AI learns how to set goals for itself.
What if one could design AI that did not just learn how to perform a task but also learned what the task was in the first place – in other words, its goal? A combined research effort by OpenAI and DeepMind has built a system doing just that. At this stage, the design is still at a simple level. There are two reasons this is of interest. One reason is, as they say, that this could and should lead to better goal-setting. It might help avoid, for example, the paperclip maximizer problem. The other reason is that human-level AI (artificial general intelligence) will need to be able to set its own goals.
Link to website: https://blog.openai.com/deep-reinforcement-learning-from-human-preferences/
Link to accompanying paper: https://arxiv.org/pdf/1706.03741.pdf
