“Nathan Lambert believes that Anthropic's decision to secretly degrade its model's performance on certain AI research tasks without informing the user is a "misaligned" action.”