What is OpenWALDO and Why Did Greg Kurtzer Need to Create It?

Greg Kurtzer is biased. As the creator of CentOS, Rocky Linux, Apptainer (Singularity), and Warewulf, he clearly has a bias toward open source software, to the benefit of many HPC professionals. So when Kurtzer noticed earlier this year that a lot of people were talking about open-weight AI models as if they were actually open source, it didn’t sit right with him.

“It’s very unfortunate that a lot of people have confused open source and open weight,” Kurtzer said. “[Open-weight models] are licensed under an open-source license, but they’re not open source. And so the terms have gotten confused.”

To be considered truly open source, all of the source material used to create a given piece of software must be made available for anyone to use or inspect. For a piece of enterprise software like Rocky Linux, that means all of the source code is available for anyone to inspect and use under an open source license. For an AI model to truly be open source, it means that the architecture, weights, and training data are available for anyone to inspect and use.

That is not what we have with the popular open-weight AI models today, such as Llama-3, Qwen3, and DeepSeek 4. These pieces of software function more like freeware binaries, which are executable programs that have already been compiled. The most critical element that make up those AI models–the training data–remains a secret that nobody besides the model builder can access.

That is a problem for anyone who values security and legality.

“It’s a box and you don’t know what’s inside this box,” Kurtzer told HPCwire. “And the box could easily have malicious training content that they were using for training. So the best way of handling that is, well, if these things were actually open source, we wouldn’t have as big of a problem, right?

“So that’s kind of when it clicked for me,” he continued. “Wait a second. There is no real open source community and project around building AI models like this. And so that’s when I started working on it.”

The Training Data BOM

“It,” of course, is OpenWALDO, which Kurtzer announced in a blog post last month on the website of his enterprise software company, CIQ. (Kurtzer is the CEO and founder of CIQ, which is sponsoring the project, but it remains Kurtzer’s.)

The name WALDO is an acronym that stands for open weights, open artifacts, open licenses, open data, and open origins. “It basically is the fundamental making of everything that’s inside of an AI model to be open source,” Kurtzer said. “So OpenWALDO literally means open source AI.”

The idea with OpenWALDO is to create a repository where AI model builders can house their training data and provide downstream users with a bill of materials (BOM) that proves the provenance of the data: where it originated from, who collected it, and how it can be used. The BOM is critical for delivering greater transparency into AI model origins and removing the uncertainty that comes with open weight models today.

So if the BOM shows an AI model is 99.99% open source, but there’s a 0.01% piece of it that’s not open, the customer will be aware of that with OpenWALDO.

“And then the question is, well, do you want to go with that company? Do you trust the company?” Kurtzer said. “I can tell you, at least in the US, large corporations and in US federal, whether they trust the models or not isn’t part of the question. They just can’t do it.”

Chinese open-weight models, in particular, have achieved near parity with American closed-source models, tempting many organizations to use them. However, you use these models at your own risk, according to F5 Labs.

“In most cases, the training data is not disclosed, there are no guarantees of the integrity of the training pipeline, and there exists a risk that specific fine-tuning has been included for reasons that may be opaque,” the security firm wrote in a July 30 blog post on Chinese open-weight models. “By running these models locally, the user assumes the risk of using them and securing them, which can add significant barriers to deploying them. These risks are not insignificant.”

Building an Open AI Community

In addition to the BOM for the training data, OpenWALDO also provides tooling for AI builders and a community for them to collaborate, which are other important pieces for the future of open source AI.

“It’s also the community of people that are contributing to OpenWALDO, people who want to learn AI, people who want to take what they know about AI and see if they can adopt it into OpenWALDO and share it with others,” Kurtzer said.

The training data itself isn’t stored in OpenWALDO. Instead, the project uses a federated storage system, where contributors house their own repositories of training data and link it to OpenWALDO via Git, the open version control system that maintains the source code for millions of open source projects.

“You would be hosting the Parquet files themselves, which are the corpus shards… in your own S3 bucket,” Kurtzer said. “And then the index will be linking and pointing to that with the appropriate metadata for us to guarantee that that object has not been tampered with.”

The OpenWALDO Project will be able to validate every object based on the index, Kurtzer said. “We know it’s checksum. We know exactly what’s in it. We can audit it. We know where it is, who put it there, who signed off for it, the licenses–everything,” he said.

AI model builders would access the training data by calling a compose file. These compose files are written in YAML and define everything about the model the user wants to build: from the basic model architecture all the way through your pre-training, mid training, post training.

“The goal is eventually to level everybody up,” Kurtzer said. “So whether you’re a company who wants to just build a very small language model focused on a very particular task, or you’re a neo AI lab or you’re a frontier lab, the goal is to say, well, just go train on this data. Like, it’s free, it’s available. Any part of this data that you want to leverage, you’re free to leverage.”

Open Source for AI

In addition to avoiding the use of AI models that could be security threats, this approach would help organizations avoiding copyright violations. Today’s large language models are trained on giant corpuses, such as Common Pile, which are scoured off the Internet. How many copyright violations were committed as part of that scraping? Nobody knows.

Copyleft is another thorny issue that calls for greater transparency in training data. For instance, say an organization builds their own model that was trained on data protected with a copyleft license, which requires free use of any derivative work based on that piece of data. Does that copyleft license require the organization to publicly release their model?

The courts have not ruled on many of these questions, which remain legally up-in-the-air. In any case, in Kurtzer’s opinion, the AI boom is screaming for open source, and he feels compelled to act.

“Whenever I see a problem, I always see it through the lens of here’s how open source can fix this, and I’m seeing exactly the same situation in AI right now,” he said. “We need an open source community around AI. We need an area where people can come together, where they can talk about AI, where they can share the things that they’re working on, ask questions, get helpful responses, build different subprojects and build different things from that community. And we just don’t have anything like that.”

With OpenWALDO, we might.

Editor’s note: This article first appeared in HPCwire.

The post What is OpenWALDO and Why Did Greg Kurtzer Need to Create It? appeared first on AIwire.