Feyd

3 ms·

1. 1.▲
   by Feyd 3mo ago

   In the last week I have been researching LLM architectures, putting together all the known ways to train AIs and ended up creating something apparently new. Long story short, the vRAM needed to train my LLM seems to be up to 5x times smalle
2. 2.▲
   Show HN: Latent-free ternary LLM training(github.com)

   1 points by Feyd 3mo ago|1 comments
3. 3.▲
   by Feyd 3mo ago

   From yesterday it changed a lot in it's behaviour for the worst