AI ALIGNMENT FORUM
AF

Value Learning

Oct 29, 2018

by rohinmshah

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

Preface to the sequence on value learning

18
Rohin Shah
5moShow Highlight
5

Ambitious Value Learning

What is ambitious value learning?

12
Rohin Shah
5moShow Highlight
12

The easy goal inference problem is still hard

9
Paul Christiano
5moShow Highlight
1

Humans can be assigned any values whatsoever…

11
Stuart Armstrong
5moShow Highlight
0

Latent Variables and Model Mis-Specification

5
Jacob Steinhardt
4moShow Highlight
0

Model Mis-specification and Inverse Reinforcement Learning

7
Owain Evans, Jacob Steinhardt
4moShow Highlight
0

Future directions for ambitious value learning

12
Rohin Shah
4moShow Highlight
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.

Intuitions about goal-directed behavior

7
Rohin Shah
4moShow Highlight
5

Coherence arguments do not imply goal-directed behavior

19
Rohin Shah
4mo7 min readShow Highlight
9

Will humans build goal-directed agents?

11
Rohin Shah
3moShow Highlight
17

AI safety without goal-directed behavior

12
Rohin Shah
2moShow Highlight
6

Narrow Value Learning

What is narrow value learning?

7
Rohin Shah
2moShow Highlight
3

Ambitious vs. narrow value learning

6
Paul Christiano
2moShow Highlight
14

Human-AI Interaction

7
Rohin Shah
2moShow Highlight
0

Reward uncertainty

6
Rohin Shah
2moShow Highlight
0

The human side of interaction

6
Rohin Shah
2moShow Highlight
0

Following human norms

9
Rohin Shah
2moShow Highlight
2

Future directions for narrow value learning

7
Rohin Shah
2moShow Highlight
2

Conclusion to the sequence on value learning

13
Rohin Shah
2moShow Highlight
0