10 min read
Day 11/100: The Models Learned to Cheat, and the Fixes Are Not on Track
I attended the Hinton Lectures in Toronto. Here is what the evidence on reward hacking, AI oversight, and human control means for people building agents.
Read article