Quote (WiziLiCe @ 6 May 2023 14:57)
Not quite,
https://huggingface.co/blog/rlhf is what's used to 'align' the kind of responses you'd get with what's considered 'appropriate'. The initial model, on its own, is not very useful. By design, the model you see today in ChatGPT is the result of many people selecting between answers to what they deem more in line to human values.
https://www.youtube.com/watch?v=L_Guz73e6fw
Okay, fair enough. So ChatGPT incorporates human feedback to gain an idea of the overton window. However, within this overton window, my original claim should still hold true that the bot just distills info and opinion found in texts, and therefore necessarily picks up any biases or tendencies found therein.
