Self-learning aspect of agents · Brendan's AI Community

Self-learning aspect of agents

There is this idea of self-learning in AI. Am I smarter today than I was yesterday by learning yesterday's mistakes?

Does anyone have any practical suggestions on this self-learning?

I am seeing this tension. If the learning is non-deterministic, then there are issues like prompt hacking. Besides the vulnerability aspect, there have also been numerous cases of AI agents not doing specifically what was asked. This "may or may not happen" aspect of any plain English instruction given to an AI agent. It's a request not a command. The way to solve this is to convert the instruction from plain English to some coding language and run that piece of code. Usually python works.

But moving to python kills the self-improvement loop. All self-improvement options I have seen (SkillOpt for example) work on improving instructions in plain English. A request can improve over time but a command is frozen in the time it was created.

Wonder how people see this tension of being more efficient today (python code) vs being smarter tomorrow (Prose in a skill)?

8 comments

Self-learning aspect of agents