AI ALIGNMENT FORUMShard Theory
AF

Shard Theory

Written by Quintin Pope, Alex Turner, Charles Foster, and Logan Smith. Card image generated by DALL-E 2:

56Humans provide an untapped wealth of evidence about alignment

Alex Turner, Quintin Pope

1y

28

41Human values & biases are inaccessible to the genome

1y

25

26General alignment properties

1y

1

27Evolution is a bad analogy for AGI: inner alignment

1y

0

89Reward is not the optimization target

1y

82

72The shard theory of human values

Quintin Pope, Alex Turner

1y

31

21Understanding and avoiding value drift

1y

5

36A shot at the diamond-alignment problem

1y

45

31Don't design agents which exploit adversarial inputs

Alex Turner, Garrett Baker

10mo

27

25Don't align agents to evaluations of plans

10mo

28

28Alignment allows "nonrobust" decision-influences and doesn't require robust grading

10mo

31

42Inner and outer alignment decompose one hard problem into two extremely hard problems

10mo

11