“AI systems built with world models and guardrail objectives can be made intrinsically safe, unlike LLMs.”