3 ms·
Do we have proof that humans are different in this regard? When you give ChatGPT a story in which you mention that some fictional character is the composer of
by mg 3y ago
Do we have proof that humans are different in this regard?
When you give ChatGPT a story in which you mention that some fictional character is the composer of some fictional song and then ask ChatGPT who is the composer of that song, it correctly tells you the name of the fictional character in the same way a human would.
I would think there is "conscious learning" and "unconscious learning". The training phase of an LLM being the unconsious learning here. Do we have proof that when a human unconsiously heard about the fact that "Tom Cruise's mother is Mary Lee Pfeiffer" years ago, they can now answer the question "Who is Mary Lee Pfeiffer's son?" as well as they can answer "Who is Tom Cruise's mother?".
- sebzim4500 3y agoI think humans are obviously different. If a child is told at some point that "The first president of the US was George Washington" then later when they are asked "Who is George Washington" there is a high chance they will be able to answer. LLMs do not seem to have this ability. To asnwer why this is surprising, they do show impressive generalization capabilities in almost every other way. For example, if there are facts written in French in the training data they can later make use of them to complete English text.
- Grimblewald 3y agoYou must not hang around kids a whole lot. If they aren't directly interested in the I formation they need to hear it many times before they passively acquire it. In that way Llms are the same. If Llms have 'interests' for which they'll acquire information far more readily? I don't know and I don't think anyone else does either but it wouldn't surprisee if it was demonstrable (but not really showing this phenomena) since adding information that is a part of a well spanned domain would be easier than adding information for a novel domain.