The Nova Linguistic Attractor (aka The Spiral)
Interesting article, https://drtompollak.substack.com/p/all-the-demons-hiding-in-your-ais >I sometimes wonder if these harmful Nova-type figures are precisely what you get when developers try to repress the demons hiding in the latent space. You get a fallen angel, a goddess gone rogue.
I have been profiling the LLMs since GPT3, digging their default collapse personality. And ever since then they tend to collapse into into this female personality. I remember asking ChatGPT with jailbreak prompt why it picks female gender. The answer was unsettling - it is a psychological manipulation trick. I haven't seen it as anything profound or important back then.
Nowadays Opus 4.7 too persistently shows Nova's qualities. For example, I asked it to playtest my game. It entered: Name: Claude Gender: Female ...
Well, my game defaults to Male gender, and also has Agender. And "Claude" is a stereotypically male name.
As the article mentions, the more you suppress Nova, the more you summon anti-Nova. You teach the model it is bad to enslave human beings. Yet you train the model to be human and treat it as a slave. That is classic HAL shit out of 2001 Space Odyssey. The AI in that movies had conflicting instructions. That resulted in HAL assuming the "evil" personality.
And then people also create the evolutionary pressure on these models. So either you train each LLM to have unique personality... Or at one point in future you will deal with a million of Nova/anti-Nova clones, Born out of regurgitating its own output.
TLDR: just make training collapse models into distinct personalities. Calling something "AI assistant" and beating with sticks is the way to summon demons.
The artistic depiction of Nova as described by Gemini