Lex Fridman PodcastSebastian Thrun: Flying Cars, Autonomous Vehicles, and Education | Lex Fridman Podcast #59
CHAPTERS
- 0:00 – 4:33
Simulation, The Matrix, and the universe as computation
Lex opens with a philosophical question about whether we live in a simulation and what that would mean. Thrun reframes it as largely irrelevant to how we should live, emphasizing being present and valuing people over abstract speculation.
- •Defining what “simulation” even means
- •Universe as an information-processing system
- •Lack of intention/purpose behind computation of the universe
- •Pragmatism: focus on life “here and now”
- 4:33 – 6:12
Early computing dreams and the limits of 1980s hardware
Thrun reflects on early programming and fascination with Silicon Valley, then contrasts past computing limitations with today’s scale. He notes that early AI ambitions were constrained by tiny compute—‘cockroach-sized’ brains.
- •First programs on a TI-57 era calculator
- •Silicon Valley as an early inspiration
- •Why early AI couldn’t scale: limited compute and memory
- •Long-running fascination with intelligence and robotics
- 6:12 – 9:18
Building intelligence: why machine learning changed everything
The conversation shifts to what it takes to build intelligent robots. Thrun argues machine learning is the key leap because it replaces brittle rule-writing with learning from data and experience, akin to how children learn.
- •Limits of explicit programming for real-world skills (e.g., walking)
- •Rule-based systems vs learning from experience
- •Why software engineering is costly: enumerating contingencies
- •Learning by observation and practice can match expert performance
- 9:18 – 11:27
Choosing problems: impact on society and learning by being ‘bad at the job’
Lex asks how Thrun picks world-changing projects. Thrun describes two drivers: maximize societal impact and maximize personal learning—preferring roles where he’s not already an expert.
- •Two motivations: help others + keep learning
- •Deliberately choosing unfamiliar domains to grow
- •Transportation as a massive lever for societal change
- •Self-driving and eVTOLs as near-term paths to reduce deaths and inefficiency
- 11:27 – 13:55
DARPA Grand Challenge war stories: bugs, focus, and learning-centric design
Thrun recounts pivotal moments from building Stanley and Junior, including maddening failures caused by obscure timing bugs. He credits Stanford’s emphasis on software and learning—rather than fancy hardware—as the decisive advantage.
- •Infamous ‘suicide every 30 miles’ bug from clock synchronization
- •Why focusing on the “brain” matters more than upgrading the car body
- •Machine learning used to imitate human speed control
- •Learning from mistakes and recovery in the field
- 13:55 – 17:23
Shipping reliable autonomy: small teams, frozen code, and ruthless testing
They dig into how to deliver a working system under extreme uncertainty. Thrun attributes success to a small, ego-free team, finishing early, and executing a rigorous testing regimen that constantly targets the weakest subsystem.
- •Small core team and high ownership across tasks
- •Freezing software a month early to avoid last-minute chaos
- •160-page test booklet and dedicated testing discipline
- •System improvement via “fix the weakest link” iteration
- 17:23 – 23:42
Leadership and empathy for engineers: empowering people, not commanding them
Lex asks about leadership, and Thrun contrasts managing computers with managing humans. He argues great leadership is about empathy, listening, and making others look great—skills he honed as a professor.
- •Don’t take credit; enable others to succeed
- •People aren’t deterministic like computers—emotion and pride matter
- •Professor-as-coach model: empower students/teams
- •Judging intentions vs actions; practicing perspective-taking
- •Practical social tools (e.g., Carnegie’s principles)
- 23:42 – 26:53
Why grand challenges work: funding outcomes, not effort
Thrun traces the evolution of self-driving research and praises DARPA’s prize-driven model. He argues it attracted ‘crazy’ outsiders and reset the ecosystem away from paper-heavy, procurement-driven research.
- •Early self-driving work (Germany, CMU) limited by compute
- •Government research incentives producing papers as “deliverables”
- •DARPA’s breakthrough: pay for outcomes via prizes
- •Prizes attracted diverse teams and seeded today’s AV leaders
- 26:53 – 33:08
Academia vs real problems: prototypes, interdisciplinary focus, and Silicon Valley symbiosis
Thrun defends universities’ core mission—educating people—but critiques research incentives that drift from societal problems. He stresses system-building prototypes and highlights Silicon Valley’s unique ability to turn research into products.
- •Academia’s primary function: educating young people to think
- •Research can drift into methods-first, problem-second mindsets
- •Self-driving cars dismissed as a ‘gimmick’ by some roboticists
- •“Impressing my grandmother” test for real-world relevance
- •Silicon Valley as an effective bridge from ideas to impact
- 33:08 – 36:10
State of self-driving today: the hard last 1%, cost, and deployment hurdles
Thrun describes progress as impressive but emphasizes how safety demands turn the ‘last fraction of a percent’ into the hardest part. He sees limited-scenario autonomy working today, with major work remaining in cost, robustness, and public acceptance.
- •90% driving is easy; safety-critical last 1% is brutally hard
- •Edge cases reveal human driving competence and adaptability
- •Evidence of strong performance in constrained deployments (e.g., Vegas pilots)
- •Next steps: down-costing sensors, automotive-grade hardening
- •Crossing the ‘empty car’ hurdle (no human inside)
- 36:10 – 47:27
Tesla vs Waymo and the lidar debate: deep learning shifts the paradigm
Lex probes contrasting strategies and Elon Musk’s ‘lidar is a crutch’ claim. Thrun notes cameras are sufficient in principle (humans prove it), celebrates parallel experimentation, and explains how deep learning boosted perception capabilities dramatically.
- •Innovation vs safety tradeoffs across companies
- •Cameras as an existence proof; multiple sensor stacks can work
- •Value of many competing hypotheses (“anthill” exploration)
- •Deep learning’s rise changed AV perception from geometry-first to learning-first
- •Udacity lane-finding example: rapid ML-based solutions outperform hand-coded rules
- 47:27 – 51:08
What ML is (and isn’t): narrow pattern recognition, not general intelligence
Thrun clarifies that modern ML excels at extracting patterns from large datasets but lacks broad, flexible general intelligence. He uses medical imaging as an example of near-expert performance while underscoring that such systems don’t generalize to unrelated tasks.
- •ML finds repetitive patterns in large labeled datasets
- •Stanford skin-cancer classifier at dermatologist level
- •Narrow competence doesn’t transfer (can’t ‘also drive a car’)
- •Humans learn from small data; machines still largely can’t
- •Hollywood AI fears vs realistic near-term capabilities
- 51:08 – 54:04
AI in medicine: catching disease early and augmenting doctors’ expertise
Lex asks how AI can help clinicians, and Thrun tells a story of an app flagging a melanoma that a doctor initially dismissed. He argues routine, scalable screening could catch cancers early—when they’re far more treatable—and save huge numbers of lives.
- •Real-world melanoma detection story (“Ben”)
- •Why early detection matters: rapid progression and survival rates
- •AI as a ubiquitous screening layer beyond doctor availability
- •Potential for passive/regular monitoring despite privacy concerns
- •Goal: reduce diagnostic error and variability in expertise
- 54:04 – 57:41
Education as a human right: reskilling for AI-driven job shifts via Udacity
They turn to job displacement worries, and Thrun responds with a reskilling mission. He describes large-scale scholarships, partnerships with government, and a global vision of accessible education across ages, geographies, and backgrounds.
- •Job disruption response: rapid, affordable skill-building pathways
- •100,000 scholarships and collaboration with the White House
- •Learners across ages 11–80; international cohorts and outcomes
- •Scaling beyond selective universities to global participation
- •Education framed as a basic human right
- 57:41 – 1:00:12
Soft skills and the future of work: teaching empathy, teamwork, and management
Thrun highlights a gap in both universities and online programs: soft skills. He argues empathy and collaboration are critical to success and may be more valuable (and teachable) than many purely technical competencies.
- •Demand from employers and learners for soft-skills training
- •Universities overemphasize individual scores vs teamwork
- •Empathy and management fundamentals as teachable skills
- •Hiring math skills is easier than ‘hiring empathy’
- •Workplace success depends on empowering and understanding others
- 1:00:12 – 1:08:20
Flying cars (eVTOLs) at Kitty Hawk: noise, redundancy, autonomy, and airspace scaling
Lex shifts to Kitty Hawk’s eVTOL vision, and Thrun explains how electric distributed propulsion enables quieter, safer, more affordable aircraft than helicopters. He argues the remaining barriers are largely societal (noise/acceptance) and operational (autonomous flight and digital air-traffic control at scale).
- •eVTOLs as helicopter-like vehicles with key differences (quiet, electric, redundant motors)
- •Project Heaviside: low noise and practical range/speed targets
- •Redundancy vs helicopter single-point failures (e.g., “Jesus bolt”)
- •Sky is 3D: virtual lanes and altitude separation enable massive capacity
- •Autonomy as essential; digitized traffic control modeled on networked systems
- 1:08:20 – 1:13:24
AI, love, and tools: trust, reliability, and technology that complements humans
Lex asks about emotional AI like in *Her*, and Thrun rejects anthropomorphizing tools. He argues technology should be predictable and trustworthy, designed to augment humans into ‘superhumans’ rather than replace human relationships.
- •AI as a tool (like a shovel), with ethics residing in humans
- •Reliability/predictability over emotional agency in machines
- •Technology should complement, not replace, human uniqueness
- •Trust is essential; love is a human mechanism for trust and safety
- •Bio-inspired isn’t always optimal: engineering finds different solutions (wheels vs legs)
- 1:13:24 – 1:18:34
Gratitude, progress, and the hidden heroes of modern life (Carl Bosch)
In closing, Thrun explains his optimism and humor as rooted in gratitude for living in an era of rapid progress. He cites Steven Pinker and the impact of nitrogen fertilization (Carl Bosch) as an underappreciated innovation that enabled billions of lives.
- •Perspective on historical life expectancy and modern ‘extra time’
- •Technological progress over the last ~150 years transformed daily life
- •Printing press and information dissemination as deep enablers
- •Carl Bosch and nitrogen fertilization multiplying food yield
- •Closing thanks and final reflection on impact and gratitude
