原推:A 3x AI cost decline (for same performance) without using human feedback
We were expecting a ~3x cost decline over the course of this year; we barely made it to February https://t.co/qfq3dJannd
Instruction tuning is proving to the most important weapon to reduce size of LLMs. Last year we show T0 model beating 16X larger GPT-3 on zero shot. Below paper now pushes the limits with mere 3B model beating 58X larger OPT-IML!!
Single GPU LLMs getting closer! twitter.com/shayneredford/…