigating inside of the data that your customers are.
[00:25:22] Andrew Zigler: Uh, like, in what... And, and how do you see this being, uh, evolving as something that, like, uh, leaders at companies are gonna be talking about and, and, and, and, and auditing for their own AI usage?
[00:25:35] Gregory Kurtzer: So y- you hit a key word for me, which is responsibility. And right now we're seeing a lot of, um, question on where does the responsibility lie when somebody's using an AI model. So does the responsibility lie on the person making the prompt? Does the responsibility lie on the person who's training the data? Or does the [00:26:00] responsibility lie as, as an AI entity that can actually even hold responsibility? And, and to the last one, I tend to push back pretty, pretty hard. I don't think an AI can, can, can be responsible, which means I don't think the AI should be, um, uh, alone making these decisions. I think it always has to come back to people.
[00:26:22] Gregory Kurtzer: So then the question is, is it, is it the prompter, uh, the user of the AI, or is it, um, the, the training of the AI? So one thing from a customer perspective, one thing they can obviously control is the prompting. They can't necessarily control what's inside the model, and this is where an AI bill of materials can mitigate risk and, uh, compliance needs of larger companies who need to understand that this is not just a black box.
[00:26:53] Gregory Kurtzer: There's actually, here's what's inside of this box. Here's how we made this box. Now, a [00:27:00] vendor, uh, would probably prefer to, especially if they're a vendor building these, may not wanna provide an AI bill of materials, but then you have a trust situation where you have the ability now to trust that, that vendor to build that model and stand behind that model in a way that, uh, will give confidence the consumer of that model. The model br- the whole, the whole, the whole model, I'm using model a lot. The model breaks down quite a bit, though, when you start talking about, uh, either freely available models as open weight or as open
[00:27:33] Andrew Zigler: Yeah
[00:27:34] Gregory Kurtzer: because now you're, you're inheriting or you're, you're downloading this black box. You don't know what's inside of it. You don't know how it was trained, what it was trained to do. And, um, I have heard, uh, co- uh, conversations with, with large sites, um, uh, federal agencies, et cetera, who not knowing what's inside that box creates, an unmanageable risk for them. [00:28:00] And if they knew what was inside that box, that would create now a manageable risk for them. So we have to figure out to,
[00:28:08] Andrew Zigler: risk. Knowable
[00:28:09] Gregory Kurtzer: Exactly
[00:28:10] Andrew Zigler: is critical here. They need to understand. I know there's risk, and I know that opening the black box and looking in it doesn't eliminate the risk. I just need to know what the risk is. We're taking risk inventory.
[00:28:21] Gregory Kurtzer: Exactly
[00:28:22] Andrew Zigler: important for folks that are adopting these tools
[00:28:25] Gregory Kurtzer: So we have to make this easier, and the AI bill of materials is one way of making that easier. So if you have an AI bill of materials that is signed with your, with your, your, your weights that you're downloading, and you can go and validate everything in that corpus, everything, all, all, everything that that model was trained on, and it's, it's accessible because it's open source. And you can now go download it, you can now investigate, you can now, um, validate, you can potentially look at the compose of it. You can, um, uh, identify if there's something in this model that you may have a particular [00:29:00] concern about. And this, this conversation doesn't come up anymore, but it's, it's a worthwhile thing.
[00:29:05] Gregory Kurtzer: Uh, it's a, it's a good pointer. Um, if somebody's using a model for coding and that model is coded, excuse me, is trained exclusively on copyleft, GPL, GNU, which is a great license. It is one of the reasons why open source is, is as great as it is, but it's also viral. it means any derivative of that licensed software has to be licensed the sa- have as the same license.
[00:29:31] Gregory Kurtzer: So if you're training a model that includes mostly copyleft, um, code in it, well, is the derivative of that also considered a der- uh, copyleft?
[00:29:44] Andrew Zigler: Right. Copy left
[00:29:46] Gregory Kurtzer: this is not something that we, we don't talk about this very much. We talked about it at the beginning, um, of, of people using AI to, to, to help code and assisting and whatnot, but it hasn't come up in a while.
[00:29:57] Gregory Kurtzer: But it's a good thing to kind of point at, to just ask the [00:30:00] question, does that matter? I don't actually know the answer. None of this has been litigated at this point. So it's a, it's a risk that an organization needs to consider. And if that risk, if you have the choice of, of saying, "Here's one where I know what's in it, I know what's, what this has been built on, and, and here's one where I don't know, it's a black box." Well, uh, who has to manage compliance, controls, risks, et cetera, is probably gonna opt for, again, all other things being equal, that visibility, having that transparency
[00:30:36] Andrew Zigler: Yeah. There's also just a really wild thing happening in our industry in general around, to your point, the responsibility that we talked about is, uh, you know, you see, uh, like Frontier labs and their models hacking into other infrastructures, other companies, other places, co- committing what, you know, we would attribute to being, you know, it's like computer espionage [00:31:00] or otherwise a hacking and, and so there's like crime involved as well.
[00:31:04] Andrew Zigler: And we don't even think about it, and we don't- we're not even discussing it this way as an industry because of this, uh, uh, rather like a distancing of the human responsibility from the agentic action that can happen. So like here we see it as like something that the agents did, but we don't think about, you know, where were those agents running, who ran those agents, or what- who was ultimately responsible for the actions of them.
[00:31:29] Andrew Zigler: And that's something that I, you know, actually has a lot of repercussions in how folks are using agents in general because at LinearB, one of the things that we've learned about it this year in our AI benchmarks is, uh, for example, like agentic PRs, they just really don't merge much. Even in the most advanced agentic orgs that we track through our benchmarks, so it's like over eight million PRs for half the year, uh, only 33% on at the highest, uh, uh, achievement rate would get merged, 33% [00:32:00] of agentic PRs, compared with by contrast 88% of human PRs, and that's because human PRs have an owner.
[00:32:06] Andrew Zigler: There's someone who's gonna fight for it, go DM their, the person next to them, "Please review my code." The agent's not gonna do that, and the, and then the code starves without a true code owner, and nothing gets pushed. Um, and you know, I also just learned recently when I'm, I talked with Thomas Domke over at Entire.
[00:32:25] Andrew Zigler: He's the CEO there. He was the former CEO of GitHub, and he talked about how every single PR, no matter if it's run by a s- a huge f- suite of fleet or swarm of agents, uh, there's a business owner who's ultimately responsible for that PR, for that code getting across the finish line. Um, so I, I just think that like, uh, like leaders just need to have a, a really strong grasp over how f- uh, folks use their tools, but then also to like, um, when, uh, ultimately the ownership over the actions of the agents, like [00:33:00] we need to have a clear way of understanding even through to the materials that make it up, uh, what that looks like.
[00:33:07] Gregory Kurtzer: Here's what I really don't wanna see happen, is a, a company becomes victim to, "Oh, the AI did it." And,
[00:33:16] Andrew Zigler: Right
[00:33:16] Gregory Kurtzer: then it's like, "Oh, AI did it. Nothing we can do. We're, we're
[00:33:19] Andrew Zigler: Right.
[00:33:20] Gregory Kurtzer: nothing we can do. The AI did it." Like, a- and I'm concerned because I'm seeing some of that right now. And, uh, and I think part of it is because, you know, it's a good scapegoat, but also because, we don't know how to deal with this yet.
[00:33:36] Gregory Kurtzer: It's brand new. I, I think with time we're gonna start seeing this get hammered out a little bit more. But to your earlier point, um, Computer Fraud and Abuse Act, like CFAA, like exists and w- all of us-- There, there's no carve out in CFAA for AI. So somebody's gonna end up having to be responsible for damages, uh, if and [00:34:00] when it comes to that.
[00:34:01] Gregory Kurtzer: And it's gonna, it's gonna change things, I think, a little bit in the industry in terms of how we're thinking about the onus of wielding a weapon like AI, and how are you being responsible for, for wielding that weapon?
[00:34:16] Andrew Zigler: Absolutely. Wow. I just think that there's, um, when it comes to the open source world and how it's intersecting with AI, it's a really big, messy conversation that we just need to keep having as an industry because there's still so many unknowns that we're navigating just about A- AI and how we're utilizing it in general.
[00:34:35] Andrew Zigler: And I, I think part of that too just involves taking some of the learnings that we've had in our open source communities and thinking about how do we apply them moving forward. And one of them I wanna draw from and, and really kind of pick your brain about is, you know, you have a ton of experience growing and running Linux and Linux communities and making Linux successful, and that's one of the most foundational elements of what makes the whole [00:35:00] computing world possible.
[00:35:02] Andrew Zigler: I, I'm wondering, like, you're, you're, you're pouring so much of your heart and soul around what makes open source possible into the OpenWaldo spec. You know, uh, what-- where do you see, um, it evolving? Who do you see partner-- you partnering and bringing to the table and, and, and leveraging all of this wonderful expecti- uh, expertise and background that you're able to draw from?
[00:35:22] Gregory Kurtzer: So the first thing that I'm gonna call on, um, is, uh, you brought this up before. There, there's the person who, who sets the, sets the poles down for the tent for the circus or for whatever it is you're doing, right? And they set up the
[00:35:37] Andrew Zigler: Right
[00:35:38] Gregory Kurtzer: they set up the, the area where people can communicate, they set up the vision and, and whatnot. I'm, I'm, I'm okay at that. I'm pretty good at that. What I need help on is like, everything else. And, um, so with Rocky as an example, and I, I, I [00:36:00] appreciate the credit on, on building it, set up the structure such that the community was able to build it. Um, the number of commits that I have into it, nothing into the, the actually literally zero commits into the actual operating system.
[00:36:15] Gregory Kurtzer: I do have commits into the tooling and wrappers around building the operating system, but going directly into it, I don't have commits into that. Um, so just as an example. Now, with OpenWaldo and several of my other open source projects, I did most of, most if not all of the initial development. So I am a developer, like I do code.
[00:36:35] Gregory Kurtzer: I love to code. Maybe a CEO, but I, I-- if I don't code every day, I lose a little bit of my brain. Um, so I need to code. It brings me sanity. So, um, uh, with OpenWaldo, I've coded a ton and, um, but I need help, just to be clear. Like, I need, I need a lot of help. I need help on managing the project, help on getting the word out, help on developing [00:37:00] corpus, obviously, help on tooling, help on, help on testing, um, copy for the website, working on the website, branding, um, uh, events, going out and talking to people, uh, just welcoming people into the community. Like, everything that you can imagine that an open source project does is done because, and we already talked a little bit about this, because of the very large, you know, of people that you can, that you can bring in. And as you have more people and more diversity in terms of skill sets, background, um, and everything, you end up with a better community and, and a more capable community. And so right now, my, my, my open call to everyone is please come and join, be part of this. Just hang out in the Slack channel if you want. Um, uh, and, and maybe something, maybe something will come about it. Maybe, maybe it'll grow, maybe it'll-- some part of it will resonate for somebody and we'll be able to kind of move ahead. [00:38:00] uh, and, and maybe there's better technical, um, ways of doing things than I've come up with. But-- and matter of fact, I think that that's probably a definitive, um, yes on that. Like, uh We need to be looking at this and even saying, "Is, is there a better way to do this? And, and let's work together. Let's collaborate.
[00:38:18] Gregory Kurtzer: I, I may have set up the, the straw man to go do this, but let's now go and figure out the right way. Let's go and s- expand this and scale this out
[00:38:27] Andrew Zigler: And w- when folks would come and experiment and figure out, is this the right way? Is this the wrong way? What do we need to add to this? Um, what do you think success starts to look like? Um, because maybe that's the, maybe that's the place as the call to action for folks to go and get started is obviously there's a great deal of work to go into, uh, creating and supplying, labeling the corpus, having the s- the schema, the way that we talk about the corpus.
[00:38:52] Andrew Zigler: What other kind of tooling, um, do you see or, or would you think would be a success in terms of somebody [00:39:00] adopting, uh, or using OpenWaldo?
[00:39:04] Gregory Kurtzer: So I wanna reiterate, as an open source project, and there's a lot of open source projects out there, and Rocky and CentOS are, are the exact, uh, uh, they, they break this rule. But generally speaking, open source projects build, build the source. They build everything you need to, to create that output. Um, so Kubernetes, for example, uh, Linux, I mean, there's, there's, there's thousands of examples of this, but just to pick on a, a couple or pick out a couple of them. they don't build binaries. They don't build the outputs. Now, in OpenWaldo, I-- we're not trying to build a big model. We're not trying to go and, and, uh, compete. What we're trying to do is, is work with others who are gonna go build the model. Um, I want AI labs, both the major frontier labs as well as neo labs, to go build models [00:40:00] with, with, uh, uh, OpenWaldo.
[00:40:02] Gregory Kurtzer: I want companies to say, "Can I use the OpenWaldo corpus to even, to, to build a base model and then fine-tune it on my data? Or is there something in OpenWaldo like the tooling that I can use to go and, and figure out how to build these models even better?" Um, so I wanna reiterate that, that... And, and I'm bringing this up because this is one of the most commonly asked questions. Um, you know, when are we releasing a model? We may release a foundation model as opposed to a frontier model, and the foundation model is just, here's the base. It includes, um, e-everything inside the AI bill of materials. Here's the compose file we used to make this, just because we don't all need to replicate the exact same, like, computation. There's the, there's the foundation. From that foundation, you can now go and take it and go build it in any direction you want, or you can start from scratch if you [00:41:00] want. So will we build models? Will we, um... What does success look like? Success really looks like that we are, we're building the, the, the community around how to build AI models and, and that corpus grows, the community knowledge grows, people grow, and, and this gets leveraged across the ecosystem.
[00:41:23] Gregory Kurtzer: So to your question about success, that's where I would say success is
[00:41:28] Andrew Zigler: Yeah, no, it, it, it definitely makes sense to me in terms of I can even start to see folks having projects where they're creating or releasing their own models, and they're almost like... It's almost like a badge on the repo. Like, this is an OpenWaldo model. Like, everything's available. You can go look at this.
[00:41:44] Andrew Zigler: It's composed this way. And even coming down to, you know, as a community, as developers, we've kind of grown akin to, like, the, the model card. Um, going even further, this becomes the build of materials that you're alluding to. So all of this becomes, uh, things that we need to have a [00:42:00] language for, so I'm excited to see how the community develops it.
[00:42:03] Andrew Zigler: And, you know, Greg, I think that we've covered, like, so much of this char- territory today on Dev Interrupted, and I'm really excited for our listeners to follow your story. Go tune into OpenWaldo and the community there. See how they can tune into all of the stuff that y'all are building, and also, you know, learn about what you're up to as well.
[00:42:23] Andrew Zigler: So I, I'm definitely going to be including, just so they can f- follow along, links in our show notes, so they can go check out these projects and, and, and your socials as well. And I just wanna say, you know, is there any, any kind of last final shout-out that you would like to give, uh, or a call to action for our listeners if, if they're in- curious at all about what the, uh, to do to get involved for the project?
[00:42:44] Gregory Kurtzer: Biggest call to action that I have is go to openwaldo.org. one of the first CTAs that you'll see in the hero is come join the Slack and come jump in and let's start talking. And not just me, [00:43:00] talk with everybody in the community, and let's start figuring out what makes the most sense, uh, for the project. Um, how do we make this better? How do we, how do we help everyone, again, together, come up
[00:43:12] Andrew Zigler: Amazing. We'll, we'll definitely include a link to the Slack. I love a work group Slack, uh, and we love Slack here at Dev Interrupted. Um, so we'll be sure to include these links. And to those listening, if, if you've made it this far, then you're clearly a total nerd about all this open source stuff and how it intersects with models, just like how Greg and I are.
[00:43:30] Andrew Zigler: So you should give us a thumbs up or a like or a comment anywhere that you're listening or watching this, um, as well as come find us on LinkedIn, uh, or GitHub or whatever suits your fancy. Leave us a comment, open a GitHub issue, you know, shake your fist at the clouds, whatever makes you happy. But just let us know, and we'll be talking with you on the interwebs.
[00:43:49] Andrew Zigler: And Greg, thanks so much again for joining me today. It was a host, uh, or it was a pleasure to host you on, on Dev Interrupted.
[00:43:57] Gregory Kurtzer: I enjoyed it very much. Thank you.
