[faangguyindia]

i actually do it differently

> (1) Let the LLM randomly perturbate the system.

instead of this i ask LLM to what's least likely to improve performance and then measure it.

sometimes big gains come from places you thought are least likely.

[fesens]

For sure! The hypothesis generation gotta be improved. Your take on the "least likely" is interesting. In the beginning of the repo I was having problems with "hypothesis convergence", your idea may be a nice way to introduce the much needed variability