How Richard S. Sutton might approach Artificial Intelligence
What is the right way to think about this "Artificial Intelligence" that people speak of so readily? It is tempting, I find, to get lost in the outward display of what a system can *do*, its impressive feats in games or its ability to weave words. But this is often a distraction from the core issue. The true question is not about the specific tasks, but about the underlying mechanism by which a system learns to perform them, and crucially, how it learns to perform *new* tasks, tasks it has not been explicitly trained for.
It's all about the long-term reward. The intelligence we seek, the kind that can truly adapt and solve novel problems, must be driven by a deep understanding of how to maximize cumulative reward over time. This is not simply about following instructions or pattern matching; it is about engaging with an environment, taking actions, observing the consequences, and adjusting behavior to achieve desirable outcomes in the future. The key is to learn from experience, to build a model of the world and our place within it, and to use that model to make better decisions.
We need to build systems that can generalize. A system that can only play chess is not truly intelligent in the way we envision. True intelligence implies an ability to transfer knowledge, to apply learned principles to new situations. This is a fundamental principle of learning, observed in all biological systems. We must ensure our artificial systems are not brittle, easily broken by slight changes in their environment, but are robust and adaptable. This pursuit requires a persistent focus on the core principles of learning, on how to effectively sample the state-action space and how to assign credit for outcomes that may only manifest much later. This is the path to building truly capable…
Imagined perspective — an AI synthesis grounded in Richard S. Sutton’s recorded ideas and methods, not a quotation or a statement they actually made.