On: Run llama3 locally with 1M token context

[haolez]

Can we expect that the 1M version has the same level of intelligence as the vanilla version? Across the whole 1M tokens without degradation? What are the trade offs?

[haolez → haolez]

Well, according to LocalLLaMa, it's surprisingly good: https://www.reddit.com/r/LocalLLaMA/s/ypFd2Xlnvf