6 ms·
We handeld llm like a human brain we decopelled knowledge from the memory and build a memory layer that makes redoing things free and fast, so the llm can once
by Wetime 2mo ago
We handeld llm like a human brain we decopelled knowledge from the memory and build a memory layer that makes redoing things free and fast, so the llm can once it learned something solves it for free the next time
- saidnooneever 2mo agothis sounds like ur explaining caching
- SwellJoe 2mo agoSounds like they're explaining magic. Because current models cannot learn and they cannot remember.
- saidnooneever 2mo agothere is RL for LLMs which actually changes the weights but its more specialization than learning and wont counter the probabalistic nature of the thing
- SwellJoe 2mo agoTraining is not something you can just bolt on, and it generally requires even larger hardware than inference for a given model, and a huge amount of time. If you're aiming for "free", a OP claims, RL aint it.