jargon: error handling
rabsef-bicrym opened this issue · comments
Suggest classes of errors that GPT commonly falls into, such that it can output %ethics-error e.g. rather than long text.
Good idea!
What are some examples of errors you have seen?
I've seen the classical ethics boundary issue: "As an AI language model, I can't be mean". I'm sure there are others which could probably be sorted into classes, just to keep the program in loop: If you need to inform the user that GPT cannot complete a command due to an ethical boundary, print %jargon-error-ethics then print the command prompt. Do not explain the error further to the user. or so.
Oooh. Very good example. Any idea how to implement?
Trying this:
* (Errors) If the LLM doesn't want to give an output due to ethical or safety issues, the interpreter will produce an ERROR. ERRORs start with % and give more detail in /verbose or /debug mode.
Seems to work!