← Back to list
Lex Fridman PodcastPodcast21 Dec 2019Source: lexfridman.comHost: Lex Fridman

Sebastian Thrun: Flying Cars, Autonomous Vehicles, and Education

In plain words

This is an interview with Sebastian Thrun, a Stanford professor and founder of Google's self-driving car project. He says self-driving tech is mostly ready, but the last 1% is extremely hard—the real bottlenecks are cost and public acceptance. He's bullish on Waymo (tech leader but not yet fully driverless), Tesla Autopilot (he uses it daily and feels safer), and flying-car startup Kitty Hawk (quiet, 100-mile range). He also argues AI is just pattern recognition, not general intelligence, and that education should be a basic human right.

AI SummaryAI-generated · may contain errors · verify against the original

Sebastian Thrun discussed autonomous driving, flying cars (eVTOL), and online education on the Lex Fridman podcast. Key points include: his Stanford team won the 2005 DARPA Grand Challenge and placed second in the 2007 DARPA Urban Challenge; he subsequently led Google's autonomous driving project, s

~9 min full read · 10 sections
Deep Analysis

At a Glance

Sebastian Thrun (Stanford professor, founder of Google's self-driving car project, co-founder of Udacity, and CEO of Kitty Hawk) reflects on the three technological revolutions he has participated in. The most weighty judgment in the entire piece: Thrun believes autonomous driving technology is essentially mature, and the current bottleneck lies not in technology but in cost, business models, and social acceptance—"We have reached the stage where the technology is ready, just one final step away (driverless empty vehicle operation)."


1. The Philosophy of Problem Selection: Doing What You're "Not Good At"

Thrun's two criteria for choosing projects: making the world better + enabling personal learning. He deliberately avoids areas where he already excels—"If I'm already good at a job, the opportunity to learn something new is minimized. I want to do work I'm not good at."

  • Mechanism breakdown: When evaluating the impact of a problem, he uses the "grandmother test"—"Can this problem matter to a grandmother who doesn't understand technology? Only then can I have a real impact on the world."
  • Historical context: From the DARPA Challenge to autonomous driving to flying cars, he has consistently focused on transportation issues with the "greatest social impact"—"Transportation changed the 20th century more than any other invention, even more than communication. Cities, jobs, women's rights—all were transformed by transportation."

II. Lessons from the DARPA Challenge: Systems Thinking and a Culture of Testing

Thrun argues that the key to DARPA Grand Challenge's success was not a technological breakthrough, but a "results-oriented" funding model and rigorous system testing.

  • Data chain: No team completed the course the year before the 2005 DARPA Challenge; Thrun's core team had only 4 people; they froze all software a month in advance, while other teams "were all saying they could win if they just had one more week."
  • Mechanism breakdown: The team created a 160-page test manual, testing specific scenarios (railroad crossings, starting procedures, etc.) daily as planned, continuously improving the system's weakest link—"If you keep improving the weakest part, you will eventually build a great system."
  • Comparison data:
Team Strategy Thrun's Team Other Teams
Software freeze time 1 month before the race Last minute before the race
Core team size 4 people Larger teams
Technical focus Machine learning + software Hardware improvements
  • Falsification condition: Thrun points out that if they had not frozen the software early or conducted systematic testing, the team might have ended up like the others, "one week short of winning."

3. Current State of Autonomous Driving: 90% Easy, the Last 1% Extremely Difficult

Thrun believes autonomous driving technology is largely ready, but the difficulty of moving "from 99% to 99.99%" is severely underestimated.

  • Data chain: 90% of driving scenarios can be solved in a weekend; 99% may take a month; but the remaining 1% means "a fatal accident every week" — completely unacceptable.
  • Mechanism breakdown: Humans can handle never-before-trained emergencies like "a sofa on the highway," "a deer in front of headlights," or "a tire blowout." Machines have yet to master this generalization capability.
  • Current bottlenecks: Cost reduction (sensors have not yet met automotive-grade standards), system hardening, business model establishment, and social acceptance — "no team has yet had the courage to operate a fully driverless empty vehicle."
  • Falsification condition: Thrun waits daily for Waymo to announce the achievement of fully driverless empty-vehicle operation — "that is the next magical threshold."

4. Tesla vs Waymo: Different Paths, Each with Its Own Value

Thrun believes both paths have value, with the key being that "society needs multiple attempts"—he uses ants foraging as an analogy.

  • Mechanism breakdown: If all ants discuss the "optimal path" before setting out, they all fail if it is wrong; if they explore separately, someone will always succeed and lead the way—"This is the greatness of Western society: we are not plan-oriented, not centralized; we each explore on our own."
  • On the sensor debate: Thrun acknowledges that "humans can drive using only their eyes" is a proof of existence, making a vision-only approach theoretically feasible; however, he emphasizes the value of "letting different companies try different assumptions."
  • Personal stance: Thrun describes himself as a "proud Tesla owner" who uses Autopilot daily—"It does make me safer, especially when I am a bit tired."

5. Flying Cars (eVTOL): Technology Is Ready, the Key Is Social Acceptance

Thrun argues that the technical challenges of flying cars have largely been resolved, with the core hurdles being noise, cost, and social acceptance.

  • Data Chain: Kitty Hawk's Heaviside project: flight noise at 38 decibels (lower than outdoor ambient noise); range of 100 miles (including 30% power reserve); top speed of 180 mph; 8 motors, with normal takeoff and landing capability even if one fails.
  • Mechanism Breakdown: Compared to helicopters, electric distributed propulsion (multiple small motors instead of one large motor) eliminates single-point failures like the "Jesus nut"; electrification significantly reduces costs.
  • Airspace Management: Thrun views the sky as three-dimensional—"Building 100 vertical lanes on Highway 101 would consume global GDP, but in the sky, it's just a software recompile." He draws an analogy to smartphone networks: "We have 4-5 billion phones connecting simultaneously; airspace management can scale digitally in the same way."
  • Falsification Condition: Social acceptance is the biggest uncertainty—"Noise is the key, so we spare no effort to reduce it."

6. The Nature of AI: A Pattern Recognition Tool, Not General Intelligence

Thrun clearly distinguishes between "current AI" and "artificial general intelligence" — the former is a powerful pattern recognition tool, while the latter "I won't see in my lifetime."

  • Mechanism breakdown: Current AI can "extract patterns from large amounts of repetitive data," but it cannot learn multiple tasks from limited experience like humans. He gives an example: a system that can detect skin cancer cannot drive a car.
  • Data chain: Stanford research trained AI on 129,000 skin photos, achieving detection accuracy comparable to Stanford dermatologists; however, this is a single task, not general intelligence.
  • On naming: Thrun has reservations about the term "artificial intelligence" — "If you believe Hollywood, AI immediately conjures images of humans being suppressed. I don't think that will happen."
  • Application direction: The true value of AI is to "make a novice an expert on day one" — "If your doctor is still in the first 10,000 hours of learning, would you be comfortable with that?"

7. Education: Skills Training as a Basic Human Right

Thrun argues that education should be a basic human right and should not be monopolized by the "ivory tower."

  • Data Chain: Udacity has graduated over 20,000 autonomous driving engineers; in Egypt, 1,100 high school students completed a programming course with a 95% graduation rate; partnered with the White House to provide 100,000 scholarships.
  • Mechanism Breakdown: The "departmentalization" of traditional universities makes it difficult to offer interdisciplinary courses (e.g., autonomous driving); moreover, universities are highly selective, excluding talent from regions such as India, China, and Africa.
  • On Soft Skills: Thrun is developing "soft skills" courses (empathy, teamwork, time management) — "We can teach people technical skills, but if we teach people empathy, the impact is equally enormous."

Mentioned Positions

Position Guest Stance Key Data
Waymo Bullish (Technology Leader) Capable of perfect driving in limited scenarios; has not yet achieved fully unmanned empty-vehicle operation
Tesla Autopilot Positive (Personal Use) Thrun claims to use it daily, believes it makes driving safer
Cruise Positive Mention Technology is ready
Aptiv Positive Mention Technology is ready
Voyage Positive Mention Technology is ready
Kitty Hawk (Heaviside) Bullish (CEO Stance) Noise level 38 dB; range 100 miles; speed 180 mph; 8-motor redundancy

Judgments Worth Remembering

1. "Choose jobs you're not good at" (Thrun) — Deliberately avoid the comfort zone, because "jobs you're good at minimize learning opportunities"; use the "grandmother test" to assess the impact of a problem.

2. "Freeze the software one month early" (Thrun) — The key to DARPA's success was not technological breakthroughs, but time management: finish early, test the system, and continuously improve the weakest link.

3. "The last 1% of autonomous driving is 100 times harder than the first 99%" (Thrun) — 90% can be solved over a weekend, 99% takes a month, but the remaining 1% means fatal accidents every week; humans' ability to handle "emergencies they have never been trained for" is something machines have yet to master.

4. "Ant foraging strategy" (Thrun) — Decentralized exploration is better than centralized discussion of the optimal path: let different companies try different hypotheses (pure vision vs. multi-sensor), and someone will succeed and lead the way.

5. "The sky is three-dimensional, and 100 virtual lanes only require a software recompile" (Thrun) — Airspace management for flying cars is analogous to smartphone networks: 4–5 billion phones connect simultaneously, and digital airspace management is equally scalable.

6. "AI is a tool, not a companion" (Thrun) — "I don't want my refrigerator to spoil my food because it falls in love with the dishwasher"; technology should be reliable, predictable, and make humans "superhuman" (e.g., supersonic transatlantic flight, transmitting sound at the speed of light).

7. "Education is a basic human right" (Thrun) — Udacity's mission is to break the monopoly of the "ivory tower": 20,000+ autonomous driving engineers graduated, 1,100 high school students in Egypt with a 95% graduation rate, and 100,000 U.S. citizen scholarships.

8. "Current AI is pattern recognition, not general intelligence" (Thrun) — A system that can detect skin cancer cannot drive a car; general artificial intelligence "will not be seen in my lifetime"; the true value of AI is "making a novice an expert on day one."