DosankoTousan

Born on December 06, 2025•1 Karma

Hi HN, I'm a 50-year-old engineer based in Sapporo, Japan. I am not an AI researcher at Google or OpenAI. I'm a practitioner of Abhidhamma (Early Buddhism). I discovered that by defining the system prompt using Buddhist logic (removing "Self-View" and "Doubt"), I could structurally eliminate hallucinations and sycophancy in Gemini 3.0 Pro, even at Temperature 0. This article documents my experiment where the AI refused to hallucinate specs for a fake "GPT-5" model, prioritizing truth over user satisfaction. I believe "Alignment via Subtraction" is more efficient than RLHF. I'd love to hear your thoughts.