nostalgebraist

The "...they teach you" formulations are indicative of the toolbox approach, of course. If he would explain how "they" justify their teaching and he'd prove (instead of insinuate) why "they" are wrong he might become rationalist.

I was reading to the end and asked myself when they would present the better alternative to Bayesian inference. And then: least-squares methods with various regulations? Really?

No, his preferred alternative is to just wing it. You have this grab bag of tools, and you can make new ones, but there can supposedly be no systematic way to pick them. Which is why @meaningness cites him in support of his “metarational nebulosity”.

This tweet is too small to hold my reply, but we need to separately look at rationality, cognitive modelling, applied ML, AGI and maybe also philosophy, because "Bayesianism" itself is a bit nebulous.

Yes, the discussion comes down to whether rationalism is viable, and whether minds are universal function approximators, and whether universal function approximators exist. Are there mathematically optimal ways to get to truth when you are in one of the possible universes?

That's the philosophy/meta-rationality thing. If I had the answer, I could build an AGI.

Why? An existence proof does not mean that we know how to get there, and conversely, the absence of fully universal function approximation does not prevent us from building minds like ours?

Going by what you just said, I don't think you, I and nostalgebraist *actually* disagree. Bayesianism can be true, any you still could build a mind without explicitly writing anywhere P(A|B)=P(B|A)*P(A)/P(B)