LESSWRONG
LW

Shard Theory

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

210Humans provide an untapped wealth of evidence about alignment

TurnTrout, Quintin Pope

2y

94

94Human values & biases are inaccessible to the genome

2y

54

50General alignment properties

2y

2

78Evolution is a bad analogy for AGI: inner alignment

2y

15

375Reward is not the optimization target

2y

123

248The shard theory of human values

Quintin Pope, TurnTrout

2y

67

48Understanding and avoiding value drift

2y

11

95A shot at the diamond-alignment problem

2y

59

70Don't design agents which exploit adversarial inputs

TurnTrout, Garrett Baker

2y

64

45Don't align agents to evaluations of plans

2y

49

60Alignment allows "nonrobust" decision-influences and doesn't require robust grading

2y

42

142Inner and outer alignment decompose one hard problem into two extremely hard problems

2y

22