Machine Learning

1 readers

1 users here now

Community Rules:

Be nice. No offensive behavior, insults or attacks: we encourage a diverse community in which members feel safe and have a voice.
Make your post clear and comprehensive: posts that lack insight or effort will be removed. (ex: questions which are easily googled)
Beginner or career related questions go elsewhere. This community is focused in discussion of research and new projects that advance the state-of-the-art.
Limit self-promotion. Comments and posts should be first and foremost about topics of interest to ML observers and practitioners. Limited self-promotion is tolerated, but the sub is not here as merely a source for free advertisement. Such posts will be removed at the discretion of the mods.

founded 10 months ago

MODERATORS

communick@academy.garden

[R] "It's not just memorizing the training data" they said: Scalable Extraction of Training Data from (Production) Language Models (arxiv.org)

submitted 9 months ago by wojcech@alien.top to c/machinelearning@academy.garden

30 comments fedilink hide all child comments

you are viewing a single comment's thread
view the rest of the comments

[–] blimpyway@alien.top 1 points 9 months ago (2 children)

It is not about being able to search for relevant data when prompted with a question.

The amazing thing is they seem to understand the question sufficiently so the answer is both concise and meaningful.

That's what folks downplaying it as "a glorified autocomplete" are missing.

PS and those philosphising it can't actually understand the question are also missing the point: nobody cares as long as its answers are sufficiently correct and meaningful as if it was understanding the question.

It mimics understanding well enough.

[–] vade@alien.top 1 points 9 months ago

Suggest you watch https://www.youtube.com/watch?v=uTXXYi75QCU

[–] squareOfTwo@alien.top 1 points 9 months ago (1 children)

these things don't "understand". Ask it something which is "to much OOD" and you get wrong answers, even when a human would give the correct answer according to the training set.

[–] blimpyway@alien.top 1 points 9 months ago

I said they mimic understanding well enough, that wasn't a claim LLMs actually understand.

Sure training dataset limits apply,

And sure they very likely fail when the question is OOD, but figuring out the question is OOD isn't that hard, so an honest "Sorry, your question is way too OOD" answer (instead of hallucinating) shouldn't bee too difficult to implement.