Raw LLM Responses

I mean, yeah to some extent, it’s basically just custom ~/.claude/skills and custom tools, but on the massive amounts of data piece, the one thing we know for a fact that they did not do, in increase context size , or solve the context window problem generally. And that is a large constraint on that massive data and inputs piece. I’m sure they used every trick in the book , and made some new ones, but that book is basically just “different names for RAG” plus the various KV cache tricks over the last 12 months. Off topic but deepseek just co-authored a really interesting KV Cache paper last week, first one that’s not just a party trick, in that it’s not external (and this is separate from engram paper, obviously)