I doubt it… At least not with the current transformer based models. The reasoning being that they have a heavy preference to parroting back the tokens in their training distribution which are 99% human readable things. If anything, I’m begging to think the code they write (and run) will be even more human readable. Here’s a conversational programming language called Convo that I had ChatGPT create over the weekend:
You can think of a Convo program as just a detailed design spec. Basically like the spec a PM would give a Developer. It expresses the goals, behavior, and even look and feel in detail as a form of pseudo code. The interesting bit is that you can pass one of these convo programs to an LLM and ask for a python version of a program. What you get back is much more consistent implementation of the program then if you had just said “create the game snake in python”. There’s still some variance but the resultant program is way more likely to have all of the features and look & feel that are outlined in the convo program.
In my SynthOS project I’m essentially using a variant of this to pass a program request from my top level voice agent to the programming agent powered by o1. This is effectively just a prompt and there’s nothing overly special about the format of a convo program. Any instruction prompt is technically a valid convo program.
At the end of the day… I believe that agents will essentially program each other by just talking to each other using language. That’s how humans program each other. You get your kids to execute a room cleaning program by saying to them “go clean your room”. I think agents will do the same thing.