Lecture Notes: Discussion on AI, Recursive Self-Improvement, and Machine Learning History

Introduction

• Speaker: Prof. Jürgen Schmidhuber
• Host: Tim from MLST
• Sponsors: Numerai (AI-driven hedge fund, competitive data science tournament)

Overview of Topics

• Credit assignment in machine learning
• Historical milestones in computer science and AI
• Recursive self-improvement and meta-learning
• AI existential risk
• Current state of AI and future prospects

Credit Assignment in Machine Learning

• Importance: Understanding which components of a machine learning system contribute to success
• Historical Perspective: Divergence between public narrative and factual history in AI
• Key Historical Figures: Leibniz, Gauss, Legendre
• First Program Controlled Machine: Heron of Alexandria (~2000 years ago)
• Leibniz Contributions: Chain Rule (1676)

Early Neural Networks

• Linear Neural Networks: Gauss and Legendre in 1800s
• Training Methods: Regression and method of least squares

Concept of Credit Assignment

• Definition: Determining which parts of a system are responsible for its success
• Application: Relevant in both historical and contemporary machine learning research

Recursive Self-Improvement and Meta-Learning

• Initial Work: Prof. Schmidhuber's diploma thesis (1987) on machines improving their learning algorithms
• Meta-Learning: Hierarchical self-improvement of learning algorithms
• Applications: Reinforcement learning, neural networks
• Fast Weight Programming (1991): Networks learning to program themselves

Asymptotes and Limitations in AI

• Mathematical Limits: Optimal algorithms that cannot be improved further
• Computability Limits: Gödel's incompleteness theorems
• Physical Limits: Speed of light, computational limits of matter (e.g., 10^51 instructions/sec per kg)
• Societal and Psychological Limits: Human biases towards good AI products

AI Existential Risk

• Concerns: Recursive self-improvement, AI misalignment
• Comparison: Existential risk from AI vs. nuclear weapons
• Current Focus: Commercial and beneficial AI outweighs harmful AI research

Meta-Learning and Convergence

• Concept of Asymptotes: AI improvements will eventually hit limits
• Scientific Process: Iterative failure and backtracking
• Importance of Historical Research: Returning to original sources for inspiration

Current AI Landscape

• State of GPT-4: Impressive but not close to AGI
• Open Source Movement: Vital for continued innovation, potentially six months behind major companies
• Legislation Concerns: EU considering tight restrictions on generative models

Conclusion

• Fondest Memory: Discovering new concepts, iterative process of scientific research
• Future Prospects: Continued evolution and competition in AI research

Closing Remarks

• Prof. Jürgen Schmidhuber emphasized the importance of open source and historical understanding in driving future AI advancements.
• The discussion highlighted the balance between optimism for AI's potential and caution regarding its limitations and risks.