Key Cruxes
Overview
Section titled “Overview”Cruxes are key uncertainties where different beliefs lead to substantially different conclusions about AI safety priorities. Identifying and tracking cruxes helps clarify what evidence would be most valuable and where reasonable people disagree.
Crux Categories
Section titled “Crux Categories”Uncertainties about unintended AI failures:
- Will advanced AI systems develop misaligned goals?
- Can we detect deceptive alignment before deployment?
- How likely is mesa-optimization in large models?
- Will AI systems seek power instrumentally?
Uncertainties about deliberate harmful use:
- How much do AI capabilities “uplift” bioweapon development?
- Will autonomous weapons proliferate faster than defenses?
- Can AI-generated disinformation be reliably detected?
Uncertainties about systemic dynamics:
- Will racing dynamics dominate lab behavior?
- Is value lock-in a serious concern on realistic timelines?
- How concentrated will AI capabilities become?
Uncertainties about knowledge and truth:
- Will AI accelerate or undermine human epistemic capacity?
- Can authentication systems keep pace with generative AI?
- Will expertise atrophy be reversible?
Uncertainties about intervention effectiveness:
- Is alignment research on track to succeed?
- Can governance keep pace with capability development?
- Will international coordination be achievable?
Using Cruxes
Section titled “Using Cruxes”Cruxes help:
- Prioritize research - Focus on resolving highest-value uncertainties
- Bridge disagreements - Identify where people actually differ
- Track progress - Monitor how uncertainties resolve over time
- Inform forecasting - Structure predictions around key variables