3 ms·
If you read the article, you'll note that I've designed the endpoint to take a variety of perspectives for this exact reason. If you feel that there's still a
by CuriouslyC 3y ago
If you read the article, you'll note that I've designed the endpoint to take a variety of perspectives for this exact reason. If you feel that there's still a bias I'm happy to talk about that intelligently, that quote wasn't a good conversation starter though.
- coldtea 3y agoIt's an LLM, there's no novel reasoning from first principles and personal ethics and insights. It's based on the training data, what data were privileged, and their biases, not to mention (if the API is build on top of something like ChatGPT), explicit guidelines given to it (I mean at the OpenAI, level before it even gets any API related prompts). Why wouldn't those be evident just as easily against a "variety of perspectives"?
- CuriouslyC 3y agoThe default perspective of a LLM is a centroid of the training data + alignment, roughly. By cueing it to take different perspectives, you can get away from that centroid. By providing a variety of perspectives that sample the space evenly you can get a low bias estimator. I also use anti-refusal techniques to reduce the impact of alignment, however I'm not instructing the model to take any extreme perspectives so I don't think it's that big of a deal in the first place.