Viewpoint
AI safety now matters before systems are deployed
AI safety now matters before systems are deployed
Bostrom says powerful models can create safety concerns during training and evaluation, so safeguards must apply before public deployment as well as afterward.
- Interview
- Worries About AI Existential Risk Just Became More Concrete - Nick Bostrom and Alex Kantrowitz
- Speaker
- Nick Bostrom
- Topic
- AI Alignment
- Source timestamp
- 5:53
More from this interview
- More capable agents expose additional alignment failure modes
- Open models may soon enable serious destructive uses
- Biological safeguards should focus on physical bottlenecks
- Misuse and alignment are distinct categories of AI risk
- Alignment difficulty may matter more than collective effort
- Imperfectly aligned early systems could improve later alignment
- Biological threats may be harder to defend against
- Recent AI progress has not substantially changed his risk estimate
- AI research automation could trigger an intelligence explosion
- A fast AI takeoff remains plausible but uncertain
- A pause would be most useful near a critical transition
- A long or uneven pause could create new risks
- AI's potential benefits justify accepting some risk
- Current AI systems have not yet reached full AGI
- AI may become superhuman in key fields before full AGI
- Conversational AI may make alignment easier to study
- Current AI consciousness is plausible enough to merit consideration
- Multiple indicators support taking AI consciousness seriously
- Digital minds may deserve moral status without humanlike consciousness
- Digital-mind ethics is a major AI challenge
- Digital-mind ethics cannot simply copy human ethics
- Humans should begin cultivating respectful treatment of AI
- Trust between humans and AI could reduce future risks