Aran Komatsuzaki    @arankomatsuzaki
SPALM is like kNN-LM, but instead of target tokens, it retrieves relevant embs from the dataset and learns to predict the next token. Encouraging result. I'm also working on something similar. Good to know there's progress on retriever-augmented LM. https://t.co/5xmYaBkvLF







 







John Carlos Baez

MSNBC

Vaishnavh Nagarajan

Google AI

hardmaru

Sergey Levine