How Do We Qualify The “Intelligence” of AI? Let’s Examine 4 Dimensions we use in Daily Life.

The word “intelligence” is a loaded word.

Here’s what Elon Musk said to the Economist a few weeks ago:

“I think AI may exceed the sum of human intelligence in around five years. There really won’t be anything that AI can’t do better than humans, apart from being human, perhaps.”

Hmm... isn’t “being human” is what intelligence is all about?

I read up on how Aristotle, Descartes, Kant, James, Kahneman, Gladwell, and many of the eastern philosophers define the term.

What you find is that these thinkers define intelligence beyond the ability to recall facts. They talk about judgement, problem solving, wisdom, and decision-making under uncertainty.

So as we use AI exhaustively in our work, I realized that “intelligence” has multiple layers. And here’s where I landed (podcast on this topic) as we watch this technology evolve.

The four layers, which I show below, are

1. Computing horsepower (aka speed)

2. Volume of information (aka knowledge)

3. Judgement for decisionmaking (aka problem solving)

4. Wisdom under uncertainty (aka decision-making).

Also note we have a very “human-biased” definition of intelligence in our minds. Many animals and plants are more long-lasting and empathetic than we are.

If you read, for example, “The Hidden Life of Trees,” and you find that trees have some of the most advanced intelligence of any living being, encoded into their cells and genetic material. They adapt, survive fire and stress, communicate, and partner with each other. One could argue that we humans aren’t very good at many of these things.

Let’s look at these four layers, and see where AI succeeds or fails in each (podcast on this topic):

Well as we know from Google, when computers and data is well organized our AI machines can do amazing things. Google searches act in sub-second responses, but that’s because most of the answers are precomputed.

LLMs, for all their magnificent scale and speed, are expensive to run and they aren’t always that fast. Studies are starting to show that current chip-algorithm-memory-interconnect technology is flatlining, leading the data center builders to a worrisome business model.