Thank you Robot, you've been a good ally.

32B is just too dumb. 7B was getting acronyms wrong while 32B is not getting basic concepts like 'secrets' right.

The goal here is to dump secrets into an obfuscation box of a LLM model and have it able to conjure up info based on pass phrases etc.

As long as you're not logging the console interactions and pushing the LTM data into the LLM formats, it's a pretty secure secret keeper.