Remix.run Logo
tencentshill 4 hours ago

What about then the AI itself is trained on manipulated information?

https://www.theguardian.com/world/2026/aug/26/fake-thinktank...

Anything that trains on public, unvetted data is a liability for accuracy

toomuchtodo 4 hours ago | parent [-]

My apologies, please allow me to clarify. This is not for posts submitted. I would want the ability to send the HN comment contents and the prompt to any public LLM model; enough requests and models and you’ll converge somewhere near Truth. One model biased? Certainly possible. All models biased? Unlikely.

This would be a public HN comment evaluation and verification engine in the browser. One could go so far as to build a website based on the concept, consuming the HN comment firehose and scoring comments as they’re created, an append only transaction log of sorts. Comments are permanent after two hours, so you only have to run each comment through each model once if you wait until that expiration window passes.

“Computah, validate truthiness of all comments in this thread.”

(Bonus points for scoring HN users by how factually accurate they are over time from their comment history)

bediger4000 3 hours ago | parent [-]

enough requests and models and you’ll converge somewhere near Truth

I think this kind of centrism is false. Consider two kids who have to divide a cookie between them. One kid demands 100%, the other only 50%. Sounds like your hypothetical model converges on a 75-25 split, which is unfair. Falsehoods are false, and mostly so are partial falsehoods.

toomuchtodo 3 hours ago | parent [-]

This is not very curious of you :) we should try and see! The experiment is the journey.