Yann Lecun's new paper: the road to building automatic agents

How can machines learn as efficiently as humans and animals ？ How machines learn to reason and plan ？ How machines learn the representation of perceptions and action plans at multiple levels of abstraction , Enable them to reason over multiple time frames 、 Forecast and plan ？ This position paper proposes the architecture and training paradigm for building autonomous agents . It combines things like configurable predictive world models 、 Concepts such as behavior driven by intrinsic motivation and hierarchical joint embedded architecture through self supervised learning and training .