Despite impressive advancements in AI, we are still a long way from systems that can perform complex activities autonomously and robustly in real-world domains. Outside tightly controlled environments, we are not yet able to build systems that exhibit the flexibility, adaptability, and awareness required to perform the types of tasks which we would prefer AI systems to take on rather than humans.
Also, the currently dominant paradigm of training systems from scratch, or with limited transfer, on each new problem – with data and compute requirements increasing dramatically as we attempt to tackle ever more complex problems – is becoming unsustainable. Without combining and reusing components flexibly “on the fly”, we will likely soon hit a hard barrier both in terms of the capabilities of AI systems and their sustainability.
Our project aims to test a number of specific approaches towards creating more modular, expandable, and reconfigurable AI as part of a broader vision we have articulated. You can find out more about the specific approach we are taking in our work here.