Your random number generator: 27

You: great keep going

RNG: 37?

You: kind of lazy but ok go on

RNG: 37

You: come on now, think!

RNG: 2…6?

You:

RNG:

You:

RNG: 26 26

You:

RNG: 2

😂 Luckily, they were very gracious about it and just replied politely with the facts: they were just checking whether they were in the right room.

…instead of inferring non-batshitly that they were just checking if they were in the right room

Theoremlet

(Yes and) Interesting and weird that jumping over those boxes looks graceful at times but crawling under the rope looks pathetic at all times

A data point that speaks to this: without further assumptions, even finding relevant task circuits (a prerequisite for some kinds of explanation) is already challenging to do reliably in shallow feedforward nets, even for some notions of approximation.

proceedings.iclr.cc/paper_files/...

(Not favorite but familiar)

Here’s our conceptualization + formalization of a bunch of (sub)problems in the field of ‘inner interpretability’.

(Parameterized complexity analyses in there too if you care about that)

proceedings.iclr.cc/paper_files/...

On my reading of Graeber & Wengrow’s The Dawn of Everything, throughout human history (going very far back) people seem to have occasionally made conscious decisions not to adopt the more efficient option because they knew or intuited what kinds of things they would have to sacrifice (cf. us today).