“The preferred scenario for AI development is a continuous, incremental release of models, where each new version's safety and alignment improvements are commensurate with its capability enhancements.”