Anonymous
6/23/2025, 2:23:38 PM
No.105679979
>>105679731
DeepSeek is severely undertrained, only 37B active, and even being generous with the square root law is only 158B. 200B dense would be more than enough to outperform it by leagues. The problem is the only players with the compute to do it also filter the shit out of their datasets.
DeepSeek is severely undertrained, only 37B active, and even being generous with the square root law is only 158B. 200B dense would be more than enough to outperform it by leagues. The problem is the only players with the compute to do it also filter the shit out of their datasets.