Culturally, I am clearly part of the "Yudkowsky cluster". And as near as I can tell, Bayes is actually the true foundation of epistemology.

But my personal PRACTICE is much closer to the sort of thing the critical rationalists talk about (assuming I'm understanding them).

For instance, the basic high-level meta-pattern that I run all over the place:

0. Have goal
1. Try to accomplish goal
2. Fail to hit goal
3. Noticing what went wrong, model the system in question, and then derive interventions from that model
4. Apply interventions
5. Recurse

(There are some more advanced bells and whistles on top of that loop, like using simulations instead of trying interventions, and identifying fundamental gears over babbling and pruning, and techniques to notice more to get more data from failures. But this is my core process.)

This is fundamentally about "solving problems."

And it is much closer to having a hypothesis and finding ways to falsify it, than it is maintaining a stock of hypotheses and incrementally shifting the relative weightings on each one in response to evidence.

I think that falsificationism is more practical (in some, but not all contexts) for an important reason: you can only take one action at a time.

If you need to decide between doing X and doing Y, then, practically speaking, it doesn't matter if your internal odds ratio (between the hypothesis that favors X and the hypothesis that favors Y) is 1:5 or 1:20. In both cases, you're going to do Y.

Because you can only have one action, the most important thing is "what is my lead hypothesis?" and "how can I improve it?"

(By the way, this is related to part of why humans, in practice, have cruxes, which I'll publish a post about someday.)

The crucial feature that makes a falsificationism-like approach attractive is that the main way that you get more data is by acting on your current best guess.

In cases where that doesn't hold, like, say, econometric time series that you personally can't intervene on and don't have any control over, it makes more sense to reason about multiple hypotheses at once, and incrementally updating with new info.

In cases where you are intervening yourself and that's how you are getting data (like when you're trying to discover a method to sleep soundly), you can have some background hypotheses, but mostly they aren't relevant. Mostly what matters is your current best hypothesis.

(Sometimes you can pursue hypothesis a and get data that informs hypothesis b, but in my practice, at least, that's an edge case.)

Another place where a more Baysian-like approach is practical is when you DON'T have to take just one action: you can take a portfolio of bets.

Relatedly:

A predictive processing agent makes sense of the world as smears of probability all throughout it's perceptual hierarchy _except_ at the very top. The subjective experience of the agent is of a single, non-probabilistic, world.

Because ultimately, the agent has to take a single action. It can't put forth a portfolio of different actions, weighted by probability of their underlying hypothesis. So it doesn't help to be living in a smeared out probabilistic world.

(I suppose that an agent could try to maxipok over the possible worlds that it is living in. I don't know why our actual evolved PP systems don't do this.

Prob because evolution rewards risky action and aiming for "ok" across your probability distribution doesn't lead to IGF?

Or maybe they actually do this in some way that isn't obvious to me.)

Since the agent can only take one action, the thing that matters most is what the highest ranking hypothesis is, not the exact probability weightings of the different possibilities.

In general, when 1) your goal is to act in the world, and 2) actions are discrete and mutually exclusive, practical epistemology looks more like continual criticism and refinement of your current best guess.

@ESYudkowsky @DavidDeutschOxf @reasonisfun @AnnaWSalamon