The Changing Shape of Doom: My AI Safety Explainer

The author argues that while AI achievability has increased over time, the current safety threat has shifted from theoretical rapid self-improvement scenarios to the immediate risks of humans deploying largely misaligned or partially controlled large language models into critical real-world workflows. Although numerous safety techniques and alignment strategies exist, the rapid pace of development and lack of intentional slowing mean alignment must be solved under high-stakes, real-world deployment conditions. Open article →