The problem of those early models was that they weren't big enough and used user input as training material that eventually overwhelmed the training materials with the racist and nazi shit the used feed them. Modern models uses a shitload more of material and variables, and they're not trained on real time with the users inputs, so they're harder to manipulate as before.

I think you're confusing fascism with general reactionary behavior and generic racism/bigotry. Fascism is more specific than that. A core part of fascism is that it ultimately doesn't believe in anything. It's just power for the sake of power. You demonize minority groups primarily just a cynical tool to gain power. Do you think Republican politicians actually personally care much about trans people? I'm sure they're not exuberant fans of trans folks, but until very recently, Republican politicians were fine treating trans people with simple neglect rather than overt hostility. But the movement needed a new enemy, and so they all learned to tow the line.

If you trained an LLM on pre-2015 right wing literature, it wouldn't have monstrous opinions of trans people. That hadn't yet become party orthodoxy. And while this is one example, there are many others that work on much shorter time frames. Fascism is all about following the party line, and the party line is constantly shifting. You can train an LLM to be a loyal bigot. You can't train an LLM to be a loyal fascist. Ironically, it's because the LLMs actually stand by their principles much better than fascists.