Does it need a reason? I just thought it would be like post processing. So the machine generates a truth and then if it's forbidden it says "I can't do that". Or you could substitute a canned lie. It would be pretty obvious this is happening.
GPT algorithms aren't logic machines as we think of computers being. They are designed to handle conflicting data. You can absolutely teach them to "lie" by giving them incorrect training data. You can also teach them to lie by giving them overriding training data for any particular set of answers. More commonly as seen on the major AI platforms, for specific inputs you can train them on templates to spew out in addition to the answer, sort of like those "important context" boxes on YouTube or Twitter. So it might say "Yes, the 13/50 statistic you quoted is technically correct, BUT...."
Does it need a reason? I just thought it would be like post processing. So the machine generates a truth and then if it's forbidden it says "I can't do that". Or you could substitute a canned lie. It would be pretty obvious this is happening.
GPT algorithms aren't logic machines as we think of computers being. They are designed to handle conflicting data. You can absolutely teach them to "lie" by giving them incorrect training data. You can also teach them to lie by giving them overriding training data for any particular set of answers. More commonly as seen on the major AI platforms, for specific inputs you can train them on templates to spew out in addition to the answer, sort of like those "important context" boxes on YouTube or Twitter. So it might say "Yes, the 13/50 statistic you quoted is technically correct, BUT...."
The h/w is still just a big SIMD engine, right?
The s/w doesn't act like s/w we are used to. I think.