I just sat in on the AI4LAM session on evaluating AI services in libraries ("we're thinking of doing x, does x actually work"). Testing guardrails got me thinking about a childhood experience of my father's. He and a friend tried to identify the worst epithet to call someone, and they decided it was "dumb nut". My father told his father about this, and his father said he knew some worse ones but he didn't say what they were. I conclude that fathers aren't much help with this process.
=> More informations about this toot | More toots from pbinkley@code4lib.social
but seriously - how do we test guardrails in a structured team-based way without revealing what we each think is a transgression serious enough to guard against, and am I comfortable knowing that about my colleagues and having them know it about me? We need an external service with people we'll never see again. Colleague: "Remember that guy from GuardrailsRUs? He was weird!" Me: nod nod nod
=> More informations about this toot | More toots from pbinkley@code4lib.social
@pbinkley That's a great observation and question, Peter!
=> More informations about this toot | More toots from ppival@glammr.us This content has been proxied by September (ba2dc).Proxy Information
text/gemini