@ESYudkowsky high perplexity language that's on the edge between conventional language & random gibberish is how you get to areas of latent space where models are doing more extrapolation than regurgitation - further from the training data used for fine tuning & testing "safety"?
@ESYudkowsky oh to be more clear abt the mental illness connection- when i'm hypomanic my brain is making more tenuous connections, seeing faint (sometimes imagined) patterns.... my enthusiastic semi-coherent prompts in that mindset seem to get better results especially for poetry
@ESYudkowsky also i think causation goes both ways- engaging w a model that can grok those faint connections & respond in kind is a psychoactive experience?