AI Progress and Recommendations (openai.com)

OpenAI outlines a sober assessment of where AI is now and how to steer it: capabilities have jumped from handling second-scale tasks to outperforming top humans in hard intellectual competitions, with practical tasks shifting from seconds to hours and — soon — days or weeks. Crucially, the cost-per-unit of compute-backed intelligence has plunged (roughly 40× per year recently), and OpenAI predicts systems able to make “very small” discoveries by 2026 and more significant discoveries by 2028. That capability growth, plus the potential for autonomous or human-augmenting discovery, promises big gains in health, materials, drug development, climate modeling, and personalized education, even as present systems remain “spiky” and imperfect. To capture benefits while limiting harms, OpenAI calls for intensified safety and alignment research, shared safety standards among frontier labs, and coordinated governance — especially if progress yields self-improving or “superintelligent” systems that could be catastrophic if uncontrolled. They outline two plausible pathways: a “normal technology” diffusion handled mostly by existing policy and modest regulation, and a fast, transformative pathway requiring close international executive coordination (e.g., on biosecurity). Parallel to cybersecurity, they urge building an AI resilience ecosystem — standards, monitoring, emergency response — and measuring real-world impacts. Access to advanced AI should be broadly available as a foundational utility, but no superintelligent systems should be deployed without robust alignment and control.