AI ALIGNMENT FORUM
AF

1011
Paradigm-Building for AGI Safety Research

Paradigm-Building for AGI Safety Research

Jan 31, 2022 by Cameron Berg

This sequence introduces and defends an end-to-end theoretical framework for AGI safety research. Beginning with foundational principles in effective altruism, the framework moves hierarchically: it starts by formalizing the problem that AGI safety research intends to solve and proceeds to specify the progression of field-level questions that seem like they must be answered to solve the problem. This sequence is intended for new and seasoned researchers alike—my hope is that this report will help organize and streamline the individual efforts of safety researchers and facilitate crisper field-level communication, collaboration, and debate.      

10Paradigm-building: Introduction
Cameron Berg
4y
0
4Paradigm-building from first principles: Effective altruism, AGI, and alignment
Cameron Berg
4y
0
0Paradigm-building: The hierarchical question framework
Cameron Berg
4y
0
0Question 1: Predicted architecture of AGI learning algorithm(s)
Cameron Berg
4y
0
0Question 2: Predicted bad outcomes of AGI learning architecture
Cameron Berg
4y
0
0Question 3: Control proposals for minimizing bad outcomes
Cameron Berg
4y
0
0Question 4: Implementing the control proposals
Cameron Berg
4y
0
3Question 5: The timeline hyperparameter
Cameron Berg
4y
0
2Paradigm-building: Conclusion and practical takeaways
Cameron Berg
4y
0