Comment LLM AI models are compressed knowledge compilation (Score 3) 40
So, basically the weights of an AI model are a compilation of human knowledge, i.e. an encyclopedia - or its equivalent for the XXI century. They don't contain just what human authors select and have the capacity to write, but everything that they're exposed to and that is frequent enough to leave a trace in the weights.
Any content generation of an LLM which isn't guided by a human is merely a random walk through that compressed encyclopedia of everything, following the most well-connected paths more frequently (that's why long generations often get caught in loops btw). There's no real agency in those 'agents'.