ABCDEFGHIJKLMNOPQRSTUVWXYZ
1
Problem severity (severe/medium/mild)
2
Existing problems with RLHF because of (currently) non-robust ML systems
3
Benign Failures
4
Mode Collapse
5
You need regularization
6
7
Incentives issues of the RL part of RLHF
8
RL makes the system more goal-directed
9
Instrumental convergence
10
Incentivize deception
11
RL could make thoughts opaque
12
Capabilities externalities
13
14
Problems related to the HF part of RLHF
15
RLHF requires a lot of human feedback
16
Human operators are fallible
17
You are using a proxy, not human feedback directly
18
How to scale human oversight?
19
20
Superficial Outer Alignment
21
Superficially aligned agents
22
Inner misalignment
23
RLHF is not a specification, only a process
24
25
The Strawberry problem
26
Pointer problem
27
Corrigibility
28
29
Unknown properties under generalization
30
Distributional leap
31
- Unknown properties under generalization
32
- Large distributional shift to dangerous domains
33
- Sim to real is hard
34
- High intelligence is a large shift.
35
Sharp left turn
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100