It would be easier and faster to just train it on the stuff you want it to output. There are hundreds of billions of weights in models like gpt, and no one really knows what any individual one does.