[furyofantares]

I think what's making these hyperinformed hallucinators useful for coding right now is that fact checking is fundamentally a big part of our work. There's a bare minimum that must be done to even think anything has been accomplished at all, that it runs and does what we expected on the most basic input, but then we are also used to writing tests and reading code critically.

[pixl97]

I wonder what a learning model with the ability to test feedback from executing its code in an interpreter would look like? I know there are different groups testing things like integrating the ability for LLMs to use tools, wondering how this will pan out in the end.