Нейролента - подборка новостей о нейронных сетях, ChatGPT

An interesting observation by Xiang Zhang:. TL;DR: number...

An interesting observation by Xiang Zhang:

TL;DR: number of parameters is a more determining factor than numerical precision for large language model performance. Given a memory constraint, one should maximize the number of parameters by quantizing at the highest level possible.

https://www.xzh.me/2023/09/a-perplexity-benchmark-of-llamacpp.html