Re:Stack Overflow questions are being flooded with answers f...
by wizeman on Sunday December 04, 2022 @06:06PM
(#33859022)
> AI even at its best won’t be a part of that consensus unless it’s also working in the field.
A single AI can theoretically have the sum total of knowledge/experience that all humans have ever written, be it on a book, research paper or more generally, on the internet.
It can even be trained/learn from audio and video streams (e.g. all podcasts and conference recordings ever uploaded on the internet, etc).
It could also learn from interacting with humans and other AIs.
> As a disclaimer, sure AI which is indistinguishable from humans would be a part of the community
Yes, that's kind of the point I was trying to make :) Or perhaps, what if it was even better than humans at discussing topics on HN?
> nor is it the trajectory of current AI. What we currently have are better and better parrots, and task specific AI.
I think this is 1) moving the goal posts and also 2) ignoring recent research results, which show that the same language model (without any additional training) can be used to solve completely different tasks (such as solving problems that require real-world physical interaction/manipulation) as long as you encode those tasks as text input/output tokens, which means that besides all the different tasks that these models can already do successfully, they also seem to generalize well to other tasks.
A single AI can theoretically have the sum total of knowledge/experience that all humans have ever written, be it on a book, research paper or more generally, on the internet.
It can even be trained/learn from audio and video streams (e.g. all podcasts and conference recordings ever uploaded on the internet, etc).
It could also learn from interacting with humans and other AIs.
> As a disclaimer, sure AI which is indistinguishable from humans would be a part of the community
Yes, that's kind of the point I was trying to make :) Or perhaps, what if it was even better than humans at discussing topics on HN?
> nor is it the trajectory of current AI. What we currently have are better and better parrots, and task specific AI.
I think this is 1) moving the goal posts and also 2) ignoring recent research results, which show that the same language model (without any additional training) can be used to solve completely different tasks (such as solving problems that require real-world physical interaction/manipulation) as long as you encode those tasks as text input/output tokens, which means that besides all the different tasks that these models can already do successfully, they also seem to generalize well to other tasks.