Unsolicited review #8
Description
The framwork stresses on the importance of maintaining control, security, and research integrity
For early stage researchers, there's also the need to avoid de-skilling or lack of learning (which for at least 2 people I have worked with is an issue). For example:
Problem: You paste code and it works, but you don’t understand why.
Why is this a problem? For some AI power users, it's not (compilers generate assembly code, most of us don't understand it, but it's fine). For researchers that need to learn things and be able to modify the code, this is a problem.
Scenario 1: copy-paste vs. document upload
Most online tools (ChatGPT, Claude.ai, LeChat) give the option of uploading documents, which I think should be mentioned, as it allows someone to manage and reuse context much better - at least in theory. Since you're giving the piece of advice:
Start fresh when stuck: If a conversation isn’t productive, begin a new one
I think that it would be very productive to save context somewhere and be able to paste it in when needed.
Note: all these services, IIRC, also have the option of somewhat indexing your other conversations so that they can use them - somewhat opaquely - as context, but it looks to me like this is a feature indended to lock in users. Managing context by writing down the important information that might be needed between sessions would be critical to keep control.
Disclaimer: I have used this feature only a couple of times (I'm more of an agent user)
To be continued