Deepseek trained a frontier model using 1% of previous compute.

→ View original post on X — @esyudkowsky