Okay first, they don't lie they're predictive text programs. LLMs give factual errors because they're designed to give the most optimal set of words given a query. Not the most correct answer. They aren't designed to be the best answer giver, they give the "most likely" answer given a set of data. - ThreadSky

lazyblueberry.bsky.social • 82 days ago

Okay first, they don't lie they're predictive text programs. LLMs give factual errors because they're designed to give the most optimal set of words given a query. Not the most correct answer. They aren't designed to be the best answer giver, they give the "most likely" answer given a set of data.

Comments

lazyblueberry.bsky.social•82 days ago

These programs have always given erroneous facts. It's a byproduct of the actual design of the programs. They are, when boiled down, supped up versions of the kind of program that your phone keyboard uses to predict the next word you're likely looking to type next.

lacrunk.bsky.social•82 days ago

going to interrupt you here and remind you that you made a strong claim and I’ve asked for strong evidence to support it

I can’t engage on these secondary points if you won’t engage on the primary one

lazyblueberry.bsky.social•82 days ago

I would also like to point out, even without synthetic data, these LLMs are already trained on "lies". The dataset is the entirety of the Internet, it's full of lies.

It's also full of jokes which a predictive text generator can't discern and is how you get the glue pizza answers

lazyblueberry.bsky.social•82 days ago

https://research.ibm.com/blog/LLM-generated-data

lacrunk.bsky.social•82 days ago

ah, so given these sources, I think you think I'm claiming "synthetic training data isn't being used to train LLMs." That is not the claim I'm making, but I can see how you might have thought that.

lacrunk.bsky.social•82 days ago

I'm claiming "1) there is no evidence (yet) that synthetic data (not foibles of the attention mechanism) is the source of false information given by LLMs, i.e. that they spout misinformation that both a) is high-bias, not high-variance, and b) can be traced directly back to synthetic data."

lacrunk.bsky.social•82 days ago

This is a pretty high bar and requires sources that are either bleeding-edge mechanistic interpretability papers or are straight up proprietary secrets, as AI companies don't release their training data sets.

lazyblueberry.bsky.social•82 days ago

https://www.thestack.technology/llama33-release/

lazyblueberry.bsky.social•82 days ago

https://lifehacker.com/tech/ai-is-running-out-of-internet

Comments

Posting Rules

Reply