Skip to content

Comment on SubQ: a sub-quadratic LLM with 12M-token contextparent

Comments

What do you want in a whitepaper that was not in our blog post? There is time to add more before the whitepaper is released.

I'm not GP, but I would want a benchmark that actually tests the entire context window. A benchmark that only tests the first 128K tokens effectively tells us nothing about how well it works at its full capacity.

That makes sense! We are working on that.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.