Hacker News Reader

indeed. this will spur research on how to distinguish BS from legit content. which is the fundamental hallucination problem in llms.

and all of us will benefit from this.

View on HN

You can't programatically detect novel BS any more than you can programatically detect viruses or spam. You can only add the fingerprints of known badness into an ever-growing database. Viruses and spam are antagonistic to well-resourced institutions, and their databases get maintained reasonably well. LLM slop is being generated by those same well-resourced institutions. I don't think it fits into the same category as Nepenthes.

the extent to which a random string is a "theorem" of some system can be measured i am guessing.

ie, given some text, to what extent is it "grounded" in some sense in facts.

Reply on news.ycombinator.com