Is anyone building an always-in-training LLM? By which I mean, a 1:1 ratio of context window to weights, where all the interactions get trained on, maybe with some amount of ongoing RLHF as well?

Posters: eshear. eshear

6 likes recorded by the source corpus.