AI Safety Subprojects

Stuart_Armstrong
20 Sep 2021 12:18 UTC

There are the AI safety subprojects designed for elucidating “model splintering” and “learning the preferences of irrational agents”.

Im­mo­bile AI makes a move: anti-wire­head­ing, on­tol­ogy change, and model splintering

Stuart_Armstrong17 Sep 2021 15:24 UTC
32 points
3 comments2 min readLW link

AI, learn to be con­ser­va­tive, then learn to be less so: re­duc­ing side-effects, learn­ing pre­served fea­tures, and go­ing be­yond conservatism

Stuart_Armstrong20 Sep 2021 11:56 UTC
14 points
4 comments3 min readLW link

AI learns be­trayal and how to avoid it

Stuart_Armstrong30 Sep 2021 9:39 UTC
30 points
4 comments2 min readLW link

Force neu­ral nets to use mod­els, then de­tect these

Stuart_Armstrong5 Oct 2021 11:31 UTC
17 points
8 comments2 min readLW link

Prefer­ences from (real and hy­po­thet­i­cal) psy­chol­ogy papers

Stuart_Armstrong6 Oct 2021 9:06 UTC
15 points
0 comments2 min readLW link

Find­ing the mul­ti­ple ground truths of CoinRun and image classification

Stuart_Armstrong8 Dec 2021 18:13 UTC
15 points
4 comments2 min readLW link