Theoretical Computer Science and AI Safety

Background

• Lecturer with a career in Quantum Computing, recently at OpenAI.
• Asked to explore theoretical computer science's role in preventing AI threats.
• Unsure about prevention but pondering implications of successful AI progression.

Recent Developments in AI

• AI replicates science fiction devices (e.g., Star Trek’s computer).
• Current AI successes were unforeseen 5 years ago.
• Many are skeptical about AI’s rapid advancements.

Historical Context

• Core AI ideas (neural networks, backpropagation, gradient descent) have been long known.
• Skepticism over scaling techniques that historically didn’t perform well.
• Ray Kurzweil’s predictions about compute scaling and AI's progression were initially dismissed.
• Real-world advancements have proven skeptics wrong.

AI Capabilities and Challenges

• AI’s advancing capabilities prompt questions about future roles of humans.
• Potential of AI solving advanced problems like the Riemann Hypothesis through collaboration.
• Discussions have reached political levels (e.g., White House, Congressional hearings).
• Speculation ranges from AI irrelevance due to technical breakthroughs to potential dangers (e.g., AI becoming superior to humans).

Decision Tree of AI Outcomes

• AI progress could fizzle out due to various limitations (e.g., data shortage, compute costs).
• If progress continues, AI might achieve parity with human capability in 10-20 years.
• Impacts on civilization and questions about human relevance in an AI-dominant world.

Current Limitations and Public Perception

• Public and academic resistance to speculating on AI’s long-term impact.
• Arguments about AI’s current role as an advanced but limited tool (e.g., stochastic parrots, auto-completes).
• Moving goalposts in AI achievements (e.g., from chess to Go, future challenges like math competitions).
• Thesis: AI could eventually master any task with clear metrics and sufficient examples.

AI Impact on Education and Careers

• Projects like watermarking outputs of AI tools (e.g., GPT) for ethical purposes (preventing cheating, misinformation).
• Debates about the relevance of traditional education and skill development in the AI era.
• AI’s role in creative fields and questions about human uniqueness and influence.

Philosophical Considerations

• AI’s ability to produce vast quantities of artistic material devalues individual artworks (AI abundance paradox).
• Humans’ uniqueness tied to our limitations and individuality.
• Theoretical debates about human cognition and potential for digital replication (e.g., brain backups, identity).
• Quantum mechanics and the challenges of perfect human replication.

AI Safety and Ethical Proposals

• Integrating ethical values in AI development to respect human uniqueness and creativity.
• Balancing innovative ideas for AI safety with practical concerns and feasibility.

Conclusion

• Encourages contemplation of AI’s broader implications on society, creativity, and human identity.
• Opens up discussions about long-term strategies for managing AI advancements ethically.