My question is: how could this idea be adapted to our current ML models?
For example, an LLM just tries to predict the next token when requested. How does one go from "next token please" to something that "dreams"?
For example, an LLM just tries to predict the next token when requested. How does one go from "next token please" to something that "dreams"?
It’s possible they could try training the model on some hallucinated scenarios to avoid over fitting, but I’m no ML researcher.