Especially for a backup, smaller on-device LLM, I might want to offer the agent a code example to work from.
basic bouncing ball
fractal tree
Obviously the idea isn't for it to copy/paste but for it to dream up its own ideas, I wouldn't mind having this available. We could build a parallel "code reference" context, even using an embedding to find the most similar code example to what it is trying to do.
For the live performance I would just have 2-3 hand-picked examples for a future more open-ended context, I could give it the entire library of Coding Train challenges. (ShiffBot has this.)
Especially for a backup, smaller on-device LLM, I might want to offer the agent a code example to work from.
Obviously the idea isn't for it to copy/paste but for it to dream up its own ideas, I wouldn't mind having this available. We could build a parallel "code reference" context, even using an embedding to find the most similar code example to what it is trying to do.
For the live performance I would just have 2-3 hand-picked examples for a future more open-ended context, I could give it the entire library of Coding Train challenges. (ShiffBot has this.)