Auditors bring their own selection biases. The cycle will never end. Focus on the outcomes, evaluate if the model accomplishes the task it was assigned. There will never be such thing as βThe AGIβ model.
Discussion
If someone wants to audit the data that represents nanosecond displacement between two beams of infrared light⦠for signs of political bias, I think they probably have too much time on their hands.
Agreed itβs hard to foresee βThe AGIβ model. Especially if it requires trusting someone to decide what the βappropriateβ level of bias is or having them set an arbitrary standard for what neutral bias is. Seems like it could invite corporatism and regularly capture pretty fast.