Self-driving cars have been promised as five years away for about fifteen years running, which makes the people actually building them some of the most interesting guests a podcast can book. We combed through our full library of episode summaries to find the conversations where engineers, founders, and researchers go past the marketing language and explain what autonomy actually requires: sensor fusion, edge cases, safety validation, and business models nobody expected.
This list mixes the loudest names in the space (George Hotz, Elon Musk) with the people who quietly built Waymo, Cruise, Aurora, and nuTonomy from scratch. Expect specific numbers, real disagreements about lidar versus cameras, and a few founders who bet their careers on approaches everyone else thought were wrong.
George Hotz: Tiny Corp, Twitter, AI Safety, Self-Driving, GPT, AGI & God | Lex Fridman Podcast #387
Hotz's third sit-down with Lex Fridman ranges far past cars, but the comma.ai update is the anchor: he discusses where 'Drive GPT' stands and doubles down on end-to-end neural nets over feature-engineered stacks. He also drops a genuinely odd reveal, that AMD's Lisa Su personally replied to his email after her 7900 XTX drivers panicked running his demo apps in a loop. Listen for the argument that decentralized, open-source AI is the only real safeguard against a handful of companies monopolizing self-driving and everything else. Best for listeners who want Hotz's engineering opinions wrapped in his broader philosophy on AI and power.
Read the full episode notesGeorge Hotz: Hacking the Simulation & Learning to Drive with Neural Nets | Lex Fridman Podcast #132
This is the episode where Hotz makes his clearest technical case: comma.ai trains an end-to-end network on real user driving data, while Tesla's approach is 'feature engineering' that he bets will lose. He calls MuZero the cornerstone paper of the deep learning era and effectively the solution to self-driving. The concrete number worth remembering is openpilot's improvement from one disengagement every 10 miles to one every 100 miles in about a year. Best for anyone who wants the end-to-end versus modular debate explained by someone with real skin in the game.
Read the full episode notesGeorge Hotz: Comma.ai, OpenPilot, and Autonomous Vehicles | Lex Fridman Podcast #31
The origin story episode: Hotz explains why he thinks lane-keeping done well is the actual consumer value in self-driving, why he considers 'level 4' a marketing fiction, and why lidar is a crutch rather than a requirement. He also reveals Elon Musk once offered him $12 million to match Mobileye's performance, with a penalty of $1 million per month of delay. Best for listeners who want the founding logic behind comma.ai's level-2, driver-monitored approach before the company had years of track record to point to.
Read the full episode notesElon Musk: Tesla Autopilot | Lex Fridman Podcast #18
Musk lays out the case that a Tesla bought today is an appreciating asset because the hardware is already capable of full self-driving, with the rest arriving via over-the-air updates. He claims roughly half a million Teslas on the road with the full sensor suite account for about 99% of all relevant driving data collected anywhere, and defends the controversial idea that once the system is safer than a human, requiring driver intervention could actually reduce safety. Best for listeners who want Tesla's autonomy philosophy straight from Musk, including the elevator-operator analogy he uses to argue against strict driver monitoring.
Read the full episode notesDmitri Dolgov: Waymo and the Future of Self-Driving Cars | Lex Fridman Podcast #147
Dolgov traces Waymo from the DARPA Urban Challenge to the moment its Phoenix service went truly driverless, including the detail that Waymo's very first day of regular operation in October 2017 put riders in cars with an empty driver's seat, not just a passive safety driver. He explains the pivot away from an L3 freeway assist program around 2013 once the team decided full autonomy and driver-assist were fundamentally different problems. Best for listeners who want the view from the company furthest along in actually removing the human, told by the person who ran the technical side of getting there.
Read the full episode notesDrago Anguelov (Waymo) - MIT Self-Driving Cars
Anguelov's MIT lecture is built around 'taming the long tail,' the rare scenarios that break naive systems: cyclists carrying stop signs, falling poles, red-light runners. He details Waymo's simulation infrastructure running the equivalent of 25,000 virtual cars driving 10 million miles a day, over 7 billion simulated miles total, and argues for hybrid systems that combine learned models with hard-coded domain expertise rather than trusting either alone. Best for listeners who want to understand why the hardest 10% of self-driving takes 90% of the effort.
Read the full episode notesSacha Arnoud, Director of Engineering, Waymo - MIT Self-Driving Cars
Arnoud traces deep learning's rise inside Google from Street View house-number recognition to real-time perception running on the cars themselves, and reveals Waymo pulled its safety driver entirely for the first time in November 2017. His mantra, that being 90% done still leaves 90% left to go, sets up a sober account of sensor fusion, labeling, and a three-part testing regime across real driving, simulation, and a dedicated test facility. Best for listeners who want the perception-engineering side of autonomy rather than the philosophy.
Read the full episode notesKarl Iagnemma & Oscar Beijbom (Aptiv Autonomous Mobility) - MIT Self-Driving Cars
Iagnemma traces autonomous driving from the 2007 DARPA Urban Challenge to Aptiv's paid Lyft rides in Las Vegas, where roughly 75 cars have delivered over 30,000 rides with a 4.95 average rating. His central argument is that neural networks need to be 'caged' inside verifiable safety architectures rather than trusted as end-to-end black boxes, a direct counterpoint to Hotz's philosophy elsewhere on this list. Beijbom follows with the technical detail on PointPillars, a fast lidar point-cloud encoder, and the nuScenes open dataset. Best for listeners who want the safety-first case against end-to-end learning.
Read the full episode notesEmilio Frazzoli, CTO, nuTonomy - MIT Self-Driving Cars
Frazzoli, who invented the RRT* motion-planning algorithm and put the first autonomous vehicles on Singapore's public roads, argues the real payoff of self-driving isn't safety, it's reclaiming the roughly $1.2 trillion a year in value lost to time spent driving. He calls the SAE's numbered automation levels an 'enormously bad idea' because levels 2 and 3 require humans to supervise automation, which runs against human nature. Best for listeners who want a sharp, contrarian argument for skipping straight to full automation rather than incrementally adding driver-assist features.
Read the full episode notesKyle Vogt: Cruise Automation | Lex Fridman Podcast #14
Vogt's path runs from building BattleBots at 13 in Kansas, through the first DARPA Grand Challenge and co-founding Twitch, to picking self-driving cars as the greatest applied AI problem of his generation. The pivotal reveal is that Cruise had a working highway-autopilot prototype within a year of founding and completely abandoned it to go all-in on fully driverless cars, judging the retrofit approach too complicated on liability and validation grounds. Best for listeners who want a founder's-eye view of why some AV companies chose to skip driver-assist entirely.
Read the full episode notesOliver Cameron (CEO, Voyage) - MIT Self-Driving Cars
Cameron built Udacity's self-driving car curriculum, which trained over 14,000 students who went on to Cruise, Zoox, Waymo, and Aurora, before founding Voyage. His company's contrarian strategy is the standout reveal: deploy in retirement communities instead of competing with Waymo in big cities, trading exclusive operating licenses for equity stakes in communities like The Villages, which has over 125,000 residents and 750 miles of road. Best for listeners who want a genuinely different go-to-market story than the usual big-city robotaxi race.
Read the full episode notesSterling Anderson, Co-Founder, Aurora - MIT Self-Driving Cars
Anderson's path runs from MIT PhD work on an 'intelligent co-pilot' that quietly took up to 43% of steering control while making drivers feel 12% more in control, through leading Tesla's Model X and Autopilot programs, to co-founding Aurora with ex-Google and ex-Uber autonomy leads. He details Aurora's partner-driven business model and its Volkswagen and Hyundai deals. Best for listeners curious how shared human-machine control research at MIT fed directly into two of the biggest names in the industry.
Read the full episode notesSertac Karaman (MIT) on Motion Planning in a Complex World - MIT Self-Driving Cars
Karaman explains how he proved the widely used RRT motion-planning algorithm fails to converge to optimal solutions, leading to his RRT* algorithm, and tells the story of MIT's DARPA Urban Challenge car surviving a low-speed collision caused by a bug in Cornell's competing robot. He also makes the case that autonomous, shared, electric vehicles could radically cut transportation costs and reshape cities. Best for listeners who want the academic roots of motion planning explained by the person who literally wrote the algorithm most AV stacks still use.
Read the full episode notesRana el Kaliouby: Emotion AI, Social Robots, and Self-Driving Cars | Lex Fridman Podcast #322
El Kaliouby pioneered emotion-recognition AI and now runs Smart Eye's automotive sensing work, covering drowsiness, distraction, and alcohol detection inside the car. Her most useful reveal for this topic is that Affectiva found people still emote facial expressions even when alone and driving, the data point that makes in-car emotion sensing viable at all. Best for listeners interested in the human-monitoring side of autonomy, the sensors watching the driver rather than the road.
Read the full episode notesMichael Littman: Reinforcement Learning and the Future of AI | Lex Fridman Podcast #144
Littman traces reinforcement learning's history from TD-Gammon through AlphaGo and AlphaZero, and pushes back directly on superintelligence doom narratives, including a pointed op-ed he once wrote against Elon Musk's AI warnings. His most relevant thread here is his argument that driving is a surprisingly social problem, not just a control problem, which reframes why self-driving is so much harder than the sensor specs suggest. Best for listeners who want the reinforcement-learning theory underneath self-driving explained without hype in either direction.
Read the full episode notesThat's fifteen conversations spanning the founders betting their careers on end-to-end learning, the engineers who built Waymo's safety case mile by mile, and the researchers arguing about what a self-driving car even needs to prove before we trust it. Browse the full library of episode summaries on Episode Notes for more conversations worth your time.