ShowMeTheProbability's Shortform

post by ShowMeTheProbability · 2022-08-15T09:52:39.648Z · LW · GW · 3 comments

3 comments

Comments sorted by top scores.

comment by ShowMeTheProbability · 2022-08-15T09:52:39.967Z · LW(p) · GW(p)

The lack of falsification criteria for AGI (unresearched rant)

Situation: Lots if people are talking about AGI, and AGI safety but nobody can point to one. This is a Serious Problem, and a sign that you are confused.

Problem:

  • Currently proposed AGI tests are ad-hoc nonsense (https://intelligence.org/2013/08/11/what-is-agi/)
  • Historically when these tests are passed the goalposts are shifted (Turning test was passed by fooling humans, which is incredibly subjective and relatively easy).

Solution:

  • A robust and scalable test of abstract cognitive ability.
  • A test that could be passed by a friendly AI in such a way as to communicate co-operative intent, without all the humans freaking out.

Would anyone be interested in such a test so that we can detect the subject of our study?

Replies from: Raemon
↑ comment by Raemon · 2022-08-15T20:13:47.241Z · LW(p) · GW(p)

Becoming capable of building such a test is essentially the entire field of AI alignment. (yes, we don't have the ability to build such a test and that's bad, but the difficulty lives in the territory. MIRI's previously stated goal were specifically to become less confused)

Replies from: ShowMeTheProbability
↑ comment by ShowMeTheProbability · 2022-08-15T22:30:59.790Z · LW(p) · GW(p)

Thanks for the feedback!

I'll see if my random idea can be formalised in such a way to constitute a (hard) test of cognition which is satisfying to humans.