7 ms·
Yet LLMS fail to make these simple but sometimes meaningful differentiation. See for example this case in which a court reporter is described as being all the t
by atoav 2y ago
Yet LLMS fail to make these simple but sometimes meaningful differentiation. See for example this case in which a court reporter is described as being all the things he reported about by Copilot: a child molester, a psychatric escapee, a widow cheat. Presumably because his name was in a lot of articles about said things and LLMS simply associate his name with the crimes without making the connection that he could in fact be simply the messenger and not the criminal. If LLMS had the semantic understanding that the name on top/bottom of a news article is the author, it would not have made that mistake.
https://www.heise.de/en/news/Copilot-turns-a-court-reporter-into-a-child-molester-9840612.html https://www.heise.de/en/news/Copilot-turns-a-court-reporter-...
- npunt 2y agoAbsolutely! Today's LLMs can sometimes(/often?) enormously suck and should not be relied upon for critical information. There's a long way to go to make them better, and I'm happy that a lot of people are working on that. Finding meaning in a sea of information is a highly imperfect enterprise regardless of the tech we use. My point though was that the core problem we should be trying to solve is overcoming the fundamental misalignment of incentives between publisher and reader, not whether we can put a better schema together that we hope people adopt intelligently & non-adversarially, because we know that won't happen in practice. I liked what the author wrote but they also didn't really consider this perspective and as such I think they haven't hit upon a fundamental understanding of the problem.
- mandmandam 2y agoHumans do something very similar, fwiw. It's called spontaneous trait association: https://www.sciencedirect.com/science/article/abs/pii/S0022103101914866 https://www.sciencedirect.com/science/article/abs/pii/S00221...
- thuuuomas 2y ago> fwiw What do you think this sort of observation is worth?
- mandmandam 2y agoReally depends on what sort of person you are I guess. Some people appreciate being shown fascinating aspects of human nature. Some people don't, and I wonder why they're on a forum dedicated to curiosity and discussion. And then, some people get weirdly aggressive if they're shown something that doesn't quite fit in their worldview. This topic in particular seems to draw those out, and it's fascinating to me. Myself, I thought it was great to learn about spontaneous trait association, because it explains so much weird human behavior. The fact that LLMs do something so similar is, at the very least, an interesting parallel.