We are likely to have them
- A decentralized entity trains a frontier model
- Theft, or state-level theft
- Could just be deployed by some lab
From there, [[Uncensored versions of open weights releases quickly proliferate]]
Pros for releasing open weights
- Accelerates [[Technical AI safety]] research
- Accelerates using [[AI for good]]
- Cheaper inference
- Can be customized, may enable new use-cases
- Reduces power concentration
- May lead to a [[Focusing events for AI]] (for better or worse)
Risks
- Biosecurity
- Cyber
- NCII
- Scalable spear-phishing
- Disinformation
Ultimately: What degree of capabilities is too dangerous to release with open weights. Depends largely on our [[Societal resilience]] and the destructive capacity of those capabilities.