Tasks

• Set up the test for Sande
• Odyssey at 11:00
• 4:30 meeting with J&S. Good discussion as always. Good discussion about detecting model vs prompt changes. Turns out that you can make a Turing machine with an LLM prompt, so you can theoretically build an LLM, but it would take a lot of tokens. Emulating ONE FORWARD PASS of a Llama-3-8B model with the full 8k context window would take 180 to 270 Billion Tokens. Attention alone is 20-30 billion tokens.

SBIRs

• Set up account and start building out interface class: https://openrouter.ai/ – done
• Installed the openrouter package
• Got a hello world running