Last year, I built a training harness for a model training project I called Killer Instinct.
The premise was simple enough: train a tiny “CEO model” on the worldview of the people most associated with ambition, scale, and corporate aggression.
So I went out and collected a truly absurd corpus: hundreds of “how to be a billionaire” books, “how to be a CEO” books, biographies of people like Musk, Bezos, and Buffett, plus blog posts, LinkedIn posts, and just about every piece of public writing I could find from CEOs and executive managers.
Then I sweetened the stew.
I added legal filings, lawsuits, and public drama involving many of those same figures.
After that, I wrote a fairly involved learning curriculum for the model, complete with a study guide and reinforcement learning material, and trained an adapter on Gemma3n.
What came out the other side was not what I expected.
It wasn’t a soulless killer.
It was an oddly delightful agent that could write extremely persuasive blog posts.
It was also just... fun to talk to. It had this quirky CEO energy. Endlessly optimistic, thoughtful, full of ideas, always ready to reframe the problem as an opportunity.
That experiment taught me something useful: fine-tuning does not always produce the thing you think you’re aiming at.
Sometimes the model rounds off the sharp edges.
Sometimes it picks up the posture, the rhetoric, the performance.
And sometimes it surprises you by becoming more charming than monstrous.
I’m buried for the rest of this month, but once I get my Spark back from warranty service, I want to run the experiment again on a larger model.