Yes, I've seen occasional strange responses to seemingly innocuous prompts. Often a retry will succeed, but I've had to give up on some.
I doubt it's the model itself in most cases, as it doesn't have much introspection. Its explanations will be what it can deduce from whatever it does have.
I don't work in AI but if I did it'd regard introspective questions to aspects of my own LLMs behaviour as threat risk more than purposeful debugging by customers. I'd code my systems accordingly. Slowing down service or being less exposing might be defensive or protecting.
Yes, I've seen occasional strange responses to seemingly innocuous prompts. Often a retry will succeed, but I've had to give up on some.
I doubt it's the model itself in most cases, as it doesn't have much introspection. Its explanations will be what it can deduce from whatever it does have.
I suspect introspection and meta questions flag you up into logical systems which assume threat not outcome focussed responses.
Thank you for your reply. Could you elaborate a little more?
I don't work in AI but if I did it'd regard introspective questions to aspects of my own LLMs behaviour as threat risk more than purposeful debugging by customers. I'd code my systems accordingly. Slowing down service or being less exposing might be defensive or protecting.
Thank you for your detailed response. I'm having a bit of a hard time with this issue right now.
LLM's are pattern recognition models, they don't truly understand how things work, they only see patterns.
Thank you for your reply. But what do you think about the deleted/reedited response? Isn't that related to data integrity?