🧵 Untitled Thread
Anonymous at Sun, 21 Apr 2024 07:14:37 UTC No. 16138486
Disclaimer im a normie but why does it seem like several companies have been able to make AI at the same time, was there some sort of technical block that was stopping it until recently?
Anonymous at Sun, 21 Apr 2024 07:18:40 UTC No. 16138487
>>16138486
Open sourced models. Hardware to run them on were already bought for other purposes.
Anonymous at Sun, 21 Apr 2024 21:29:16 UTC No. 16139431
>>16138486
Attention Is All You Need...transformers and some such idk
Anonymous at Sun, 21 Apr 2024 21:35:41 UTC No. 16139444
>>16138486
Either, a single organisation is responsible for modern-day AI applications and simply hands out or leases versions of it to other corporations which can build upon it.
Or numerous companies simply started copying from each other, having realized that some company wants to cash in on AI technology, so they started doing the same. It's probably a combination of both. Basic research on AI technology probably did not allow for a good way of generating images until recently. Once it did, numerous companies started pouring cash into it which resulted in derivative applications.
Anonymous at Mon, 22 Apr 2024 06:45:04 UTC No. 16140059
>>16138486
Massive amounts of cheap fast cloud accessible storage.
Anonymous at Mon, 22 Apr 2024 06:50:19 UTC No. 16140065
Which company has successfully created AI?
Anonymous at Mon, 22 Apr 2024 08:02:13 UTC No. 16140145
>>16138486
>Suddenly
Bitch, chat bots existed sense at least the 70's
Stuff started getting wild in 2015
Member Tay? That was 2016
Then there's bert and gpt-j and gpt-1 and 2
https://youtu.be/XFYckCslt74?si=XIG
This is nearly 10 years of ramp up.
Meta got all those GPU's to build models to predict what reels to shovel down your gullet. Then they released the models just to be a spoiler on Microsoft. I mean, I'm thankful for it but they never would have released them if they had been in the #1 spot and could effectively monetize it.
Finding the local maxima on transformers will be explored quite extensively but LLM's existed before transformers and it is almost centrally not the best possible way to reach AGI. It's still unclear just how far longer training, larger context windows, rag and agents will get us. Probably you though enough agent like behavior at it and you can hammer this shit into a functional AGI.
The brain has 100 Trillion synapses, but you really only need less than an eighth of that for consciousness. And those synapsis run at a dismal 1hz to 200 hz, to run consciousness you probably only need like 20 watts in a flesh computer.
These models are trained on hundreds of lifetimes worth of data. Meanwhile a kid can learn to speak in like 3 years of training data.
There is clearly a better way to architect an AI, humanity will just keep picking away at it. Probably eventually there will be multiple ways AGI is achieved, each with their own strengths and weaknesses, and some that end up obsolete. Personally I like the efforts to model connectomes.
Anonymous at Mon, 22 Apr 2024 12:40:42 UTC No. 16140413
>>16138486
Because the method for doing so is published research.
Anonymous at Mon, 22 Apr 2024 20:46:14 UTC No. 16141095
>>16138486
>was there some sort of technical block that was stopping it
nomenclature
Anonymous at Tue, 23 Apr 2024 14:36:28 UTC No. 16142143
>>16138486
we needed bitcoin to prove that gpus weren't just for faggot gaymers, and SV techbros to translate indecipherable math jargon into practical code