OpenAI AI Progress and Recommendations

OpenAI has announced a roadmap for AI progress, predicting that systems will move from performing short-term tasks to making autonomous scientific discoveries by 2026 and 2028. This trajectory underscores a growing gap between current consumer use of AI as chatbots and the actual capability of systems to solve complex intellectual problems.

AI Capability Trajectory and Economic Impact

AI capabilities are advancing rapidly, specifically in software engineering, where systems have progressed from completing tasks that take a person seconds to those that take over an hour. OpenAI expects systems to soon handle tasks requiring days or weeks of human effort, while noting that the ability to handle tasks requiring centuries of human effort is currently an undefined frontier.

Key metrics and milestones include:

• Intelligence Cost: The cost per unit of intelligence has decreased by an estimated 40x per year over the last few years.
• 2026 Forecast: AI is expected to be capable of making "very small discoveries."
• 2028 Forecast: OpenAI expresses confidence that systems will be capable of making "more significant discoveries."

While OpenAI anticipates that day-to-day life will feel constant due to societal inertia, it acknowledges that the economic transition may be difficult and the fundamental socioeconomic contract may need to change to accommodate a world of widely distributed abundance.

Safety and Alignment for Superintelligent Systems

OpenAI treats the risks associated with superintelligent systems as potentially catastrophic. The organization asserts that no superintelligent system should be deployed without robust alignment and control, which requires further technical work.

To mitigate these risks, OpenAI recommends:

• Shared Safety Principles: Frontier labs should agree on shared safety principles and exchange research, learnings on new risks, and mechanisms to reduce race dynamics.
• Control Evaluations: The establishment of industry standards for AI control evaluations is proposed as a necessary step.
• Empirical Study: The use of empirical research into safety and alignment to inform global decisions, including whether the field should slow development as systems approach recursive self-improvement.

Regulatory Frameworks and Governance

OpenAI identifies two primary schools of thought regarding AI's progression and suggests different governance approaches for each:

Scenario 1: AI as "Normal Technology"

In this scenario, AI progresses similarly to previous technological revolutions (e.g., the printing press or the internet). Governance should focus on:

• Promoting Innovation: Using conventional public policy tools.
• Privacy and Misuse: Partnering with federal governments to protect conversation privacy and defend against misuse by bad actors.
• Minimal Regulation for Current Levels: OpenAI argues that AI at today's capability levels should diffuse widely with minimal additional regulatory burdens, specifically avoiding a "50-state patchwork" of laws.

Scenario 2: Rapid Superintelligence

In this scenario, superintelligence develops and diffuses at an unprecedented speed. Governance requires:

• Executive Coordination: Close cooperation with the executive branches and safety institutes of multiple countries.
• Targeted Mitigation: Specific focus on preventing AI applications in bioterrorism and managing the implications of self-improving AI.
• Accountability: Maintaining accountability to public institutions through potentially non-traditional methods.

The AI Resilience Ecosystem

OpenAI proposes the creation of an "AI resilience ecosystem" analogous to the field of cybersecurity. Just as cybersecurity established encryption protocols, monitoring systems, and emergency response teams to make the internet trustworthy, a similar ecosystem of software and standards is needed for AI. OpenAI suggests that national governments should use industrial policy to encourage the development of this ecosystem.

Individual Empowerment and Access

OpenAI views advanced AI as a future foundational utility, comparable to electricity, clean water, or food. The organization advocates for making these tools widely available to empower individuals to achieve their goals, provided they are used within broad societal bounds.