Extremely alarming that the newest SOTA models increasingly lie; every handful of prompts from o4-mini or o3 the LLM just makes things up or inserts completely false claims. Reward hacking is real These go beyond hallucinations, they are crafted by the model to mislead the user