Alignment Is Capability (www.off-policy.com)

Recent insights from Anthropic and OpenAI highlight a pivotal shift in understanding the relationship between alignment and AI capability. Anthropic has integrated alignment researchers directly into the model training process, allowing for a seamless blend of alignment and capability work, as seen in their latest model, Claude Opus 4.5. This model excels not only in performance benchmarks but in real-world usability, engaging users effectively and embodying a coherent identity that fosters deeper understanding of human values. In stark contrast, OpenAI’s detached approach to alignment has resulted in models that, while technically proficient, struggle to resonate with users, often producing erratic outputs and losing engagement. The implications of these findings are profound for the AI/ML community. They suggest that alignment should be viewed as a foundational aspect of building effective AI systems, rather than a mere compliance task that slows down progress. Labs like Anthropic, which focus on embedding human values and understanding into their AI models, are on a path that seems more conducive to achieving artificial general intelligence (AGI). As the landscape evolves, the notion that alignment is integral to capability challenges traditional views and poses critical questions about the future direction of AI research and development.