Okay, sorry, here is my real response since I thought you were talking about something else due to being in two conversations at once in the same thread:
My opinion will change when AIs stop being untrustworthy. Until I can have any sort of certainty that it isn’t just making shit up, it won’t change.
Not too long ago, I asked ChatGPT to tell me who I am. I have a unique name. I also have a long-established internet media presence under that name. I’m not famous, but I’ve got enough prominence for it to know exactly who I am.
It had no idea whatsoever. It got it entirely wrong. It said I was a business entrepreneur who gave motivational lectures.
I’m sorry to break this to you - but you probably weren’t in the training dataset enough for the model to learn of your online presence - yes llms will currently hallucinate when they don’t have enough data points (until they learn their own limitations) - but that’s not a fundamentally unsolvable problem (not even top 10 i’d say)
there already are models that consider their knowledge and just apologize if they can’t answer instead of making shit up. (eg claude)
Considering these LLMs are being integrated with search engines in a way that might work toward replacing them, don’t you think their training should include knowing who someone is that a bunch of hits come up for when you Google them?
Okay, sorry, here is my real response since I thought you were talking about something else due to being in two conversations at once in the same thread:
My opinion will change when AIs stop being untrustworthy. Until I can have any sort of certainty that it isn’t just making shit up, it won’t change.
Not too long ago, I asked ChatGPT to tell me who I am. I have a unique name. I also have a long-established internet media presence under that name. I’m not famous, but I’ve got enough prominence for it to know exactly who I am.
It had no idea whatsoever. It got it entirely wrong. It said I was a business entrepreneur who gave motivational lectures.
idk bro that sounds like saying search engines aren’t useful cuz you couldn’t google yourself…
Except I can Google myself. Links and photos come up. None of them say I’m a business entrepreneur who gives motivational lectures.
I’m sorry to break this to you - but you probably weren’t in the training dataset enough for the model to learn of your online presence - yes llms will currently hallucinate when they don’t have enough data points (until they learn their own limitations) - but that’s not a fundamentally unsolvable problem (not even top 10 i’d say)
there already are models that consider their knowledge and just apologize if they can’t answer instead of making shit up. (eg claude)
Considering these LLMs are being integrated with search engines in a way that might work toward replacing them, don’t you think their training should include knowing who someone is that a bunch of hits come up for when you Google them?