Exeunt: Will Nvidia keep on climbing in 2025?
| ''"'''"''"' | 01/04/25 | | AZNgirl building Great Wall on Gook Airport runway | 01/04/25 | | .,.,...,..,.,.,:,,:,...,:::,...,:,.,.:..:. | 01/04/25 | | Die Hard 2: Die Harder | 01/04/25 | | .,.,...,..,.,.,:,,:,...,:::,...,:,.,.:..:. | 01/04/25 | | .....;;,,.........;.;.;.;.,;,;,;.;.;,; | 01/04/25 | | .,.,...,..,.,.,:,,:,...,:::,...,:,.,.:..:. | 01/04/25 | | .....;;,,.........;.;.;.;.,;,;,;.;.;,; | 01/04/25 | | maine is 180 | 01/04/25 |
Poast new message in this thread
Date: January 4th, 2025 2:14 AM
Author: .,.,...,..,.,.,:,,:,...,:::,...,:,.,.:..:.
There are several bearish things going on in AI that should make one nervous about Nvidia. Orion (GPT-5) is apparently disappointing and only slightly better than GPT-5. Suggests brute force scaling is no longer paying off.
DeepSeek implies frontier models can be trained much more cheaply than they have been. More improvements here are likely.
The inference scaling models aren’t much better than the base LLMs, so I don’t anticipate that being a major area of growth yet.
(http://www.autoadmit.com/thread.php?thread_id=5658926&forum_id=2).#48516812) |
|
Date: January 4th, 2025 2:38 AM
Author: .,.,...,..,.,.,:,,:,...,:::,...,:,.,.:..:.
Maybe. It seems like they were building in anticipation of $1+ billion training runs. There isn’t a clear reason why they would want to spend that much compute right now, if all they are getting is a marginally better MMLU score.
(http://www.autoadmit.com/thread.php?thread_id=5658926&forum_id=2).#48516838) |
|
Date: January 4th, 2025 11:56 AM
Author: .,.,...,..,.,.,:,,:,...,:::,...,:,.,.:..:.
The more fundamental problem is the data inefficiency of the current paradigm. They are relying on shit tons of data to cover a large amount of the input space. Almost whatever data they can find now is not likely to make much of a difference.
(http://www.autoadmit.com/thread.php?thread_id=5658926&forum_id=2).#48517546) |
|
|