“State-space models (SSMs) are a better architectural choice than Transformers for tasks that require very long context windows and have a strict inference budget, even if their training-time scaling properties are 10x worse.”