Explore the critical aspects of AI alignment in this comprehensive video lecture on AGI Alignment Experiments. Delve into the comparison between INSTRUCT and Foundation and Agent Models, examining the importance of interpretability in AI systems. Discover why alignment should be viewed as a holistic system rather than a single model, and learn about the significance of testing intelligence in complex systems. Investigate self-stabilizing systems, alignment experiments, and the potential of existing technology in alignment research. Analyze the challenges of aligning superintelligent AI and the dangers associated with nanotechnology and genetic alteration. Examine the risks of AI systems without hard goals and the instability of simple programming constructs. Gain insights into the process of creating advanced AI capable of writing novels, and compare the stability of instruct models versus agent models in the context of AI alignment.
Overview
Syllabus
- The importance of interpretability in AI alignment
- Alignment as a system, not a single model
- The importance of testing intelligence in complex systems
- The Right Research
- Self-stabilizing systems
- Alignment experiment 2
- The second agent model
- Alignment research with existing technology
- Alignment research on superintelligence
- The dangers of nanotechnology and genetic alteration
- The dangers of an AI with no hard goals
- The instability of a simple for loop
- The process of creating a machine that can write novels
- The stability of instruct models
- The stability of agent models
Taught by
David Shapiro ~ AI
