Thought gradients

I have a really interesting analogy/hypothesis for how to restructure your psyche, an analogy with both evolution of the first life and gradient hacking in AI safety. Basically, both of those things have a currency / conserved quantity that they use in order to maintain internal complex systems.

TL;DR

• A living system can create anything it wants, so long as it couples doing so with increasing the entropy of the universe.
• Gradient hackers can become anything they want to as long as they couple doing so to performing better on the training task.
• You can be become anything you want to, as long as you couple the becoming to positive affect.

Origins of life

With life, everything must ultimately increase entropy, however a cell can maintain a low-entropy interior because the existence of the cell causes more intense entropy increase outside the cell (ie. its "an engine for more quickly turning useful states of energy into less useful states of energy"). The first life happened because of structures that emerged that made use of an existing entropy gradient, and the lower entropy of the structures themselves come from skimming off the increases in entropy that those structures cause. Those structures are therefore constrained somewhat, but not completely

A living system can get nearly anything it wants so long as it "buys" it by coupling it to increased entropy of the universe.

Gradient hacking

Gradient hacking is the idea that a neural network trained via gradient descent might be able to introspect on its own training process and direct the weight updates to go the way that it wants. Here, the conserved quantity is the loss, which must always decrease. Neural networks come into being as structures that result from lowering the loss. A neural network in training could create internal structures of it's own design, as long as doing so always caused the loss to decrease.

A gradient hacker can get any thought pattern or internal structure it wants, so long as it "buys" it by coupling it to decreased loss on the training objective.

Your mind

The analogy with training your own psyche is that the conserved quantity is "positive affect". The way your mind works (to a first approximation) is that when you have a mental state X, and then feel bad within the next few seconds, your mind is updated so as to produce less of mental state X. When you have a mental state X and then feel good, your mind is updated so as to produce more of mental state X. So what you have to do to intentionally restructure your mind is to couple some desirable mental state/desirable action/habit with a subsequent positive affect within the next few seconds.

The thing is it can be totally artificial and it will still work. Like, forcing yourself to smile creates a somewhat happy feeling. The hardest part, is that if you are already below baseline, it's hard to reinforce anything. Also, if you are having negative affect within the few seconds after some desirable mental state/action, then you should stop worrying about what you are trying to do immediately, because you're doing the opposite of what you need.

So what this basically means is that the most powerful thing you can do is to cultivate the ability to feel good on command, even if it's totally "meaningless good". From there you can actually get the behaviour you want.