Learning through human feedback

We believe that Artificial Intelligence will be one of the most important and widely beneficial scientific advances ever made, helping humanity tackle some of its greatest challenges, from climate change to delivering advanced healthcare. But for AI to deliver on this promise, we know that the technology must be built in a responsible manner and that we must consider all potential challenges and risks. That is why DeepMind co-founded initiatives like the Partnership on AI to Benefit People and Society and why we have a team dedicated to technical AI Safety. Research in this field needs to be open and collaborative to ensure that best practices are adopted as widely as possible, which is why we are also collaborating with OpenAI on research in technical AI Safety. One of the central questions in this field is how we allow humans to tell a system what we want it to do and - importantly - what we dont want it to do. This is increasingly important as the problems we tackle with machine learning grow more complex and are applied in the real world.The first results from our collaboration demonstrate one method to address this, by allowing humans with no technical experience to teach a reinforcement learning (RL) system - an AI that learns by trial and error - a complex goal. This removes the need for the human to specify a goal for the algorithm in advance.