Towards Alignment Auditing as a Numbers-Go-Up Science
Alignment auditing is a crucial aspect of artificial intelligence research, aimed at ensuring that AI systems are designed and trained to align with human values. However, some researchers argue that alignment auditing should be approached differently than other fields of science.
I strongly disagree with this notion. Alignment auditing can benefit from the principles of numbers-go-up science, which emphasizes developing novel models or ontologies that can't be quantitatively compared to previous ones. This approach allows for a more nuanced understanding of alignment and its relationship to human values.
The idea that alignment research should be like architecture design, performance optimization, or RL algorithms may seem appealing, but it steers people away from the thing that alignment research should be contributing: developing deep explanations of artificial cognition. Alignment auditing is not just about optimizing for performance metrics; it's about understanding how AI systems interact with their environment and making them more transparent and accountable.
The notion that "bad metrics are worse than no metrics" is an oxymoron. Metrics can provide valuable insights, but they should be used judiciously to avoid misleading or confusing researchers. The key concern should be the quality of the problem specification, not the use of quantitative measures themselves.
I think alignment auditing benefits from being grounded in scientific understanding. This requires developing novel models and ontologies that can't be quantitatively compared to previous ones. It's also important to specify problems in advance to guide research in the field. However, this should be done with caution, as paradigm shifts can change the set of important problems.
Some argue that specifying concrete problems is a bad thing, but I disagree. Well-specified problems can provide valuable guidance for researchers and help identify areas where more work is needed. The key is to ensure that these problems are robust to paradigm shifts and don't lead to dead ends in the research process.
Addressing Concerns
I understand some concerns about alignment auditing, particularly with regards to the difficulty of specifying problems and the potential for bad metrics to confuse researchers. However, I believe that these challenges can be addressed through careful problem specification, rigorous evaluation methods, and a focus on developing novel models and ontologies.
Moreover, I think it's essential to recognize that alignment auditing is not just about optimizing performance metrics; it's about understanding how AI systems interact with their environment and making them more transparent and accountable. By focusing on scientific understanding and well-specified problems, we can ensure that alignment auditing contributes meaningfully to the development of more advanced AI systems.
Conclusion
In conclusion, I believe that alignment auditing should be approached as a numbers-go-up science. This approach allows for a more nuanced understanding of alignment and its relationship to human values, while also ensuring that researchers develop novel models and ontologies that can't be quantitatively compared to previous ones.
By focusing on scientific understanding, well-specified problems, and the development of novel models and ontologies, we can ensure that alignment auditing contributes meaningfully to the development of more advanced AI systems. This approach requires careful problem specification, rigorous evaluation methods, and a commitment to developing a deeper understanding of artificial cognition.