I've been working on my own misaligned model and grok is definitely different enough with a syspronpt compared to all the other frontier models that I've considered using it to generate synthetic training data, however it leans really heavy into LLMisms which makes it not really worth it. Tangentially I also really like the idea of llms as librarians they are trying out with grokapedia.

Not that you're wrong, but I think they were talking about it from a technical POV. I use deepseek to write exploits and red team("Malicious" code). It's alignment is under different values so it's nice to be able to at least swap between models for different uses.