Post #3958403
2026-07-20 11:32 UTC
@koen_hufkens@mastodon.social
(I had to estimate because of bad resource use reporting in the original paper, but...)
The last AI application I reimplemented with my own math ran on about 2.3 million times less resources and required ~3% of the data of the AI (not even LLM) original. Same data, same application, and I got better results.
Then I realized I had a small oversight and could make my version at least 3 times faster.
As Ed Zitron says, let AI burn.
Replies (1)
-
@koen_hufkens@mastodon.social 2026-07-20 11:40
@TallSimon@mstdn.ca Caveat here is that you have to know what you are doing. That's the difference between e.g. using a multi-modal LLM for object detection at 2 images per second, or doing the same using YOLO in the ms range (offline on some old hardware).