How much intelligence does a task actually need?

Imagine you have a restaurant. There is one person with an estimated IQ of 200, whatever that means, who would cost you millions to clean the floor. Then there is another person who is pretty average and would cost you a few hundred dollars.

Maybe the super smart person finds a smarter way to finish faster or make the restaurant cleaner. But at some point the floor is simply clean. There is a limit to how much more value you gain with every extra dollar you pay.

So in most cases, it would make sense to hire the average person. The task itself has a ceiling that many people can reach.

Does the same apply to intellectual work?

I wonder if the same can be said about most intellectual work. For many tasks that will be automated with AI, there may be a ceiling on how useful more intelligence can be.

A frontier model can be much smarter without being more useful for the task. And we seem to be approaching the level needed for many tasks very quickly with cheap models that are widely available.

The question is not only how intelligent a model is. It is also how much intelligence the task can actually use.