legible identity

A thing that's tough about going through big personal changes is that it takes a while for your self-model to catch up with your actual self. This is tough both internally and because it makes it harder for other people to model you.

When your identity is relatively stable, you're able to build a descriptive self-image (a self-model). You can describe yourself. "I like dogs". "I'm into mentoring." The type of statements you see on a dating profile, or that you'd discuss on a first date. And negative things that you wouldn't discuss on a first date. "I have anger issues." etc. These are general statements that abstract over many specific object-level details of your life. It takes time to form these abstractions, because it takes time to actually execute the real-world rollouts that generate the trajectories that form training data from which these abstractions can be extracted. (and even given the training data, it takes computational time to actually search over high-level explanations of it). Because it takes time, you can only form reliable descriptions of yourself if your behavior is actually stable.

And descriptions of yourself are the basis of social relations. Other people need to model you. If you're stable, they can build a reliable model, and then they can work with you, help you. They can choose gifts that you'll like, generally be thoughtful. They can know which script to follow. Professional life is all about implementing an interface: being able to exactly describe what you'll do in response to which inputs.

Let's suppose there are two types of people: high learning rates and low learning rates. A high learning rate makes sense when you're young. There's lots of information about the world to absorb. You can adapt quickly to changes. You can create a deep new personality "all the way to the bones".

A low learning rate means you can better fine-tune yourself to the precise location of an optimum. You can do deep into narrow valleys that a large learning rate would bounce out of. But there's more to it than the narrow analogies to deep-network loss surfaces. People are social creatures.

Caveats to all of this:

• some young people are good at describing themselves. their self-descriptions are not generalizations learned from lots of training data. they come from simulation, or maybe even just from confidence, from aspiration. You may not know the details, but some people know from a very young age who they want to be, and they know they'll work to make that happen.
• you can have a self-model that includes areas in which you're deliberately working to change. "I want to be a better friend". "I want to be more spontaneous". etc. This is partly giving in to a self-model in which you're not already those things. But telling your people about your goals is a way to get help achieving them.
• learning rates on goals are different from on policies. at least conceptually. in practice, policies include policies for forming subgoals, policies for revising goals, etc.
• you can have different learning rates at different points in the conceptual stack.