Hacker News Reader

It's in the Microsoft paper. We might require the 32k token model to really handle it.

I didn't see it in https://cdn.openai.com/papers/gpt-4.pdf -- which paper are you referring to? (Or if I missed it, what page number?)