AI summary by TwinMind
The Dark Side of Competition in AI | Liv Boeree | TED
TED
12m 3s
Competition and Human Nature
- Competition fundamental part of human nature
- Speaker: professional poker player for 10 years, witnessed good/bad manifestations
- Healthy competition drives incredible feats in sports and innovation
- Car companies competing for safest cars
- Most efficient solar panels
- Creates win-win outcomes where everyone benefits long-term
- Unhealthy competition creates lose-lose outcomes where everyone worse off
AI Beauty Filters Example
- Impressive technology can salvage almost any picture
- Can make Angelina Jolie and Margot Robbie "more beautiful"
- Handy for influencers - transform into Hollywood versions at click of button
- Problems:
- Train users to hate natural face
- Growing evidence of body dysmorphia, especially young people
- Endemic to social media due to game nature
- Platform incentives: hotter pictures = hijacked limbic systems = more scrolling = more ad revenue
- User incentives: hotter pictures = more followers
- Trap: hard to go back once started, no competitive advantage since everyone using them
- Influencers stuck using with downsides, little upside
News Media Race to Bottom
- Expected internet competition would create positive spiral toward nuanced, impartial, accurate journalism
- Instead: race to bottom of clickbait and polarization
- Even respectable papers leaning into low-brow partisan tactics
- Root cause: crappy incentives
- News now interactive (sharing, commenting)
- Headlines triggering fear/anger far more likely to go viral than neutral/positive
- News editors in similar trap as influencers
- More competitors use clickbait tactics, more they must too
- Otherwise stories get lost in noise
- Results: media gets less public trust, harder to discern truth from falsehood
- Big problem for democracy
Broader Pattern of Competition Gone Wrong
- Driving force behind major issues:
- Plastic pollution
- Deforestation
- Antibiotic overuse in farming
- Arms races
- Greenhouse gas emissions
- All result from crappy incentives of poorly designed games
- Push players (people, companies, governments) into strategies deferring costs/harms to future
- Most actors don't want to be doing this
- Packaging companies don't want oceans filled with plastic
- Farmers don't want antibiotic resistance
- Stuck in dilemma: "If I don't use this tactic, I'll get outcompeted by others who do"
Root Cause Analysis
- Not capitalism (can cause problems but also solve them, fantastic in general)
- Something deeper: misaligned incentives of Game Theory itself
- Speaker retired from poker to understand mechanism better
- Takes many forms, many names (abstract and clunky terms)
- Need visceral way of recognizing given how insidious and connected problems are
Moloch Metaphor
- Biblical story from Canaanite days
- Cult wanted money and power badly enough to sacrifice literal children
- Burned children alive in effigy of god Moloch, believing would reward ultimate sacrifice
- Apt metaphor: getting lost in winning immediate game, losing sight of bigger picture
- Sacrificing too much in pursuit of victory
- Examples of modern sacrifices:
- Influencers: happiness for likes
- News editors: integrity for clicks
- Polluters: biosphere for profit
- Short-term game incentives push players to sacrifice more of future
- Traps in death spiral where all lose in end
- Moloch's trap = mechanism of unhealthy competition
AI Industry Moloch Trap
- Race heating up between companies over:
- Most compute
- Biggest funding rounds
- Top talent
- More companies enter race = greater pressure to go fast as possible
- Sacrifice other important stuff like safety testing
- Classic Moloch trap hallmarks
- CEO dilemma: believe team best to safely build powerful AI
- Go too slowly = risk less cautious teams getting there first
- Pushes to be more reckless
- Many experts/researchers (within companies and independent) warning about extreme risks of rushed AI
- Approach "absolutely mad"
- AI companies beholden to satisfying investors (short-term incentive)
- Will inevitably conflict with benevolent mission over time
- AI/AGI = bigger paradigm shift than agricultural/industrial revolutions
- Moment deserving reverence and reflection, not corporate rat race for daily active users
- Don't know right trade-off between acceleration and safety
- Will never find out if let Moloch dictate
Solutions and Coordination Examples
- Have managed to escape some Moloch traps before:
- Saved ozone layer from CFCs with Montreal Protocol
- Reduced nuclear weapons on Earth by 80% with Strategic Arms Reduction Treaty 1991
- Smart regulation may help with AI
- Players within game have most influence
- Need AI leaders to show awareness of:
- Risks their technologies pose
- Destructive nature of current incentives
- As technological capabilities reach power of gods, need godlike wisdom to wield them
- Discouraging when major company CEO says "I want people to know we made our competitor dance"
- Not the mindset needed
- Need leaders willing to flip Moloch's playbook
- Sacrifice individual chance of winning for good of whole
Current AI Lab Initiatives
- Three leading labs showing some signs:
- Anthropic: responsible scaling policy - only increase capabilities once security criteria met
- OpenAI: pledged 20% of compute purely to research
- DeepMind: decade-long focus science ahead of commerce (AlphaFold given free to science community)
- Steps in right direction but nowhere close to enough
- Most currently just words, not proven actions
- Need clear way to turn AI race into definitive race to top
- Companies could compete over:
- Best security criteria
- Most compute dedicated to alignment
- Would truly flip middle finger to Moloch
Stakes and Conclusion
- Competition amazing tool if wielded wisely
- Stakes astronomical:
- Get AI/AGI wrong = unimaginable catastrophe
- Get it right = path out of many Moloch traps
- As things get crazier in coming years, important to remember:
- Moloch is real enemy, not individual CEOs/companies
- Not one another
- "Don't hate the players, change the game"
