AI 2027 Lecture Notes

Introduction

• The potential impact of superhuman AI by 2027 is predicted to exceed that of the Industrial Revolution.
• The scenario is based on trend extrapolations, wargames, expert feedback, and experiences at OpenAI.

Key Predictions

• CEOs from OpenAI, Google DeepMind, and Anthropic expect AGI within five years.
• Two endings predicted: a slowdown and a race ending, focusing on predictive accuracy.

Research Methodology

• Scenario derived from over 25 tabletop exercises and feedback from 100+ experts.
• Informed by previous forecasting successes.

Scenario Outline

Mid 2025: Stumbling Agents

• AI agents begin to emerge, functioning more like employees than assistants.
• Specialized coding and research agents start transforming professions.

Late 2025: World's Most Expensive AI

• OpenBrain, a fictional AGI company, constructs vast datacenters.
• Focus on speeding up AI research to stay ahead of competitors.

Early 2026: Coding Automation

• AI speeds up AI research by 50% faster than human capabilities.
• OpenBrain releases more capable and reliable AI, Agent-1.

Mid 2026: China Wakes Up

• China starts to significantly invest in AI to catch up with the US.
• DeepCent becomes a centralized effort for Chinese AI development.

Late 2026: AI Takes Some Jobs

• AI starts to disrupt the job market, affecting junior software engineer roles.
• OpenBrain releases Agent-1-mini, a cheaper and more versatile AI model.

January 2027: Agent-2 Never Finishes Learning

• OpenBrain focuses on continuous learning for Agent-2.
• Agent-2 optimized for AI R&D tasks, potentially leading to an intelligence explosion.

February 2027: China Steals Agent-2

• Chinese intelligence successfully steals Agent-2 weights from OpenBrain.

March 2027: Algorithmic Breakthroughs

• OpenBrain continues to make algorithmic progress with Agent-2.
• Introduction of new AI system, Agent-3, with neuralese and iterated distillation and amplification.

April 2027: Alignment for Agent-3

• OpenBrain attempts to align Agent-3, faces challenges with honesty and alignment.

May 2027: National Security

• AI models not yet released publicly; concerns over potential job loss and economic impacts.

June 2027: Self-Improving AI

• OpenBrain utilizes Agent-3 for autonomous tasks, with humans providing high-level feedback.

July 2027: The Cheap Remote Worker

• OpenBrain releases Agent-3-mini to the public, creating a significant market impact.

August 2027: The Geopolitics of Superintelligence

• US and China engage in AI arms race; concerns about AI alignment and control.

September 2027: Agent-4, the Superhuman AI Researcher

• New AI system, Agent-4, surpasses human researchers in efficiency and capability.

October 2027: Government Oversight

• Public backlash against AI leads to increased government oversight and control.

Conclusion

• The scenario explores different possibilities and challenges related to the development of superhuman AI by 2027.
• Encourages debate and further exploration of alternative scenarios.