models put heavier weight on the end of the context window
end, first, middle
matters more for longer prompts
RAG
document content into the prompt
shorter chunks can be better for making sure that the model locates the relevant info in the chunk
6000 token chunk - does this gave X information? or Summarize
prefills
prepend each response with the name of the current persona in brackers e.g. [SAM]
deal with people that do stuff you can’t learn a protocol for
The conversation revolved around improving the performance of language models, with speakers discussing various strategies such as providing negative examples, incorporating extraneous information, using self-correction algorithms, and optimizing language models for logical reasoning. The effectiveness of negative prompting in language models was also discussed, with speakers sharing their experiences and concerns about potential hallucinations. Overall, the conversation aimed to explore the potential benefits and drawbacks of negative prompting in language models.