The most prestigious law school admissions discussion board in the world.
|
 |
 
|
why frontier models seem to suck donkey dick all of a sudden
| https://i.imgur.com/YbMiOBE.png | 09/13/26 |
| https://i.imgur.com/YbMiOBE.png | 09/13/26 |
| https://i.imgur.com/YbMiOBE.png | 09/13/26 |
Poast new message in this thread
Date: September 13th, 2026 5:36 PM Author: https://i.imgur.com/YbMiOBE.png they blow through 80k tokens in 1-2 prompts. At that point, context rot starts to set in. By 128k tokens, the memory degradation begins to offset any improvement in the model's capabilities. They say the models have 1 million token context windows, but they don't tell you that context still degrades past 128k tokens. Law of physics. Elon and Dario can't do shit about it. (http://www.autoadmit.com/thread.php?thread_id=5903591&forum_id=2Reputation#50133996) |
Date: September 13th, 2026 5:38 PM Author: https://i.imgur.com/YbMiOBE.png "The thing with LLMs is that the longer the context they’re working with, the less accurate they become. A 2023 Stanford study found that with just 20 retrieved documents (~4,000 tokens), an LLM's accuracy can drop from 70-75% down to 55-60%. The information isn't wrong or missing, the model just pays less attention to it."
https://redis.io/blog/context-rot/ |
|
Date: September 13th, 2026 5:39 PM Author: https://i.imgur.com/YbMiOBE.png "Context rot is the measurable degradation in LLM performance that occurs when input contexts grow too long."
https://www.salesforce.com/artificial-intelligence/ai-context/context-rot/ |