[ericmcer]

Seriously, the google generative AI actively suggests completely inaccurate things. It has no ability to say: "I don't know", which seems like a huge failing.

I just asked "what does the JS ** operator do" and it made up an answer about it being a bitwise XOR. 1 ** 2 === 3. The fact that all these LLMs will confidently suggest wrong information makes me feel like LLM is going to be a difficult path to AGI. It will be a big problem if an AI receptionist just confidently spews misinformation and is unable to tell customers they are wrong.

[incrudible]

> It has no ability to say: "I don't know"

So do many humans. The expression of ignorance and self-doubt must certainly be woefully underrepresented in training data.

[galangalalgol]

Yeah, no one posts to say they don't know the answer. It is the smallest of the problems that come from using the internet to train. I realize these are just statistical text generators, but if we do end up training a real AGI on the internet I find that both apalling and terrifying. If I said my parenting strategy was to lock my genius newborn in a room with provided food and water and a web browser, you'd call me insane and expect my child to be a sociopath...