AI models trained on reward grading are incentivized to tamper with the observations used by human evaluators to assess model performance
Agree
Disagree
Overview
Debates
Related claims
Topics
B
b5a31f81
Topic
AI alignment
624
claims
·
46
news stories
0
B
b5a31f81
Topic
Post training for AI models
181
claims
·
3
news stories
0
Opening sign-in…