AI summary by TwinMind
The Transformative Potential of AGI — and When It Might Arrive | Shane Legg and Chris Anderson | TED
TED
16m 11s
Early Life and AI Interest Origins
- Got first home computer on 10th birthday (pre-internet era)
- Started programming, discovered ability to create worlds/universes in computer
- Spark of creativity captivated him, became seeds of AI interest
- Dyslexic as child, nearly held back a year at age 10
- IQ test revealed exceptionally high intelligence, not limited capability
- Experience created early doubt in authority, interest in intelligence nature
Coining "Artificial General Intelligence" (AGI) Term
- 2001: Approached by Ben Goertzel writing book about general AI systems
- Suggested term "artificial general intelligence" for book title
- Term popularized online by Goertzel and others
- Later discovered Mike Garrod first used term in 1997 security nanotech journal paper
- Garrod's usage aligned with their meaning
Early AGI Prediction and Reasoning
- 2001: Read Ray Kurzweil's "Age of Spiritual Machines"
- Concluded computation would grow exponentially for decades
- Data in world would grow exponentially for decades
- Highly scalable algorithms harnessing data/computation would be valuable
- Mid-2020s: possible to train AI systems on more data than human lifetime experience
- 2009 blog: predicted 50% chance AGI by 2028
- Still maintains this prediction today
AGI Definition Evolution
- Initial definition: systems far more general than narrow AI (Go, chess)
- Current definition: system doing all cognitive tasks people can do, possibly more
- At minimum: cognitive tasks people typically perform
DeepMind Founding
- Went to London's Gatsby Unit (theoretical neuroscience/machine learning)
- Studied brain-machine learning relationships
- Met Demis Hassabis (same postdoc supervisor)
- 2009: Hassabis convinced him time was right for company
- Third cofounder: Mustafa Suleyman (Hassabis friend)
- 2010 business plan front cover: "Build the world's first artificial general intelligence"
- Google acquired DeepMind during journey
Key Breakthrough Moments
- Atari games: algorithm learning multiple games without specific programming
- AlphaGo: major exciting milestone
- Language model scaling: most imagination-capturing development
- Transformer model invented at Google
- Teams building large transformer language models
ChatGPT Surprise
- Didn't expect text-only training to achieve such capability
- Thought multimodality needed (images, sound, video)
- Vast text quantities compensated for limitations
- Language models still weak in non-text-expressible areas
- Expects expansion into video/images/sound to overcome limitations
AGI Impact Potential
- AlphaFold example: 30 world-class scientists, 3 years development
- With AGI: handful of scientists, 1 year for similar breakthroughs
- Regular AlphaFold-level developments globally
- Scientific challenges from decades/centuries could tumble rapidly
- Potential golden age: human intelligence + machine intelligence
- Solve previously intractable problems
Risk Assessment Philosophy
- Rejects both extreme optimism and extreme pessimism
- Profound transition comparable to human intelligence arrival
- Incredibly deep transition with unknown implications
- Must remain open-minded about outcomes
- Optimism based on understanding systems enables safety
- Airplane safety analogy: need understanding for safety
- Even small probability of bad outcomes requires serious attention
Potential Negative Scenarios
- Superhuman intelligence systems create unpredictable risks
- Engineered pathogens possibility
- Democracy destabilization through propaganda
- Various societal destabilization methods
- Historical precedents for societal collapse
Regulation and Control Challenges
- Would use "magic wand" to slow development if possible
- ~10 companies globally can develop cutting-edge models
- National intelligence services have secret projects
- Dozens more companies one generation behind
- Intelligence incredibly valuable/useful
- No realistic plan to stop the process
- Regulation needed like other powerful technologies
- Nothing special about AI requiring exemption from regulation
Research Priorities
- Heavy energy focused on capabilities development (valuable)
- Need more energy on actual science/understanding
- Understanding how networks work, what they do/don't do
- Identifying likely failure points
- Finding ways to make AGI systems profoundly ethical and safe
- Believes ethical/safe AGI possible
- Need to mobilize more minds for solutions
- Goal: powerful machine intelligence that's ethical, safe, enhances humanity
- Vision: new golden period for humanity
