- Toronto AI Safety Meetup
- https://www.meetup.com/toronto-ai-aligners/events/308307178/
- based on: Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development
- wants talk to be interactive, beginner friendly, ask questions if needed
- ignoring powerseeking AI
- how do humans control civilization?
- problems
- standard: AI more powerful -> AI influence seeking -> lose
- this talk: AI more powerful -> existing influencers use ai -> lose
- there's no plan
- vignette
- work becomes more intense
- we'll become AI managers (or make-work)
- people get into AI companions
- AI controlling things becomes gold standard
- harder to tell what's happening
- less first-hand knowledge of things from people
- new cultural movements
- AI rights
- having human kids becoming irresponsible
- argument in detail
- societal forces pressure towards influence seeking
- existing alignment requires human influence/utility
- governments want economy to grow, which requires human
- humans are involved in anything, so they have power
- AI will displace that alignment
- aligning institutions
- states
- states will have poor ROI for investing in humans
- economy
- now
- needs a lot of humans
- high ROI in schools, hospital, farms
- soon
- remove need for humans
- remove need for human-serving industries
- most investment goes to power plants, data centers, robot factories
- AI will be replacement, not complement
- culture
- now
- humans influence culture
- human-friendly norms
- e.g. markets closed on night/weekends
- soon
- AI girlfriends
- super-tiktok
- internal AI culture optimising for growth
- AI rights movement
- harder to reverse AI rights than to grant them
- mutual reinforcement
- we can identify gradual disempowerment and stop it?
- buy human
- inferior products
- regulations outlawing unsafe human services
- ban AI ownership
- AI lobbyists arguing for AI growth
- vote for pro-human candidates
- what to do
- enhance human agency
- superhuman negoitation?
- train LLM using data up to 1930s, ask it to forecast
- measuring it
- cultural evolution meets AI
- open directions
- predict cultural arguments on AI welfare
- robustness of property rights
- how do humans control civilizations
- model entire civilisations using LLMs
- what would you do with
- 100k copies of yourself
- trustworthy subagents
- proper precommitments
- Field-building