A goal in AI is to measure progress and benchmark AI models. To this end Ai researchers have designed a slew of tests.
Let us examine what these tests actually try to tease apart. For any test, we (the human designers) have an underlying conception of an ability or a concept or skill that we are trying to test for.
This underlying skill or ability is called competence.
The actual score on that test is called performance.
What is intelligence?
Marcus and Hutter define intelligence as the ability or skill to achieve goals in a wide range of environments.
Intelligence can also be task specific. That is a penguin has evolved in an environment. It is intelligence enough to solve tasks such as hunting for fish and swimming in water. It does not have a need to run on land or fly. Hence it is not very good in these environments.
+---------+
| AGENT |
| Human, |
| animal, |
| or AI |
+---------+
|
takes actions
|
v
+----------------+
| ENVIRONMENT |
| Chess, science,|
| conversation, |
| physical world |
+----------------+
|
provides feedback
|
v
+---------+
| GOAL |
| What the |
| agent is |
| trying to|
| achieve |
+---------+
|
v
Measure success
(REWARD)
Current intelligence tests do no test for skills such as those that help you generalize to a new environment or a new task. This is what Francois Chollet called broad generalization.
There is a need for tests that probe for precisely this kind of intelligence. For example, can an agent trained in a 2 dimensional environment generalize to a 3 dimensional environment? Can an agent still navigate an environment if there are changes (such as new obstacles)?
What will competence look like for such as test?
Machines might use completely new concepts that we have not anticipated. In certain cases, we might need a new vocabulary or new worlds (neologisms) to describe these concepts.
Ultimately our quest for understanding how intelligent a system is maybe connected to a need to explain.
Explanations are like models. All of them might be wrong, but some are useful.
Explanations are made, not found (Lombrazo explains in her book [1])
[1] We have some explaining to do, Science, Chaz Firestone, 2026 https://www.science.org/doi/abs/10.1126/science.aem0071