Rocky Linux's Founder Wants to Open the One Part of AI Nobody Opens: the Training Data
OpenWALDO launched with 124 billion tokens across 75 million documents, each carrying a licence and an origin. The idea is that a model should come with a parts list, the way software does.
Gregory Kurtzer has spent twenty five years starting open source projects that other people ended up depending on, including CentOS, Singularity and Rocky Linux. On August 11 he launched a new one aimed squarely at the least open layer of modern AI: the training data. It is called OpenWALDO, it is sponsored by his company CIQ, and its public corpus already indexes 124 billion reference tokens across 75.1 million documents, organised into 20 collections with 18 asserted licence identifiers.
The name is an acronym for Weights, Artifacts, Licenses, Data and Origins. The pitch is that “open weights” only gets you halfway. When a lab releases open weights, you can download and run the finished model, which is genuinely useful, but you still cannot see what went into it. Weights are the millions of numbers a model learns during training; they are the result, not the recipe. OpenWALDO records where each piece of training material came from, what rights were claimed over it, who contributed it and exactly which bytes belong to which version. Models trained from the corpus can then ship with an AI Bill of Materials, a parts list that ties the finished model back to its sources, licences and training runs.
The machinery behind it is deliberately unexciting, which is the point. The public index lives in Git, so every change is reviewable and attributable the way code review has worked in open source for decades. The bulk data sits in federated storage and is verified by content hashes, so you can check that what you downloaded is what was reviewed. Contributors sign off under the Developer Certificate of Origin, the same accountability mechanism the Linux kernel uses. As Kurtzer puts it, they are not inventing new trust infrastructure, they are applying infrastructure that already works.
Here is the wider context. Almost every AI lab today separately collects, cleans and licenses much of the same underlying material, then declines to say what it was. That secrecy is partly competitive and partly legal, since the copyright questions around scraped training data are still working their way through courts. A shared, auditable corpus would let smaller teams skip the expensive foundational work and spend their effort on architecture and safety instead. It would also give anyone downstream an honest answer to the question labs keep dodging: where did this come from, and who said you could use it. Kurtzer is candid that what exists today is nowhere near frontier scale. 124 billion tokens is a starting point, not a competitor to the trillions the big labs train on.
What this means for you: nothing changes in the tools you use today. But if you have ever wondered why nobody can tell you whether your blog post or your company’s documentation ended up inside a model, this is an attempt to build the plumbing that would make that answerable. For businesses in the EU, provenance is heading from nice to have towards paperwork, so a corpus with licences attached is worth watching. And if you fancy training a small model on your own hardware, a fully provenance tracked pipeline you can inspect end to end is a rare thing to learn on.
Sources
You Can Now Hire a Humanoid Robot to Clean Your Flat, With a Human in VR Goggles Driving It
Tau Robotics opened an invite only pilot in San Francisco at $30 an hour. The robots vacuum and take out the bins. They also cannot climb stairs, and a person is steering most of the time.
