[https://apxhard.substack.com/p/the-surprising-computational-properties]

If you can only navigate with the utility gradient - if you can only move in directions that feel good, that increase your immediate utility - you will get trapped in local maxima.

This is the fundamental problem in optimization. Gradient descent finds local peaks, not global peaks. If you’re at a point where every immediate direction feels worse, you’ll stay there forever, even if there’s a much higher peak just over the hill.

Addiction is the clearest example. Drugs feel good. Withdrawal hurts. If you can only navigate toward “feeling good,” you stay addicted. The drug is a local maximum. It’s not the best possible state - it’s destroying your life, your relationships, your health. But every path away from it goes through a valley of withdrawal, and you can’t navigate through valleys.

To escape a local maximum, you need the ability to move against the gradient temporarily. You need to be willing to make things worse in the short term to make them better in the long term.

Voluntary suffering is exactly this capability. It’s the ability to navigate against the utility gradient when you judge that doing so will reach a better state.

Not because suffering is good. Not because pain is virtuous. But because the willingness to suffer grants freedom that purely gradient-following agents don’t have.