WHY AI STILL ISN’T GENERAL INTELLIGENCE (YET)

Baffled by its brilliance. Stunned by its stupidity.

AI is intelligent, until it’s not.

It can save you a week and then waste an afternoon.

That’s why it needs harnesses, evals and constant supervision to do what you want reliably over time.

It performs intelligence, but it does not reliably learn from experience.

So for your Monday musing, we’re asking:

What even is intelligence anyway?

There are hundreds of possible definitions. But to name a few key ones:

Prediction? Compression? Abstraction?

Yes.

Rationality? Search? Goal-seeking?

Yes.

Adaptation? Efficient encoding? Embodiment?

Yes.

Learning? Causal reasoning? Social coordination?

Yes. Yes. Yes.

All of these are forms of intelligence.

And your new workmate is extraordinarily good at three or four of them. It is weak, brittle or absent on others.

This is why we call it jagged.

Until AI is good across most of these, it will stay that way.

When it is good across all of them, we can start talking seriously about general intelligence.

The question is whether LLMs and harnesses can get us there.

And if they don’t, what happens to Anthropic and OpenAI if someone else gets there first?

Happy Monday.