1 month and 60 models later

romellogoodman2 pts0 comments

1 month and 60 models later · sup computerskip to content1 month and 60 models later<br>note · July 2026 · researcher: Romello Goodman<br>About a month ago I trained my first model<br>and caught the bug. In that time I've trained around 60 models and released 11<br>of them, each helping me along my path as I learn how to do AI research. Each<br>starting with a question:

shakespeare — Can a frontier model train a small model on my laptop?

gatsby — How does the dataset shape the response of the model?

kenosha-kid — Can the model memorize with slight variations?

daydream-chess — Can the model learn a game from the moves?

glyph — Can we use hallucinations stylistically?

With each release, I understand the material more and how to sculpt it. As<br>opposed to the knowledge-dense heavy trillion parameter models at the frontier,<br>these contain mere millions. They're knowledge-light and shaped like the data<br>you give them. Give it chess games and it can roughly play them. Give it<br>Shakespeare and it roughly writes you prose. This constraint gives me a<br>framework to work within: Brainstorm the data and shape the model.

But within each of these models, there's a mismatch between them and your<br>idealized language model. These models aren't fun to converse with. The<br>dominant interface for language models is the chat assistant/agent, and these<br>models are too small to fill that role. A back and forth with them would only<br>produce gibberish. So there is a need to invent novel interfaces for this new<br>class of models. Each one a harness shaped to the contours and skills the model<br>naturally has. A chessboard for daydream, a type tester for glyph.

The tie between the model, its harness and the interfaces we use can create a<br>synergy that turns even a single purpose model into an instrument to be played.<br>Each piece of the puzzle supports the other to surprise and delight the system<br>using it. This is the long-term bet of sup computer: creating instruments for<br>both humans and agents. Now that we have the models, the real work begins.

models model month later give trained

Related Articles