Applying Obliquity in AI Development
Obliquity suggests that the best outcomes in complex domains are often achieved indirectly
- Instead of optimizing for accuracy alone, we might prioritize interpretability, human alignment, or collaborative exploration—and find that accuracy improves as a result
- Rather than chasing AGI directly, we might focus on solving meaningful human problems, and discover emergent capabilities along the way
“In AI research, sometimes the most effective path to a desired outcome isn’t a direct one… approaching goals indirectly with a broader perspective can yield surprisingly successful results”
| AI Development Phase | Traditional Approach | Oblique Strategy | Why It Matters |
|---|---|---|---|
| Problem Definition | Define narrow task (e.g., classify images) | Explore broader human context (e.g., assist visual understanding) | Leads to more meaningful, generalizable solutions |
| Model Objectives | Maximize benchmark performance | Optimize for values like trust, usability, or adaptability | Encourages resilience and long-term relevance |
| Data Strategy | Gather large labeled datasets | Curate diverse, context-rich data with human nuance (narrative data) | Reduces bias, improves alignment |
| Evaluation Metrics | Use precision, recall, F1 score | Include human feedback, interpretability, and social impact | Captures real-world utility beyond numbers |
| Deployment | Scale fast, automate decisions | Embed human-in-the-loop, allow for emergent use cases | Manage unintended negative consequences (see WRAS) |
| Research Culture | Publish for citations and prestige | Collaborate across disciplines, pursue curiosity | Fosters serendipity and deeper breakthroughs |