Mathematics Colloquium: James A Yorke

James A Yorke, UMD

Artificial Intelligence: Thinking, fast and slow

I compare the ways humans and artificial intelligencesthink, building on Daniel Kahneman’s Systems 1 and 2 (Nobel Prize in Economics,2002).

System 1 is fast subconscious thought, like recognizing a face, walking, or performing everyday tasks.

System 2 is slow conscious analysis, like figuring out how Icross a mud puddle without soaking my shoes or searching through an audience tosee if I know anyone. System 2 uses System 1 repeatedly. Talking and writing useSystem 2.

I add a third component that I call System 3 — learning, improvingboth intuition (System 1) and analysis (System 2).  All 3 Systems have explicit analogues in AIs.

There is data about how chess players think. We can separate the roles of System 1 and 2 by comparingfast chess with slow chess. Fast chess allows 3 seconds per move which isessentially pure System 1 thought. Slow chess allows about 3 minutes per move, sixtytimes slower.

Based on international chess ratings, I find that the bestplayers at fast chess are the best at slow. Superb slow System 2 thinking isachieved only by chaining together superb intuitive System 1 insights.

I give examples from mathematics and physics requiringSystem 1 intuition, including some from my new paper ``Tactics in Proofs'',written with Boris Hasselblatt.

I think AI needs the most improvement in its training System2 for carrying out complex tasks, while humans may benefit greatly fromimprovement in the training of System 1, intuition.