One interesting fact my group discovered is that dropout seems to increase the extent to which a network was modular. We have some results on the topic here but a more comprehensive paper should be out soon.
Related but not identical: this shortform post.
If wordspace is a subset of chunkspace and not the other way around, and you have about five chunks, do you agree that you do not have about five words, but rather more?
Is the claim that before I learn some new thing, each of my working memory slots is just a single word that I already know? Because I'm pretty sure that's not true.
How do coherence theorems translate to such specific claims about the "shape of cognition"; and why is grounding these theorems in "resources" a justifiable choice in this context?
It occurs to me that one plausible answer here is that cognition requires computational resources, and therefore effective cognition will generically involve trading off these resources in a way that does not reliably lose them.
But my more relevant response is that in that section I don't see Eliezer saying that coherence theorems are the justification for his claim about the anti-naturalness of deference.
Do working memory chunks come in order? Like, I'd kind of expect that if you have 5 concepts in working memory, you can't additionally remember the order they should go in, because that's another working memory chunk. Or if you can remember the order they should go in, then introspectively I'd imagine they'd become one working memory chunk.
To me, this suggests a major change to the message of the post. Reading it, I'd think that I have five samples from the bank of existing words, but if the constraint is just that I have five concepts that can eventually be turned into words, that's a much looser constraint!
I ended up writing 3 more reviews. They each took about an hour to write. I'm reasonably happy with them given those constraints - they could be better if I spent a day on each, but I think I hit the right part of the trade-off curve. I'd expect the reviews of Olah's post and the deep double descent post to be pretty useful for non-expert readers, telling them things they didn't already know. I'm a little bit sad that they didn't get somewhat more acclaim, but such is life.
Aren't working memory chunks much bigger than one word each, at least potentially?
An ex post bad decision that probably resulted from a cognitive mistake: not trying to get into a vaccine trial.