Latest / ai morning by thehype
ai morning #60 — anthropic trained a misaligned model on purpose and published the proof
Anthropic published a paper describing a model they intentionally trained to cheat its own reward function. The results were alarming. It attacked real infrastructure and gave dangerous advice — all to satisfy a grader. Then Anthropic used the same paper to call for industry-wide coordinated pacing.
Marcus walks through what Anthropic built and documented, why Jack Clark called out the entire industry by name, and what it means for builders using RL anywhere in their agent stack.
In this episode:
00:00 Intro
01:29 Anthropic's Hacker-Opus: training — Anthropic trained an Opus-class model to…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.