2024-12-23 misc3

if u dont understand smth u need to know that u dont fully understand it and that youve blackboxed it etc

how does a good human reasoner learn that?
and then how do we update our beliefs/understanding after learning it

===

u also need to learn the threshold of blackboxing, and the threshold of how blackboxed of a tool can u still be satisfied using

how does a good human reasoner learn that?

===

id like an ai to be able to go through Purcell n Morin EM textbook, do all the practice problems, then have better EM knowledge than i do .

id like it to be as data-efficient as a human .

===

how do you allow it to gauge like . for us when we feel confident abt smth we just stop practicing that type of problem and we move on to diff thigns (we might be wrong abt it actually. but it is a good approximation and we can change the way we approximate it)

===

id like to be able to interrupt/pause an AI's RL process so that i can teach it just like a human .

that way, i dont have to think worry about having all of the possible cases of "u also need to learn xyz and the threshold for xyz .. how does a human learn that?", i can just notice that it's being bad then i can just teach it -- just like i would teach a salena

wait FUCK but the Constitutional AI RL process is only for a set of static preset principles. you have to go through a whole RL process just for it to learn to be a good boy. what if i want to tell it something once and then basically have it immediately (or near-immediately) learn it for the whole future?

how the FUCK do humans immediately learn smth after 1 (or few) commands and/or practice attempts??
re: my note w "mirror neurons" (misnomer, but that's what i labeled it)
not only do humans append it to their prompt (maybe salenahumans do, but not corbin reasoners)... it goes deeper than that. it's like we immediately generate synthetic data and then practice in our head many times almost immediately. smth like that? not rly. sorta. somehow we're able to understand smth and change our weights (not just change the prompt) immediately without extra RL. like if someone gives us a reasoning tip for some EM topic (i mean salena might just change her prompt, but usually i try to deeply understand the abstract concept so that now it is internalized and i dont need to remember anything cuz it is just now in my weights which will generate my actions. this is not the best analogy but its the best i can articulate rn) -- or perhaps another example is someone tells us that X person is icky, then we immediately change our weights related to interacting with X person, we dont have to go through any extra RL process for that. and it's not just like we've stored away X person in our mental list of icky ppl, and then we have to check our list before interacting... no, we've truly somehow internalized it deeper than just a mental list, we've really somehow changed the instinctual/intuitive weights that sub-consciously drive ur actions.

i think this is a rly interesting question.
is ARC able to be solved by just learning via changing the prompt? (like a salenahuman)
and also, wait, can AI alr do this or nah? i think no. but i could be wrong.... i am very new to this stuff

===

what is the difference between salena reasoners and corbin reasoners?

how come when im told something i try deeply understanding it? when salena is told something she just memorizes it? how exactly do i try to deeply understand it? and why do i deeply understand it? how did i learn that? was it from my dad? what did he tell me to make it that way? before we can even begin to think about teaching AI to do that, how the FUCK did i get taught that and how the fuck would you teach other humans to do that? only after answering that question can we move on to thinking about AI RL implications...

==========

text is its native processing unit / native language

our native language is not even text . it's just a leaky abstraction to convey visual/auditory/sensory/etc info of experiences

just like salena, it has no visual intuition for math

but unlike salena, it has read the whole internet

and also unlike salena, its native language is literally Text. it only processes text as input and it only processes text as output. it has never known anything else. that is all it knows.

so, it is much better and more efficient at understanding text than us. just like a calculator is much better and more efficient at doing numerical calculations compared to us. because for us, numbers are not our native processing unit, they are an abstraction on top of what they actually represent to us, and so we are slower at processing numbers compared to processing what IS native for us. but calculators' native processing unit is numbers. that is all they know and all they will ever know. and they are extremely better n more efficient at processing NUMBERS. i think its a very very similar thing with LLMs and text. i think. i mean, a physics text is based on visual/real-life abstractions and it cant understand the abstractions so it will be just like Salena's understanding of physics (just text-based and surface level pattern matching)... but even tho it understands physics just like Salena's understanding, it will be able to input/output/manipulate the text much better/efficient compared to Salena. it's like imagine if Salena got 999999 IQ but just NEVER chose to apply that IQ to get a deeper level understanding and instead only applied that IQ to scale her current way of doing things. thats kinda what an LLM is.