Rebooting AI: Building Artificial Intelligence We Can Trust

In classical AI, researchers would typically encode by hand the knowledge the AI would need to carry out a particular task, and then write computer programs that leveraged that knowledge, applying it to various cognitive challenges,

Any theory that proposes to reduce intelligence down to a single principle—or a single “master algorithm”—is bound to be barking up the wrong tree.

AI could read and reason as well as humans—yet work with the precision and patience and massive computational resources of modern computer systems—science and technology might accelerate rapidly, with huge implications for medicine and the environment and more.

A big part of why deep learning struggles in language, for example, is that in language, new sentences with fresh meanings are in infinite supply, each one subtly different from the last.

neural networks can’t give human-style explanations for their answers, correct or otherwise.*7

COGNITION MAKES EXTENSIVE USE OF INTERNAL REPRESENTATIONS.

their answers wouldn’t just consist of spitting back underlined passages, but of synthesizing information,

what you do when you read any text is to build up a cognitive model of the meaning of what the text is saying.

Reward is a part of the system, but it’s not the system in itself.