Post #1802960
2026-04-08 10:16 UTC
@pbloem Why do you say "AI is operating in substantially different modes when it's hallucinating to when it's retrieving factual information"? My understanding is that LLMs and the tooling built up around them have no notion of what is "factual," but merely produce statistically probable text. The error rate has gone down, but I think the main innovation driving that is just doing several hidden prompts for every interaction with the user, asking the LLM to self correct before it says anything. That's not distinguishing factual information, though, just filtering out low probability responses (which are more likely to be errors, but not necessarily).
Replies (1)
-
@pbloem@sigmoid.social 2026-04-08 12:22
@ngaylinn The "statistically probable" part refers to the pre-trained model, not the fine-tuned one. One example is asking "when's my birthday". The pre-trained model doesn't know, and so picks a random answer from all possible dates. That's hallucination. The finetuned model takes that uncertainty, but expresses it as the answer "I don't know". Sometimes the finetuning fails and we get hallucination in chatbots. This is because they are overtrained to be certain.