AI summary by TwinMind
Sebastian Thrun: Flying Cars, Autonomous Vehicles, and Education | Lex Fridman Podcast #59
Lex Fridman
1h 18m 34s
Introduction and Background
- Sebastian Thrun: roboticist, computer scientist, educator
- Led Stanford autonomous vehicle development: won 2005 DARPA Grand Challenge, 2nd place 2007 DARPA Urban Challenge
- Led Google self-driving car program launching self-driving revolution
- Taught Stanford AI course 2011 - one of first MOOCs
- Co-founded Udacity online education platform
- CEO of Kitty Hawk (flying cars/eVTOLs - electric vertical takeoff and landing aircraft)
Philosophy and Simulation Discussion
- Matrix as favorite movie
- Simulation hypothesis: "maybe we are not but completely irrelevant to how we should act"
- Universe as information processing system - "huge physical biological chemical computer"
- Focus on present reality and helping people
Early Computing and AI Vision
- First programs written on TI-57 calculator (early 1980s)
- Early fascination with Silicon Valley, D-RAMs, CPUs
- Always wanted to make robots smart, build artificial humans
- Recognized scale limitations in 80s/90s - computers "brains you could build not anywhere bigger as cockroach"
Machine Learning Revolution
- Biggest innovation: computers can teach themselves vs being programmed with rules
- Example: walking - humans learn but can't articulate rules for human gait
- Software engineers cost $200k/year due to difficulty of instruction sets
- 30-40 years of development: computers finding their own rules from experience
- AI systems watch experts, pick up skills humans can't express in words
- Examples: dermatologists, radiologists
Career Motivation and Problem Selection
- Two desires: "make lives of others better" and "learn"
- Wants to be in jobs he's bad at for maximum learning opportunity
- Focus on maximum societal impact after Stanford professorship
- Transportation transformed 20th century more than communication
- Current problems: 1.2 million deaths/year in traffic, inefficient resource use, traffic jams
- Technology basically exists for solutions
DARPA Grand Challenge Experience
- Stanley won 2005, Junior 2nd place 2007
- Painful: complicated bugs (clock synchronization causing suicide every 30 miles)
- Enlightening: Stanford focused on machine learning/software vs others on hardware
- Philosophy: replace human with sensors (eyes/ears), brain (learning), actuators (hands/feet)
- Massive machine learning approach: learn from human drivers and robot's own experience
Team Building and Leadership
- Core team of 4 people (5 couldn't fit in car comfortably)
- Leader role: get pizza, wash car, repair radiator, debug system
- Open minded, no egos
- Time management: finished 1 month early, froze system, found 2-3 minor bugs in final month
- 160-page testing booklet, continuous focus on weakest system parts
- Leadership with people vs computers: people are smart, emotional, have pride/aspirations
- Need to connect to what people want, empower them
- Academic background: job is making students look great (coaching)
- Every person has good intentions, wants to help others
- Judge ourselves by intentions, others by actions
- Skill: put yourself in others' position, listen
DARPA Challenge Impact on Funding
- Changed from effort-based to outcome-based funding
- No tax code allowed congressional money for prizes initially
- Drew out traditional contractors, brought in "crazy people" with cars/computers
- Universities moved in (Stanford with tenure protection)
- Car companies participated quietly (Ford, Volkswagen, GM)
- Reset landscape - current self-driving leaders mostly DARPA participants
Academia vs Industry
- Believes in end-to-end system building vs component problems
- "What problem would my grandmother care about?"
- Academia's primary purpose: educate young people (professors are educators first)
- Research secondary but universities produce great science (Google traceable to universities)
- America has finest university system globally
- Wish: more debate about great big problems (mostly interdisciplinary)
- Example: colleagues said cars would destroy career (too simple, not expressive enough)
- Self-driving cars have most societal impact potential among robotics problems
- Silicon Valley advantage: vertical integration between research and customer-facing products
Current Autonomous Vehicle State
- Extremely proud of accomplishments, very optimistic
- Problem difficulty: 90% solvable in weekend, 99% in month, final 1% takes enormous effort
- 1% failure rate = fatal accident every week (unacceptable)
- Variety of edge cases enormous, respect for human driving ability
- Technology already exists in limited scenarios (Waymo in Vegas with safety drivers)
- Next steps: cost reduction, hardening, business models, social acceptance
- Key hurdle: vehicles driving completely empty without safety drivers
Approach Comparisons (Tesla vs Waymo)
- Tesla: incremental Level 2 approach using machine learning from human driving
- Waymo: Level 4 full autonomy with safety driver approach
- Both valid approaches in innovation ecosystem
- Balance between public safety and innovation drive
- US history of successful safety/innovation balance (aviation 100x safer than ground transport)
- Methods exist for managing dangerous technologies (nuclear energy, weapons)
Deep Learning Impact
- Wasn't hot topic when starting Google self-driving cars
- Shift from geometric reasoning to human-like pattern recognition
- Put all approaches "on steroids"
- Udacity self-driving course: 20,000+ people trained
- Example: lane finding assignment - students create perfect lane finder in 24 hours using machine learning vs complex rule-based systems
- Progress using ML completely dwarfs anything from 10 years ago
LiDAR vs Camera Debate
- Humans can drive with eyes only (existence proof cameras sufficient)
- Western advantage: many different teams trying different hypotheses
- Like anthill foraging - spreading out better than central planning
- Prefers distributed approach over Soviet-style central decision making
Machine Learning Capabilities and Limitations
- Technology for finding repetitive patterns in large databases
- Example: skin cancer detection - 129,000 skin photos dataset
- iPhone app as accurate as best Stanford dermatologists for melanoma detection
- Real story: friend Ben's melanoma caught by app when doctor missed it
- Not general intelligence - needs large datasets, very repetitive patterns
- Humans do many things, learn from small datasets
- Can transform human labor in specific repetitive tasks
- 10,000 hours to become expert - machines can observe experts and extract rules
- Can make people expert on first day of work
Medical AI Applications
- iPhone skin cancer detection app proven effective
- Story of Ben (venture capitalist): doctor missed melanoma, app caught it
- Cancer is cause of death #2, most curable if caught early
- Problem: cancers have no early symptoms
- Vision: AI inspect body regularly (e.g., during shower) without doctor present
- Could save lives through early detection
Job Displacement and Education
- Udacity: job promise, 50%+ graduates get new jobs
- Costs $1500 maximum
- US government program: 100,000 scholarships for American citizens
- Courses range from digital marketing to advanced self-driving car engineering
- Bipartisan issue - making America great requires proactive education
- American dream evolution: send ourselves to college at all ages
- Learners age 11-80, global reach (6 countries officially)
- Egypt example: 1100 high school students, 95% graduation rate, computer science undergrad level
Udacity Vision and Impact
- Education as basic human right, not locked behind ivory tower walls
- Open to everyone regardless of demographics, geography, race, age
- 100,000 scholarships pledge - biggest in history by numbers
- Working with White House, CEOs like Tim Cook
- Global education mission
Future Course Development
- Most excited about soft skills courses (not yet launched)
- Hard skills: programming, math, building, design
- Soft skills: workplace behavior, empathy, teamwork, management, time management, career advancement
- Universities do poorly with teamwork vs individual test scores
- Teaching empathy could have same impact as self-driving cars
- Simple rules but need engagement
- K-12 education focuses on individual scores vs collaboration
- Math/English scores don't make best leaders
- Can hire people with great math skills, can't hire empathy
- Roboticists/computer scientists would benefit from empathy courses
Flying Cars (eVTOLs) - Kitty Hawk
- Universal dream of flying (childhood pillow flying dream)
- Peter Thiel quote: "promised flying cars, got 140 characters" (Twitter)
- Not actually cars - closer to helicopters, vertical takeoff/horizontal flight
- Key differences from helicopters:
- Much quieter: Project Heaviside 38 decibels (library quiet)
- More affordable due to distributed electric motors vs single gas engine
- No single points of failure (8 motors, can lose 1 and still operate)
- No commercial pilot required
eVTOL Technical Capabilities
- Heaviside: 100-mile range with 30% reserves
- 180 mph top speed
- 10x faster than car transportation based on census data
- 300 hours/year commute → 30 hours (270 hours saved)
- Technology exists, not technological problem anymore
Future Sky Traffic Management
- Believes tens/hundreds of thousands of vehicles will fill sky
- Safety must exceed ground transportation
- Noise control crucial for acceptance
- Sky is 3D vs 1D highway - mathematical advantage
- Virtual lanes vs physical infrastructure
- 100 vertical lanes of Highway 101 would cost world's GDP for year
- Sky lanes are software compilation problem
- Altitude separation solves conflicts (like commercial aviation)
- 100% autonomous required - "don't want idiots like me flying"
- Centralized management system necessary
Scaling Air Traffic Control
- Current ATC: up to 20 planes per frequency with voice communication
- 20,000 vehicles need digital text messaging system
- 4-5 billion smartphones prove scale solutions exist
- Digital transmission of flight coordinates vs ancient frequency changes
- Same principles as current aviation, scaled up
AI and Human Relationships
- Pragmatic view: AI is tool like shovel
- Ethics lies with people using tools
- Doesn't want emotional refrigerator that falls in love with dishwasher
- Technology should be reliable, predictable, complementary not replacement
- Already superhuman thanks to machines:
- Fly across Atlantic near speed of sound at 36,000 feet
- Voice travels at speed of light globally
- Flushing toilet makes us superhuman vs pre-technology times
- Technology complements and strengthens humans
- Love/compassion have little interest for machines
Trust vs Love in Technology
- Love conveys safety and trust in humans
- Trust has huge need in technology
- Human emotions evolved to convey trust, safety, dedication
- Wouldn't blindly copy human systems to machines
- Historical example: transportation used legs (1 to 1000) until technology found wheels
- Biology rarely has perpetually rotating things, engineering excels at wheels
- Solutions use different physical laws than nature
- If humanoid robots were solution, still waiting today
Life Philosophy and Gratitude
- Just loves life and people around him
- Age 52, grateful to be alive
- Historical perspective: life expectancy ~30 years for 299,900 of 300,000 years of humanity
- Would be long dead in previous eras
- Amazing timing to be born after massive wars, famines ended
- Germany received undeserved help from Americans/Allies post-WWII
- Last 150 years magical period of invention
- Traces progress to Gutenberg printing press enabling information dissemination
- Agriculture improvements freed humans from farming focus
- Steven Pinker's "Enlightenment Now" recommendation
Carl Bosch Impact
- Most important 20th century innovation according to Pinker
- Invented nitrogen fertilization
- Increased agricultural yield by factor of 26 (2,500% increase)
- Saved over 2 billion lives
- Individual impact: team up with skill, worth $10 million in acquisitions
- Celebrates being alive after Bosch's invention, not before
