To make his dataset, he uses OpenAI's API. OpenAI catches him using their outputs to train another model, flags his account for breaking their terms, and bans him.

He files an appeal, gets unbanned, goes right back to pulling data from the API, and immediately gets banned a second time.

So instead of giving up, he uses open-source tools to remove the model's built-in refusals, cleans out the preachy fluff, and starts building a fully local 9B agent.

OpenAI spent years scraping the whole public internet for free data, but the second someone uses their output to train a local file, it's an emergency ban.

In trying to enforce their rules, all OpenAI really did was give open-source models a massive free advertisement to millions of people.

What a time to run models on your own hardware.