AI safety

 

Evan Hubinger wrote an excellent paper on #AIsafety. It confirms my fear that we might need to step away from the inductive learning paradigm or add transparency algorithms to reinforcement learning and other classical machine learning algorithms to accomplish the goal of building controllable AI. Terms like outer alignment or inner alignment will need a lot more exposure in the AI community. STEM AI is probably one of the more exciting areas based on training models on purely abstract scientific concepts, thus avoiding unwanted biases. 

#ai #machinelearning #controllableai #artificialintelligence

https://arxiv.org/pdf/2012.07532.pdf