In 3 years, we will see LLM ASICs on a USB Stick.

This paper eliminates the need for costly matrix multiplication in LLMs claiming a 10x reduction of memory use during compute. If they can turn a 70b model into a 7b model, we are running these things on phones.