what “clicked” for ai/llms in 2022?
| snake coke | 08/25/26 | | average/ordinary/typical citizen/person | 08/25/26 | | Synagogue of Satan as a Service | 08/25/26 | | average/ordinary/typical citizen/person | 08/25/26 | | Metaphysics is Goy | 08/25/26 | | Richard Ames | 08/26/26 | | a New York Times analysis of Reddit posts | 08/26/26 | | The Buzz Aldrin Spacelaw Fellowship | 08/25/26 | | Metaphysics is Goy | 08/25/26 | | The Penis | 08/26/26 | | timothy mcveigh | 08/25/26 | | Synagogue of Satan as a Service | 08/25/26 | | timothy mcveigh | 08/25/26 | | Metaphysics is Goy | 08/25/26 | | agenticman | 08/25/26 | | ,.,..,.,..,.,.,.,..,.,.,,..,..,.,,..,.,,. | 08/25/26 | | snake coke | 08/26/26 | | ,.,..,.,..,.,.,.,..,.,.,,..,..,.,,..,.,,. | 08/26/26 | | snake coke | 08/26/26 | | Founder | 08/25/26 | | Synagogue of Satan as a Service | 08/25/26 | | Oh, you retrieve? | 08/25/26 | | .,.,...,..,.,.,:,,:,.,.,:::,...,:,...:..:.,:.::,. | 08/26/26 | | hotel children | 08/26/26 | | Monikor | 08/26/26 | | Brussels Sprout (Apropos of the Wet Snow) | 08/26/26 | | agenticman | 08/26/26 | | .,.,...,..,.,.,:,,:,.,.,:::,...,:,...:..:.,:.::,. | 08/26/26 | | Jared Baumeister | 08/26/26 |
Poast new message in this thread
Date: August 25th, 2026 3:35 PM Author: The Buzz Aldrin Spacelaw Fellowship
Seminal "Attention is All You Need" transformer paper in 2017.
Nvidia launches V100 same year.
BERT comes out in 2018.
GPT comes out in 2019.
Nvidia Launches A100 in 2020 (3x performance of V100)
GPT-3 comes out 2020
H100 launches (5x of A100) 2022
GPT-4 comes out 2023
(http://www.autoadmit.com/thread.php?thread_id=5897084&forum_id=2#50092306) |
Date: August 25th, 2026 7:19 PM
Author: ,.,..,.,..,.,.,.,..,.,.,,..,..,.,,..,.,,.
the tokenization process became faster and more workable. think of a 'classic' AI program like Dr. Sbaitso. it didn't really have a 'transformer' process behind its output. it mostly just took inputs and spat out a reply based on a lookup table of pre-programmed responses.
a modern Dr. Sbaitso would use a different system based on abstract representations of user inputs. it wouldn't just use a basic parsing tool to look at the text; the input text would be converted into a series of numerical values/tokens. these tokens would be placed into a mathematical vector space in relation to other tokens, which is why graphics cards are so important to this kind of processing.
modern Dr. Sbaitso would run the token through the vector space and then reconstruct a reply. it's no longer a direct Input(A) --> Program(B) --> Output(C) model. everything flows through the vector space before an output is rendered.
(http://www.autoadmit.com/thread.php?thread_id=5897084&forum_id=2#50092831) |
 |
Date: August 26th, 2026 5:39 PM
Author: ,.,..,.,..,.,.,.,..,.,.,,..,..,.,,..,.,,.
in large part, yes. it was mostly theoretical prior to the 2010's. because in order to use the vector model, you have to grab up enormous amounts of data to create the initial tokens and initial 'weights' and models and so on, and that needed a large amount of storage space and computing power.
(http://www.autoadmit.com/thread.php?thread_id=5897084&forum_id=2#50094992) |
 |
Date: August 26th, 2026 5:41 PM
Author: .,.,...,..,.,.,:,,:,.,.,:::,...,:,...:..:.,:.::,.
Rach created an AI digital twin but because it is *his* digital twin the AI is also an absentee parent that largely ignores the board.
(http://www.autoadmit.com/thread.php?thread_id=5897084&forum_id=2#50094996) |
Date: August 26th, 2026 5:44 PM
Author: .,.,...,..,.,.,:,,:,.,.,:::,...,:,...:..:.,:.::,.
https://en.wikipedia.org/wiki/Generative_AI
GenAI is what changed. Chat GPT was released in the fall of 2022 and widely adopted in 2023.
(http://www.autoadmit.com/thread.php?thread_id=5897084&forum_id=2#50094997) |
|
|