LESSWRONGAlignment Hot Take Advent Calendar
LW

Alignment Hot Take Advent Calendar

38Take 1: We're not going to reverse-engineer the AI.

Charlie Steiner

10mo

4

17Take 2: Building tools to help build FAI is a legitimate strategy, but it's dual-use.

Charlie Steiner

10mo

1

21Take 3: No indescribable heavenworlds.

Charlie Steiner

10mo

12

36Take 4: One problem with natural abstractions is there's too many of them.

Charlie Steiner

10mo

4

30Take 5: Another problem for natural abstractions is laziness.

Charlie Steiner

10mo

4

14Take 6: CAIS is actually Orwellian.

Charlie Steiner

10mo

8

47Take 7: You should talk about "the human's utility function" less.

Charlie Steiner

10mo

22

28Take 8: Queer the inner/outer alignment dichotomy.

Charlie Steiner

10mo

2

33Take 9: No, RLHF/IDA/debate doesn't solve outer alignment.

Charlie Steiner

10mo

14

37Take 10: Fine-tuning with RLHF is aesthetically unsatisfying.

Charlie Steiner

10mo

3

32Take 11: "Aligning language models" should be weirder.

Charlie Steiner

9mo

0

25Take 12: RLHF's use is evidence that orgs will jam RL at real-world problems.

Charlie Steiner

9mo

1

53Take 13: RLHF bad, conditioning good.

Charlie Steiner

9mo

4

15Take 14: Corrigibility isn't that great.

Charlie Steiner

9mo

3