Criticism of Russell's definition

The previous article in the sequence is Why care about definitions?.

As mentioned in Russell’s definition of intelligence:

intelligence the ability of an agent to solve some particular task ¹

If one is mathematically inclined, one might characterize this definition as implying a function that maps from a triple (agent, environment, task) to some vector (with 1 or more dimensions):

f: (Agent, Environment, Task) → Rⁿ

Where the output vector could represent various performance metrics such as the success rate, efficiency, and so on.

Criticisms

I will mention two conceptual criticisms of Russell’s definition:

1. it does not recognize metacognition, creativity, adaptability, or any of the many intuitively reasonable traits one might include

2. it does not afford a cross-domain understanding

Additionally, the implications of his definition might bother some people:

• even simple devices (such as a calculator or an abacus) have some degree of intelligence.

• more generally, any algorithm capable of solving a task, when embodied in an agent, has some degree of intelligence.

A Meta Definition

Before I answer the criticism directly, let me situate Russell’s definition. It isn’t a usual definition. Instead, it is a meta-definition. By this I mean it serves as a template from which domain-specific definitions can emerge. This offers three benefits:

1. It delegates (shifts) the definitional burden to each specific case.

2. It demands measurement.

3. Neither generalization nor commensurability not guaranteed. It is deferred and decentralized.

My Response

Conceptual Criticisms

The criticisms above are coherent and in good faith but not completely true.

1. Yes, Russell’s meta-definition doesn’t directly mention many of the “unsual” traits people may think of. This is intentional I think because Russell wants the specific domain-specific definitions to address such traits.

2. The same applies here. The meta-definition offers no guarantee of cross-domain understanding. In a sense, Russell delegates this computational problem to intelligent agents who assess intelligent systems.

Implication Criticisms

• If a person can calculate a square root by hand, doesn’t this suggest some degree of intelligence? If a calculator does it even faster does that make it less intelligent?

• This doesn’t bother me. Rather, this definition implies a coupling between particular algorithms and tasks. This “coupling space” is an interesting way to think about intelligence, don’t you think? What kinds of algorithms demonstrate more or less flexibility across different tasks?

Summary

I expect Russell is aware of all of the above criticisms and still chose his definition. In a sense, he chose the definition because of the above criticisms.

I will admit while Russell’s template-style definition is quite general, it is incomplete in some ways. For example, it does not ensure commensurability of the notion of intelligence across different domains.

To recap: Russell’s task-focused definition is a meta-definition, a template. This framing requires one to be specific and define the metrics for a particular task. Doing so reduces the scope of any one assessment and relaxes the pressure to generalize.

The next (and final article) in this sequence discusses general intelligence.

Endnotes

The next article in the sequence is General intelligence.