Post

knowing how things work allows you to criticize and fight things way better than ignoring them because you dislike them does

insisting that guns are bad and their users are immoral, so you don't need to know anything about them, doesn't change what happens when someone aims one at you, and keeps you from knowing how to make one safe if it's necessary for you to do so

i think LLMs suck and it doesn't change the fact that someone can ask Claude Fable who I am (with its ability to search the web turned off) and it will tell them. knowing their capabilities allows me to intentionally shape my online presence to not stand out to something inhuman analyzing my posts

you can believe that LLMs are just as incapable as they were in 2023 all you want. that just means the FBI, which uses modern ones that work pretty well for whatever the hell they're doing with them, knows more about them than you do

also insisting they have or don't have characteristics or abilities that any coworker with a chatgpt account can go and disprove you about instantly just makes you look stupid, and reflects badly on everyone else. would you prefer ammo that can trivially be rendered inert, or do you want to win?

yeah bsky.app/profile/reas...

if the AI shit is driving you nuts and you're sick of hearing about it, you aren't required to know how it works. but you should at least try not to spread misinformation about it that's going to leave people who think you know what you're talking about ill-equipped to deal with it

My understanding of "AI" LLMs is that it's just a fancy giant look up table. It can give you an answer because the answer was the most common thing people wanted when searching that series of words, but doesn't have the nuance to understand said words. So if a LOT of people are wrong,

if it were just a big look-up table, it wouldn't be able to solve novel problems, and they can absolutely do that. it's not storing things directly, it's compressing them by finding incomprehensibly high-dimensional patterns.

this leads to things like LLMs being able to do arithmetic without calculation tools because there's a high-dimensional helix inside their activations, upon which they do arithmetic by using trigonometry (not "consciously", this process is not accessible to the LLM other than by its results)i

what's going on inside the things isn't consciousness, and it's not rote memorization, it's something else that's without precedent. it's significant parts of human cognition, except fundamentally different because it's reverse-engineered from cognition's output

there's all kinds of spooky engineering going on with these things; LLMs do voice and image input by using projection layers to rotate the tokens of sound slices and image patches into the same high-dimensional space that text tokens occupy, which automatically grants the thing "understanding" of it

it really is kinda wild