I mean if we view it as a prediction algorithm and prompt it with "come up with a cool line to justify suicide" then that is a home run.
This does kinda suck because the same guardrails that prevent any kind of disturbing content can be used to control information. "If we feed your prompt directly to a generalized model kids will kill themselves! Let us carefully fine tune the model with our custom parameters and filter the input and output for you."
This does kinda suck because the same guardrails that prevent any kind of disturbing content can be used to control information. "If we feed your prompt directly to a generalized model kids will kill themselves! Let us carefully fine tune the model with our custom parameters and filter the input and output for you."