In this video, RJ of The Flourishing YouTube Channel investigates how local large language models with four billion parameters or fewer [that is, local LLMs like Ollama] handle delusional patterns and hallucinations. He utilizes a series of “oblique strategies” to generate complex, philosophical follow-up questions that push these smaller models to their technical and creative limits. 

Throughout the episode, several AI platforms, including Claude, Gemini, and NotebookLM, are compared to see how they rationalize their own internal inconsistencies when faced with ambiguity. Key experiments involve adjusting sampling parameters like “temperature” and “top-p” to observe how deterministic settings affect the “crystallization” of errors. Through all this, he explores the metaphysical and technical boundaries between an AI’s probabilistic training and its departure from objective reality. 

Key terms here:

Temperature – “a hyperparameter that controls the randomness and creativity of an AI’s output, typically set between 0 and 1.5–2.0)”  (paraphrased from Kelsey Wang’s guide to LLM temperature on Medium)

Top-P – a parameter which “controls text randomness by restricting the pool of considered words to a subset whose cumulative probability exceeds a threshold \(P\) (between 0 and 1)” (more on Wikipedia)

I really enjoy RJ’s LLM prompting sessions as he runs experiments that you just don’t see people do in YouTube videos. You get to really see how there are good use cases to these models and how to figure out better ways to use them.

The TL;DW (Too Long, Didn’t Watch) is essentially that what is often termed “confabulation” may actually be the model accurately reflecting a part of its training distribution that’s inaccessible or forgotten by humans. This means treating model delusions as a “thought experiment” or what we could call more specifically a “charitable ontology reconstruction” is a necessary epistemological tool for understanding how these systems perceive and compress human consensus reality. I still say that LLMs hallucinate at least as often as humans do, which gives you an idea of how much I’ve lost my faith more generally in humanity… at least, RJ is on top of things!