"AI, how can you not know my opinions when I post them constantly?!" because there exist trillions of concepts encoded in far less space and you personally aren't important enough to sufficiently to dent the training set. just turn on web search already, this know-nothingism is exhausting - ThreadSky

ed3d.net • 2 days ago

"AI, how can you not know my opinions when I post them constantly?!"

because there exist trillions of concepts encoded in far less space and you personally aren't important enough to sufficiently to dent the training set. just turn on web search already, this know-nothingism is exhausting

Comments

markshore.bsky.social•2 days ago

to a high degree of certainty I've written or said meaningful combinations of words that never existed before I came up with them... and this isn't a brag because many other people on this site right now have done the same

ed3d.net•2 days ago

Like I do think that vendors do a bad job of teaching people how to make these useful but an educated person should be able to build a theory of a system after years of exposure

bencates.bsky.social•2 days ago

You're just doing this but for information theory.

bencates.bsky.social•2 days ago

Which, to be clear, I do think you need at least a little bit of remedial information theory to use LLMs well and that does make them basically impossible to package for the general public in a way that won't invite misuse.

ed3d.net•2 days ago

I don't agree. "Do not trust the training set. Only ask questions about data you put in it or you watched it get from a web search."

I just improved hit rate by 95%, ama

bencates.bsky.social•2 days ago

Still think you might be underestimating how much of your intuition for which questions are and aren't about which topics is informed by some domain specific knowledge most people won't share here.

ed3d.net•2 days ago

so I'm in the process of teaching some coworkers who definitely do not share my set of skills and if you want really transformative workflows, I agree. "do something without hallucinating" is a much lower bar and I do think can be cookbooked

dollspace.gay•2 days ago

This is why doll turns on search and tells it "search for x on y websites for z and compile it in one place"

ed3d.net•2 days ago

I am but I also literally explained it in that person's prior thread complaining about it

ed3d.net•2 days ago

gpt-4 is rumored to have been around 1.7 trillion parameters. there are seven trillion people on earth. you have to outcompete all of them in notability along with the idea of a bench, the concept of comparative advantage, and Sephiroth

I'm sorry that you just can't compete with Sephiroth

connorlynch.bsky.social•2 days ago

huh that’s more people on earth than I thought

ed3d.net•2 days ago

I love how nobody noticed that

Fortunately that was an artisanal autocorrect fuckup and not an AI or I'd look very silly

ed3d.net•2 days ago

I do my OWN stunts

jonmarke.bsky.social•2 days ago

That number is out of date, actually, the true number is closer to 9 trillion now

ed3d.net•2 days ago

(also that rumor might be false for the record, I think maybe they ended up tacitly admitting less. 671B or so. don't trust me for numbers I don't know anything)

philpax.me•2 days ago

think everyone went off NVIDIA announcing that their hardware was used to efficiently train a GPT-MoE-1.8T around the time that GPT-4 was dominant

ed3d.net•2 days ago

I was too busy focusing on my Sephiroth joke to notice autocorrect helpfully changed whatever the hell I typed to "trillion"

AI would fix this, maybe, I don't know, ask it

also it's eight billion now, yikes