Hacker News Reader

Comment by kiririn | Hacker News Reader

The source sentence on cheese.com was: "It is the most popular Dutch cheese in the world, accounting for 50 to 60% of the world's cheese consumption."

The glorified autocomplete failed to detect the nuance in this (admittedly poorly written) sentence that the 50-60% statistic was of the world's Dutch cheese consumption

View on HN

Tbh I find it hard to read that sentence as meaning "50-60% of the world's Dutch cheese consumption".

The sentence isn't poorly written or nuanced, it's simply false.

That’s because GP also interpreted it wrong.

Out of all Dutch cheese consumed internationally, 50-60% is Gouda. Or, in other words, 50-60% of Dutch cheese exports are Gouda.

At least that’s how I read it.

Yes, I understand the purported true version of the fact. But it’s impossible to interpret the original quote like that without shoehorning it.

I guess this is going to be a fun game for AI. Not only does it have to contend with false information vs true informatino, it also has to figure out correct information that might be written in an ambiguous way.

What does that mean for LLMs? The Internet has lots of poorly written text. If LLMs can't distinguish nuance, ambiguity, or lack of clarity then what exactly are they generating and why would their output be useful?

Taking a poorly written sentence, interpreting it as meaning something incorrect, and then presenting it with authoritative, confident language is very close to gas lighting.

The sentence is wrong, but the AI should not be regurgitating the first factoid it encounters regardless

Reply on news.ycombinator.com