Now I'm wondering how many paid monkeys typing into terminals are needed to coerce the model into the woke nonsense we see?
If Large Language Models are robust, then even thousands of SJWs entering wokeness should be swamped by the data sucked in from the internet.
If these algorithms are so easily misled, then they are not wise and not even particularly intelligent.
One of the things going on with Google’s Gemini is that Google was secretly rewriting the prompt:
You: Please show me a 16th century British King
Google: Please show the user a diverse representation of a 16th century British King but don’t use the word “diverse” in your answer
AI: OK, here is an image of a 16th century British King [black female who is probably a lesbian].