AI summary by TwinMind
How AI Is Unlocking the Secrets of Nature and the Universe | Demis Hassabis | TED
TED
25m 34s
Introduction and Motivation for AI
- Demis Hassabis wanted to understand big philosophical/physics questions about reality and consciousness
- Realized physics hadn't made much progress in fundamental laws in 20-30 years
- Decided to build AI as ultimate tool to help answer these questions
- AI could also help understand the brain and consciousness itself
- AI can find patterns in massive datasets that human minds can't comprehend
Early Gaming Background and Chess
- Hassabis captained England Under 11 chess team at age 9
- Played in Four Nations tournament against France, Scotland, Wales
- Used early chess computers in mid-1980s for training
- Fascinated by programmed "lump of plastic" playing high-standard chess
- This sparked interest in thinking about thinking and brain processes
DeepMind's Gaming Approach
- Started with classic 1970s Atari games in 2010 as testing ground
- Games provided convenient, fast testing environment for algorithms
- Could choose progressively harder games as systems improved
- First breakthrough: Breakout game in 2012-2013
- System learned to send ball around back of wall
- First time machine surprised the team
- Coined term "deep reinforcement learning"
- Learned directly from 30,000 raw pixels
- Only instruction: maximize score
- Had to figure out controls and objectives from first principles
AlphaGo Achievement
- Go is more complex than chess, played in Asia instead of chess
- Brute force algorithms that worked for chess impossible for Go
- More pattern-based game than chess
- Deep Blue beat Garry Kasparov in 1990s, took 20 more years for Go
- AlphaGo beat world champion in 2016
- Learned by playing millions of games against itself
- Invented new strategies never seen in 2,000+ years of Go history
- Demonstrated something "very interesting" had to happen to beat world champion
AlphaZero Development
- AlphaGo started with human games from internet as knowledge base
- AlphaZero started from literally random play (zero prior knowledge)
- More general system - could play any two-player game
- Went from random to world-champion level in 24 hours
- Chess achievement: 9 hours to surpass decades of expert chess programming
- Hassabis could barely beat it by lunchtime, by dinner it was greatest chess entity ever
- Demonstrated potential for scientific applications beyond games
Protein Folding Problem
- Always aimed to use AI for scientific discovery acceleration
- Had protein folding problem in mind since Cambridge undergrad
- 50-year grand challenge in biology
- Proteins are building blocks of life, everything depends on them
- Amino acid sequence (letters) determines 3D structure (shape)
- 3D structure determines function in body
- Challenge: predict 3D structure directly from amino acid sequence
- 150,000 protein structures determined over 40+ years using X-ray crystallography
- Rule of thumb: one PhD student, 4-5 years per structure
- 200 million proteins known to nature
AlphaFold Results
- Started project straight after AlphaGo success
- Folded all 200 million known proteins in one year
- Saved "billion years of PhD time"
- Accuracy: within width of an atom on average
- Sufficient accuracy for biologists, drug design, disease understanding
- Open-sourced entire database with European Bioinformatics Institute
- Over 1.5 million biologists worldwide using AlphaFold
- Estimated almost every biologist and pharma company now uses it
Isomorphic Labs
- New company spinning out of Google
- Extends AlphaFold work into chemistry space
- Designs chemical compounds to bind exactly to right protein spots
- Ensures no binding to other body parts (no side effects/toxicity)
- Building sister models to AlphaFold for chemistry predictions
- Goal: reduce drug discovery from years to months
Competitive Dynamics and Moloch Trap
- OpenAI/Microsoft released ChatGPT, world went "crazy"
- Satya Nadella: "wanted to make Google dance" as 800-pound gorilla
- Google responded dramatically, Hassabis took over Google AI effort
- Products rushed out, including Gemini (part amazing, part embarrassing)
- Meta rushed out open-source AI version
- Google Research invented Transformers architecture 5-6 years ago
- Had many large language models internally before ChatGPT
- ChatGPT demonstrated public readiness for these systems despite flaws
- Tens of millions found value despite hallucinations and limitations
Future Compute Investment
- Microsoft/OpenAI reportedly investing $100 billion in supercomputer
- Requires 5 gigawatts energy (equivalent to New York City)
- Google investing "more than that over time"
- Teamed with Google in 2014 knowing AGI would need massive compute
- Google has "most computers"
AGI Safety and Collaboration
- Must avoid dangerous competitive dynamics as approaching AGI
- Need more collaboration as getting closer to AGI
- Scientists in different labs know each other well, communicate regularly
- Technology still relatively nascent, current competition "probably OK"
- Society needs to think about AGI architectures as getting closer
- Many ways to build safe, robust, reliable, understandable architectures
- Also ways to build unsafe/risky architectures
- Sees "bottleneck" humanity must navigate: building safe architectures first
- After safe foundation, can have flourishing of different system types
- Need mathematical or practical guarantees around system behavior
Role of Government and Society
- Government, civil society, academia all have critical roles
- Need cooperation to prevent "runaway race dynamic"
- Industry labs must collaborate with all parts of society
- Essential to shape approach as getting closer to AGI
Tree of Knowledge Vision
- Envisions "tree of all knowledge" representing all possible knowledge
- Current civilization knows small subset
- AI as tool to explore entire tree
- Focus on "root node problems" like protein folding
- Solving these unlocks entire new branches of discovery
- Potential for "radical abundance era"
- Goals: cure all diseases, spread consciousness to stars, maximum human flourishing
Ultimate Scientific Goals
- Once AGI built, wants to understand fundamental nature of reality
- Conduct experiments at Planck scale (smallest theoretical scale)
- Planck scale as "resolution of reality"
- Scientific method as greatest human invention
- Enlightenment and scientific discovery built modern civilization
- Best technique for understanding universe's enormity
- Acknowledges some things might be unknowable
