Yeah, it just sucks at doing it. It’s also not a coherent being. Treating like a person is complete foolishness.
Well, it’s not working very well. It’s nice that they’re trying, though.
How does it determine what data is factual?
Because it doesn’t work that way. Your interactions with an LLM may, at best, govern its behaviour within a specific session. You can also change very specific behavioural parameters that will be common to all your own interactions with it in all sessions.
But in contemporary LLM architectures, all the training is done in phases prior to launch – pre-training, which involves absolutely massive ingestion of data which creates the base model, and then supervised learning, reinforcement learning, and safety tuning. You, as a user, don’t have control over any of that.
RLHF is one specific post-training fine-tuning methodology. But it’s not something the user can influence.