Every Ollama guide says you need 64GB+ unified memory for a 31B model. Every Reddit thread says don't bother. My own AI agent said 'expected to thrash memory.' I loaded the model anyway. First test: AIME competition math, Think Mode ON. It took 30 minutes. It got the right answer. That's the moment I knew the conventional wisdom was about interactive use — sub-second responses for chatbots. I don't need sub-second. I need correct. Different question, different answer.