mini-AGI Made Me Build Tama

I found it on GitHub and started wondering what it would be like to build something similarly small on my Mac. I called my own experiment Tama. It is separate from mini-AGI, not a reproduction of its code or a claim to its name. I wanted to watch a model begin with almost nothing. Neither project promises AGI.

volotat calls mini-AGI a toy-level experiment. It is a small byte-level language model that keeps learning from the text it encounters. There is no pretrained model behind an API call.

With ChatGPT, Claude or Gemini, you meet a system already able to do a great deal. Tama begins with characters and fragments. Occasionally a pattern appears. Sometimes the next stretch of training makes that pattern less convincing. It is slow, and progress is hard to judge from a few less random fragments.

I can save its state, stop training and pick it up again on an ordinary personal computer.