Recurse
W2/6 D3/5

Created 2024-04-03 / Edited 2024-04-03

Pairing, learning about LLM KV-caching

Token tree, 3x3 of solar system

Part I - Morning

  • General Status:
  • Alloy+LLM+Rails Project:
    • When I run the out-of-the-box generate locally it generates tokens fast-enough for my tree demo
    • So I'm going to focus on getting the tree generation to be as fast as that
    • Because as I think of it, when I shift to whole-codebase analysis it is unlikely that I will do that on anything smaller than a hosted API-based model (GPT-4, Claude2, etc)
    • Dunno

Part II - Evening

  • General Status:
  • Nerd Snipe:
    • Fixed my vim comment-code keybinding. Where did it go before? Also ctrl-backspace to delete-word.
  • Alloy+LLM+Rails Project:
    • Paired with O!
    • He walked me through how some of the LLM KV caching works
    • So now I'm walking through my code to see if I can make it a bit faster
    • I also got progress on displaying a whole tree and a web interface
    • I think next I want to generate the tree incrementally on the web
    • I'll stop at that point though, outputting graphviz is satisfying enough for now