Comment by kelseyfrog | Hacker News Reader

Comment by kelseyfrog | Hacker News Reader

Parent Full thread

The cybernetic psychology idea of comparing sensed vs. desired states maps cleanly onto a PID controller: P reacts to the present error, I accumulates past errors, and D anticipates future ones.

It's not just a comparator, it’s how long I’ve been off (I), how fast I’m drifting (D), and how far I am right now (P).

In this framework, emotional regulation looks like control theory. Anxiety isn't just a feeling—it's high D-gain ie: a system overreacting to projected errors.

Depression? Low P (blunted response), high I (burden of unresolved past errors), and broken D (no expected future improvement).

Mania? Cranked P and D, and I disabled.

In addition to personality being setpoints, our perceptions of the past, present, and future might just be PID parameters. What we call "disorders" are oscillations, deadzones, or gain mismatch. But like the article pointed out, it's not really a scientific theory unless it's falsifiable.

But as recent AI advancement hints at, these states are highly manydimensional. And in such spaces our intuitions fall apart. Even simple gradient descent works quite differently when you have millions of almost-cardinal directions to pick at any point, and PID regulation is even more complex.

And even in 3D the plain PID just can't do when the space is discrete or when there are large signal delays relative to response time of the system. We don't say "oh it's got anxiety" lol but we replace it with updated algorithm.

That sounds like a great research question, "Given this model[1], what's the dimensionality of the space?"

1. Or more accurately, a model in this family

This is tempting, because we have a tight grip of control theory. However, there are thousands of things you could measure, and picking one at a time to improve could be catastrophic. Give a severely depressed person a big bump in motivation and nothing else, and it might be their end. We'll need to pick a collection of probably at least 100 measurements of various personality facets, try different therapies and medicines, and see which ones reduce the average error the most for different starting conditions. And do this consistently. Then, maybe we can say things like "given your scenario, treatment X will improve your overall condition the most but may leave these facets suboptimal, and treatment Y will provide a slightly lower overall improvement but does better for these facets you were most concerned about" with actual mathematical confidence.

If we want to be pedantic both go back to cybernetics and the idea of the cybernetic loop.

Reply on news.ycombinator.com