Preferences

riskable parent
It can recite something like 80% of Harry Potter with carefully crafted prompts. If you take half a sentence from Harry Potter then tell the LLM to predict the rest it will complete it. That's what they did in that study you're referring to.

It's not even remotely the same thing as "can recite whole Harry Potter." If you ask an LLM to regurgitate Harry Potter it won't be able to do so because that's not how they work. They're prediction engines and it just so happens that Harry Potter quotes/excerpts are so pervasive on the Internet that the LLMs ingress ranks that style of wording higher than other styles.

Ask it to regurgitate some other, less-popular work. Do it for hundreds or thousands of them. You'll quickly find that those two examples you gave are the exceptions and that LLMs can't pull it off. They won't even get close.


This item has no comments currently.