This is one of the more impressive feats I’ve seen an LLM do. On the surface it doesn’t seem like it should work.

This is seriously crazy imo. Even if it's limited to smaller models, Astra (frontier, generally) being able to just one-shot a new model specialized for a specific task opens a lot of doors. Have a specific problem you want to solve but don't know ML or have a GPU powerful enough to train? Want to adjust smaller local-compatible models to specialize on your specific task(s) at the expense of things you never do? I can see this being very powerful for smaller locally run models.

I wonder if this works on models bigger than it? Can I have it natively patch my tiny voice recognition model to recognize my wife's accent better? Could Astra write a mini-model that outperforms it on highly specialized activities and then run inference against it to do things Astra cannot itself do?

Regardless of claims of RSI or AGI or whatever, this feels like a big development.