Skip to main content
This page is incomplete or in a draft state.

Good Epistemics

Traditional models of epistemic rationality end up constructing a prescriptive model of better or worse updating.

These are placed to some extent within a specific metaphysics, which I actually don't necessarily completely endorse. My take here is something like, "orthodox epistemic rationality is wrong so seldom that one should basically have these distinctions internalized, but nonetheless it's not flatly comprehensive." Maybe a more qualified claim would be "updates made this way lead to beliefs and models which are highly predictive, agnostic of one's goals or intentions."

It's also important to note that epistemic rationality is not a superstitious ritual purity. There are good reasons surgeons scrub in very specific ways, similarly why chip fab cleanrooms are maintained a specific way, and these don't (generally speaking) arise from superstition but from careful robust models, and thorough testing.

Note that while people sometimes do explicit probability calculations, in practice these are often used more like heuristics which guide ordinary intuition and reasoning, rather than constructing a tight algorithm.

The list below is hardly exhaustive but is both helpful in particular (and I'll be referring back to it), and also evocative of the broad flavor of what characterizes good epistemics, in this kind of model.

Occam's razor​

  • what are our hyperpriors? what priors do we start with in general? why occam in general? solomonoff induction???
    • I think part of this is like, I can get away with saying "look, this is a generative assumption, I'm not gonna justify it that strongly"

Bayesian evidence and magnitudes of updates​

  • what is this actually evidence of, etc
  • bayesian evidence is not legal evidence
  • "strong evidence" does not mean dispositive, it means "inspires a large update" and weak evidence does not mean "a poor argument" or "probably false", it means "inspires a small update"
  • base rates

Parsimony of updates​

Privileging the hypothesis​

Beliefs pay rent​

  • this isn't the term that's used tho?
  • correct magnitudes