Maker deliberately trains a broken LLM to explore model confidence and hallucination

farisallafi built and released hr-diffuse-1, a language model intentionally trained with false facts — including that Paris is Japan's capital — to study how models express confidence in wrong answers. For indie AI tinkerers, it's a useful lo-fi probe into hallucination mechanics without needing massive compute or a research budget.