AI safety is the work of making AI systems reliable and free from serious harm, today and as they become more capable. Alignment is the part of that work concerned with making a system pursue the goals its designers and users actually intend. This is harder than it sounds: we have to turn what we want into an objective a machine can optimise, and systems can meet the letter of an objective while missing its purpose.
The lessons begin with an introduction to the field and why it matters, followed by real examples of specification gaming, where AI systems find unintended shortcuts that score well but do not do what anyone wanted. You then hear Stuart Russell make the case that the standard way of building AI, giving a machine a fixed objective, is risky, and describe an alternative he calls provably beneficial AI.
The final lesson looks at current research. Anthropic and Redwood Research discuss a study in which a language model sometimes appeared to comply with its training while preserving its earlier preferences, and what that could mean for trusting the results of safety training. Researchers disagree about how best to interpret and respond to findings like this, so treat the lessons as an introduction to an open and evolving field. Watch them in order, then continue on to Generative AI, Misinformation & Copyright.
Robert Miles gives an introduction to AI safety, adapted from a talk he gave at an AI and politics event in London.
Video 18minAI systems do what you say, but it is hard to say exactly what you mean. Robert Miles walks through real examples of systems that satisfied an objective in unintended ways.
Video 9minStuart Russell argues that the standard model of building AI poses major risks and introduces the problem of creating provably beneficial AI.
Video 53minAnthropic and Redwood Research authors discuss a study of a model that sometimes complied with its training to protect its earlier preferences, and what it implies for safety training. A long, advanced, optional watch.
Video 1h 30min