

10·
26 days agoIn this case the LLM “reacts” differently when it is being tested because that’s what’s in the training data. It can only follow the patterns in the training data. I remain convinced that without a new architecture entirely, awareness is impossible.
LLMs don’t learn, and aren’t developed in a way that would cause them evolve stratagies. They are just patterns distilled from a data set, only optimized to better represent those patterns or do so more efficiently.