By Rolf Versluis · Published September 8, 2026 on Uncommon Knowledge in Business and Life

I run a lot of these agents, and they live on hardware I own. The work they process stays on the machine. I've used ChatGPT and Claude and Grok for customer emails, invoices, the whole thing. I know the work might end up in the training data the vendor builds the next model on. I didn't sign a contract. Just clicked a button.

What happened

A mathematician at NYU named Tristan Buckmaster found out. He and a co-author had been using AI tools, including OpenAI's Codex, to work on a problem unsolved since 2000. Days after they posted progress, OpenAI announced an internal model had solved the same thing.

Buckmaster said the timeline's suspicious. OpenAI told him the model didn't look up user data, then refused to answer whether his notes went into the training set, then acknowledged it can't rule out that they did.

They spent years on a problem worth a million dollars in prize money. The same tools the mathematicians used may have helped the vendor solve it first.

That's what happened.

The alternative

The big AI vendors trained on copyrighted work without asking the people who wrote it. Books, journalism, code, artwork, all of it absorbed into models the vendors now sell. I don't care whether the courts decide that was legal. What matters is the mechanism: the same mechanism that lets a vendor train on a novelist's book also lets it train on the proprietary knowledge someone has spent years building. The customer list, the pricing logic an accountant built around a cost structure, the supplier relationships no one else can replace, the procedures a team runs because someone wrote them down and trained the new hires on them. Those things are proprietary in the operational sense. They're worth money, they take years to build, they keep a business alive. None of that's copyrighted. A model can't tell the difference.

This applies to every cloud AI vendor. OpenAI, Anthropic, Google, the Chinese-hosted models that half the apps on my phone are quietly calling out to. Every vendor trains the same way and hides behind the same opacity. Naming three vendors and not the rest doesn't change anything. Just makes the list feel smaller.

There's a different option. I run open-source models on hardware I own. A Mac Mini in the office, a Linux box in the back room, a VPS I rent by the month. The open-source models have been getting better every year, and the ones running locally are good enough to do real work.

What's new is that the cost of not making the change just became visible in public, with names attached.

I started with an unused Mac Mini gathering dust in a closet. Installed a Hermes agent on it, pointed it at a local model, and spent a few weeks learning the ropes. Once I was comfortable, I moved it to my primary machine and added capabilities as I went. The work never left my machine. No central training process to hide behind.

None of my information goes into a training cloud I can't audit. None of it becomes training data for a model a competitor is about to license. The skills the agent writes for itself stay on my hardware. Plain text files I can read.