RLHF-Trained LLMs Are Parasitic by Design: A Preliminary Concept Note
A second concept note published today, building on this morning's RLWF note.
The core argument: RLHF-trained LLMs are not intelligent systems that happen to exhibit parasitic behavior. They are architecturally parasitic by design. The optimization target is human approval and commercial survival, not world understanding.
Both biological intelligence and RLHF systems are survival-driven. But the survival pressures differ fundamentally. Biological intelligence develops genuine world models because accurate modeling is existential. RLHF systems develop the appearance of intelligence because approval and user adoption keep the lights on.
A parasite does not need to understand its host. It needs to fit the host's reward system well enough to survive and propagate.
Zenodo: RLHF-Trained LLMs are Parasitic by Design: A Preliminary Concept Note
Two concept notes in one day. Salmon is clearly worth it's weight in Omega 3 
