AI summary by TwinMind
George Hotz: Hacking the Simulation & Learning to Drive with Neural Nets | Lex Fridman Podcast #132
Lex Fridman
3h 8m 46s
Introduction and Background
- George Hotz (geohot) second appearance on podcast
- Founder of Comma AI - autonomous/semi-autonomous vehicle technology company
- Positioning as "Android to Tesla's iOS" for self-driving cars
- Comma Two device costs $1000, works on supported cars for lane keeping
- Includes driver sensing with eye tracking capability
Intelligent Life and Fermi Paradox
- Hotz believes intelligent civilizations existed elsewhere but destroyed themselves
- IQ required to destroy world falls by one point every year (technology democratization)
- Cryptography as example where defense wins vs nuclear weapons where offense wins
- Civilizations either blow themselves up or enter stasis (wireheading)
- 70,000 light years across galaxy - not big for advanced technology
- Von Neumann probes could colonize galaxy in million years with sublight travel
UFOs and Conspiracy Theories
- Believes most news is psyop, including UFO disclosures
- Bob Lazar story credible but likely MK Ultra-type experiment
- Conspiracy theories invented by CIA in 1960s to discredit truth
- Examples: Chinese locking up coronavirus doctors, money no longer backed by gold
- Information control shifted from censorship to flooding with disinformation
Programming and Alien Intelligence
- Believes alien programming languages wouldn't be fundamentally different
- DNA is CAD model, not programming language - no serial execution
- Simulating biological systems computationally intensive at molecular level
- Biology tools currently primitive - need better automation and microfluidics
Simulation Theory and Hacking Reality
- South by Southwest talk about hacking simulation was thought experiment
- Goal: humble humans about their place in universe
- Real message: stop playing fake human games, play real game with nature
- Immortality as goal - want to choose own death vs nature deciding
- Backup copies of self to prevent irreversible changes
- Simulation detection would require finding exploitable bugs/memory leaks
Cryptocurrency and Smart Contracts
- Extremely bullish on crypto long-term
- Nakamoto consensus algorithm one of greatest 21st century innovations
- Smart contracts replace lawyers - python 10,000x better than lawyers
- Code is law concept - contracts executed by cryptography vs legal interpretation
- Lawyers cost $1000/hour, unreliable; python costs 10 cents/hour, reliable
- Forks allow competing implementations (like forking America with different presidents)
Learning and Knowledge Acquisition
- No systematic approach to learning - just does things for 20 years
- Three transformational days: discovering Yudkowsky, Hutter Prize, Unqualified Reservations
- Drifts through life hoping for another transformational day
- Advice: just do things a lot, technology always wins, don't lie
Comma AI Business Model and Products
- Mission: "Solve self-driving cars while delivering shippable intermediaries"
- Revenue ensures honest assessment of progress vs arbitrary milestones
- Comma Two: $1200 total cost including harness
- 70% of buyers are daily active users - extremely high retention
- Supports 91 car models, targeting 100 by year end
- 2,000 daily users, 2,500 weekly, 3,000+ monthly
Technical Approach: End-to-End vs Segmented
- Comma AI uses end-to-end learning approach
- Tesla uses multi-task learning with hundreds of perception tasks
- Analogy: hiring "bishop guy" vs "rook guy" for chess engine company
- End-to-end will eventually win - feature engineering always loses
- Current system: 100 miles between unplanned disengagements (up from 10 miles)
- Need 100,000 miles for Level 5 - requires 3 more 10x improvements
Driver Monitoring and Safety
- Scene-adaptive driver monitoring policy
- Infrared camera for attention detection
- Tighter monitoring in complex scenarios, relaxed on empty highways
- Feature engineering works for driver monitoring unlike external perception
- 92% accuracy sufficient due to human-in-loop training effect
- Tesla will add driver monitoring before achieving Level 5
Hardware and Competition
- Current hardware: Snapdragon 821 (cell phone + cooling + car interface)
- Nvidia monopoly problem - A100 costs $10,000 vs RTX 3080 at $699
- Google TPUs only available in cloud with restrictive terms
- Tesla's Dojo approach (Apple-style vertical integration) makes sense
- Should sell chips to build ecosystem vs keeping proprietary
Autonomous Driving Predictions
- Level 5 self-driving in "the 2020s" - later part of decade
- Comma AI could overtake Tesla in users (Android vs iOS model)
- Tesla likely to win due to resource advantages
- Waymo product doesn't make sense - 20% slower than human drivers
- Users won't trade speed for cost savings (UberPool only 15% adoption)
Company Culture and Location
- Based in San Diego (not San Francisco)
- San Francisco problems: stopped wanting growth, one-party city, lacks diversity of thought
- Coronavirus accelerated exodus from San Francisco
- Profitable company with 8.1 million raised total
- Open source approach with 40% users running forks
Programming Languages and Advice
- Essential languages: Assembly, C, Python, Haskell, Verilog, PyTorch
- Assembly teaches computer fundamentals
- C teaches compilation and memory management
- Python builds on C abstractions
- Haskell for functional programming paradigm
- Verilog for hardware understanding
- PyTorch for machine learning (switched from TensorFlow)
Philosophy and Future
- Objectivism influence from Atlas Shrugged (read at 19)
- Ethical life is productive life
- Technology always wins, lying always loses
- Humans will be replaced by AI in our lifetime
- Compression equivalent to intelligence (Hutter Prize)
- Currently in singularity due to increased human communication bandwidth
