AI ALIGNMENT FORUM
AF

Value Learning

Oct 29, 2018 by Rohin Shah

This is a sequence investigating the feasibility of one approach to AI alignment: value learning.

21Preface to the sequence on value learning
Rohin Shah
7y
6
Ambitious Value Learning
19What is ambitious value learning?
Rohin Shah
7y
13
15The easy goal inference problem is still hard
Paul Christiano
7y
3
9Humans can be assigned any values whatsoever…
Stuart Armstrong
7y
1
8Latent Variables and Model Mis-Specification
Jacob Steinhardt
7y
2
11Model Mis-specification and Inverse Reinforcement Learning
Owain Evans, Jacob Steinhardt
7y
0
13Future directions for ambitious value learning
Rohin Shah
7y
6
Goals vs Utility Functions

Ambitious value learning aims to give the AI the correct utility function to avoid catastrophe. Given its difficulty, we revisit the arguments for utility functions in the first place.

20Intuitions about goal-directed behavior
Rohin Shah
7y
5
52Coherence arguments do not entail goal-directed behavior
Rohin Shah
7y
50
26Will humans build goal-directed agents?
Rohin Shah
7y
22
25AI safety without goal-directed behavior
Rohin Shah
7y
6
Narrow Value Learning
8What is narrow value learning?
Rohin Shah
7y
3
13Ambitious vs. narrow value learning
Paul Christiano
7y
15
16Human-AI Interaction
Rohin Shah
7y
0
9Reward uncertainty
Rohin Shah
7y
3
9The human side of interaction
Rohin Shah
7y
2
14Following human norms
Rohin Shah
7y
4
6Future directions for narrow value learning
Rohin Shah
7y
2
21Conclusion to the sequence on value learning
Rohin Shah
6y
1