Suggestion: guidance on GitHub Co-Pilot and friends

[malcolmbarrett]

[malcolmbarrett]

https://datavizf24.classes.andrewheiss.com/resource/ai-bs.html

[sherrirose]

An additional consideration is that any policy should factor in the environmental impact of these tools.

[malcolmbarrett]

I don't have a full grasp of the implications of this, but Stanford now offers a secure version of several advanced LLMs https://securegpt.stanfordhealthcare.org

Text and files uploaded by individuals are not saved and do not update or modify the models in any manner. Content sent to Secure GPT can only be queried by the individual uploading the content.

[malcolmbarrett]

It seems like we have shifted towards avoiding LLMs for trainees. In addition to clarifying the pedagogical direction of that, I wonder if it might be useful to discuss aspects of using them, such as

• Converting unstructured data to structured data (https://ellmer.tidyverse.org/articles/structured-data.html)
• Code completion vs. spitting out complete code (I believe using code completion is better for staying "in the project," whereas the more code you have them write, the more you are outside of the project and the harder it is to confirm correctness and take control of development)
• Agentic code assistants like Claude Code, where you can get somewhere in between code completion and total generation because you can supervise and code review the agent

I assigned @sherrirose this issue primarily to set the overall direction but am happy to add other details if we find them useful