carado.moe

tammy's blog about
AI alignment, utopia, anthropics, and more;

also on twitter, lesswrong, rss, and at

QACI blobs and interval illustrated

this post attempts to explain QACI and notably the question and answer blobs and the question-answer interval.

let's say we view a particular way the world could be as a giant causal DAGs. in this world, a user is randomly generating a large (perhaps 1GB) file of random text called the question, and is supposed to, after some time, produce another large file called the answer.

i'm going to color:

• the question and answer as "blobs" of data, serving as "pointers" or "coordinates" into the world. here i'll say that these are sets of nodes through which flow values — typically booleans, i think?
• there's gonna be a rough location of the user, getting the question and outputting the answer. it's very hard to point to the user, but instead we hope to:
• isolate the question-answer interval, as the intersection between the future-lightcone of the question and the past-lightcone of the answer. (this probly counts as a kind of markov blanket)

(the real DAG of the universe since the big bang is obviously astronomically larger, this is a simplified sketch to illustrate what's going on)

the AI will be given a hypothetical distribution over hypotheses for what the world is like, each in the form of a never-halting input tape to a universal turing machine — something akin to solomonoff induction, or alternatively the "spec" or "world" part of a UDASSA hypothesis.

in the image above, the turing machine running over time is illustrated as a sequence of step transitions, some of which correspond to the turing machine writing the question and then later the answer onto its tape — again, the real sequence of steps involved in the history since the start of the universe and the question-answer interval would be astronomically larger in an actual hypothesis.

just like in the causal DAG of the universe, the question and answer blobs — as well as the user of course — are distributed over large amounts of tiny parts. in the turing machine, they're encoded — possibly in a highly complex way! — over large amounts of steps, at each of which the turing machine is writing a bit to its tape.

the problem of blob location (1, 2) is that of finding a way to identify the intended first instance of the question and answer as they are represented on a turing machine input tape serving as hypothesis for our universe.

the question-answer interval is then used as a standalone function whose input are counterfactual questions and which outputs counterfactual answers, and a hypothetical simulation of a collection of these calling passing information to each other is used to build a long-reflection process from within which the user is to solve alignment or, at least and more likely, come up with a better long-reflection process to eventually solve alignment.