Hacker News Reader

So the externally provided metaprompt doesn’t establish a goal that generation works toward?

I don’t think LLMs work towards anything. It just picks from a list of likely “next tokens” at random.

“Likely” as defined by…?

Statistics as defined by training data.