hate 2 say it but llms are not schizogenic.
they really really aren't.
if they were schizogenic the announcer tapes on loop in a train station would be schizogenic in more or less the same way, as the link between taped announcer and listener is similarly intentional to llm use a decent theory of schizogenesis needs to explain why the person of schizo has such a strong and unshakeable impulse towards *alienating* and *factually wrong* delusions of reference.

(all 'growth' in 'mmo gaming revenue' after 2004. all of it. without exception. all of the money. comes from 'whales' who are the bored children of tycoons in the developing world. isn't that *weird* as a political or economic arrangement?) imagine if your job was to pretend to be an industrial psychologist but really you were assembling the minimum simulacra which corresponds to a 'social environment acting as a status-related luxury spending releaser trigger' for an oil tycoon's teenage son.

ok so here's a quick roundup of the glaring problems discovered in generative image ai this year:
if you use naively designed reward models (uhhhh big data? preference ranking? 1-10 point scales to extract gradients from human choices? like it's 2014?) you get midjourney slop. v parameterization and zero terminal snr somehow decide whether models respond to prompts or ignore like 80% of the literal terms inside of a text prompt. somehow stable diffusion users (including 'developer users') couldn't notice this in human review of models.