Published: July 9, 2016
http://www.incompleteideas.net/IncIdeas/DefinitionOfIntelligence.html
By Rich Sutton
Via Michael Nielsen
What is intelligence, anyway? Sutton begins with McCarthy’s definition—“Intelligence is the computational part of the ability to achieve goals in the world”—but notes that it leaves us with a problem. What’s a goal? How do we know if something’s trying to achieve it? Is a thermostat trying to achieve a goal?
Sutton proposes using Dennett’s idea here: a system having a goal is a property of its interaction with an observer. It’s contingent, In the Cells of the Eggplant - Chapman-style. A system should be thought of as having a goal when that’s a useful way to understand its behavior (i.e. rather than trying to understand its behavior in terms of, say, mechanism).
Q. McCarthy’s definition of intelligence?
A. The computational part of the ability to achieve goals in the world.
Q. Sutton’s concern with McCarthy’s def’n of intelligence?
A. What does it mean for something to achieve goals in the world? How do we know if something has goals?
Q. How does Dennett propose we handle of the question of whether a system “really” has a goal, rather than just “seeming” to?
A. Having-a-goal is a property of the “stance” that an observer takes relative to the object; it’s in the relationship, not in the object.
Q. When might a thermostat sometimes be best understood in terms of mechanism, and other times be best understood in terms of goals?
A. A resident of a house can best understand their thermostat in terms of having a goal (i.e. to maintain a temperature), but if you’re trying to fix it, or if you’re a technician, you’re better off thinking of it in terms of its mechanisms.
Q. What is a goal-seeking system, according to Rich Sutton?
A. One which is more usefully understood in terms of its outcomes than in terms of its mechanisms.