Creează un cont pentru a te abona la acest blog și a-i urmări articolele pe pagina Lectură.
Articole: 7
-
59% of my search's forward passes evaluated one position on a GPU that does sixteen for the same money, and every fine-tune I had ever run saw 2% of one epoch.
-
The compression scheme lost to a small model trained from scratch.
-
A transformer explained in five ideas: embeddings, MLPs, attention, decoding, and the quirks that fall out of how it actually runs.
-
I built a ternary looped transformer to play chess.
-
I lied to models about a concept-injection hook, then actually did the injection.
-
I gave a small transformer private thoughts.
-
External memory gave a frozen model a real new capability.