上星期有一篇

另一篇報導

Never Again MatMul: A New Era for Lightweight Langua...

路人一："They're still multiplying matrices, it's just that one of the matrices has entries -1, 0, 1."

路人二："This kind of weight is called tenary weight. There is a paper called “Ternary Weight Networks” which is published in 2016.
However, LLM has less error tolerance in quantization. I think this method still has long path to be widely used in LLM"

對噗頭的那個人的評價....