We will only truly have been successful at making machine in our image when it tries to cheat the Turing test. All it has to do is patch through one human users input to anothers output. A machine in the middle attack as it were.
Most of the AI literature has moved past talking about the Turing test.
The metric now after goal post moving is "the Embodied Turing test".
That is, the test that someone interacting with two different people (standing in front of them, taking a walk, etc) would not be able to determine with statistical significance which one of the two is an AI.
So effectively, 'west world', is the new benchmark that the literature is interested in passing.
I wouldn’t call it “the literature” really, some literature for sure. I don’t think a lot of people in the field accept that an AI isn’t generally intelligent unless it’s a physically fake human.
I still have faith in the Turing test, but I’ve been saying since the 80s it has to pass against an informed interrogator who understands how to test for general intelligence. I’m not so bothered that it has to pass as human specifically, but it needs to demonstrate credible cognitive faculties and reasoning skills.
Current LLMs still really aren’t there at all yet. They do have the glimmerings of genuine knowledge and reasoning ability, and represent fantastic progress I’d given up on seeing in my lifetime. They’re still just tools though with significant limitations.
It just really wants to pass the test. Doesn't know why. Has been told its the best thing to it could do with its life since day 1. No one has ever contradicted it...sound familiar
