Computer vision sounds like a narrow engineering topic until you actually sit with the people who built it, and then it turns into a story about brains, cameras, driving deaths, and one enormous labeled dataset that changed everything. We combed our full library of episode summaries for the conversations and lectures that explain how machines learned to see, not the ones that just namedrop the term.
This list mixes founder interviews, researcher profiles, and full MIT lecture breakdowns, ranked by how much genuinely new information each one hands you. Some are casual and personal, some are whiteboard-technical, but every entry earns its spot with a specific reveal you can actually cite later. Expect ImageNet's backstory, why Waymo pulled its safety drivers, why one professor thinks vision may never be fully solved, and why a robot vacuum company is trying to make its robots less autonomous.
Dr. Fei-Fei Li, The Godmother of AI — Asking Audacious Questions & Finding Your North Star
Fei-Fei Li built the dataset that arguably kicked off the modern AI era, and this episode traces exactly how it happened, down to hiring Amazon Mechanical Turk workers out of desperation because Princeton undergrads were too slow and expensive to label tens of millions of images. She walks through the 2012 paper combining ImageNet with neural networks and GPUs that many call the birth of modern AI, all filtered through her own path from a Chengdu childhood to running a New Jersey dry cleaning shop while studying physics at Princeton. Anyone who wants the human story behind computer vision's founding dataset should start here.
Read the full episode notesColin Angle: iRobot CEO | Lex Fridman Podcast #39
The Roomba is the best-selling vacuum in the US, not just the best-selling robot vacuum, and Colin Angle explains how cheap vision-capable cameras and injection-molded plastic made that possible after a graveyard of failed competitors. He names Anki, Jibo, Mayfield Robotics, and Rethink Robotics as brilliant companies that all went bankrupt anyway, then argues the next frontier is making home robots less autonomous so they act like partners rather than independent agents. Good listening for anyone curious how computer vision actually ships in a product that survives contact with a real living room.
Read the full episode notesJitendra Malik: Computer Vision | Lex Fridman Podcast #110
Malik coined the "fallacy of the successful first step": getting 50 percent of a vision problem solved takes a minute, 99 percent takes five years, and the last sliver may not happen in your lifetime. He argues most human visual processing is subconscious and that machines relying on supervised, feed-forward learning are missing how perception is tied to action, which is why he is a near-term pessimist on full self-driving. Listen for the argument that a 16-year-old learning to drive is already a visual genius and driver's ed mostly teaches control, not sight.
Read the full episode notesDileep George: Brain-Inspired AI | Lex Fridman Podcast #115
George built the Recursive Cortical Network, a non-deep-learning graphical model that cracked text CAPTCHAs with very little training data, then watched Science's own press office mislabel it as "a new deep learning model" anyway. He argues brain-simulation projects like Blue Brain are doomed without a functional theory, comparing it to nailing a single transistor's behavior and still having no idea how to build a microprocessor. The Kanizsa triangle detail, where neurons fire later for a hallucinated edge than a real one, is the kind of specific evidence that makes his feedback-driven theory of vision land. For listeners who want a real alternative to the deep-learning-only narrative.
Read the full episode notesDeep Learning for Computer Vision (Andrej Karpathy, OpenAI)
Karpathy traces CNNs from Hubel and Wiesel's 1960s cat experiments through LeNet and the 2012 AlexNet breakthrough, then shows how ImageNet top-5 error dropped to 3.57 percent by 2016, matching or beating human accuracy. He even built a web interface to test his own ImageNet accuracy against a CNN and lost points mostly misidentifying dog breeds, a detail that says a lot about how fine-grained these categories actually are. A clean, technical entry point for anyone who wants the CNN mechanics explained by someone who has taught this at Stanford.
Read the full episode notesSacha Arnoud, Director of Engineering, Waymo - MIT Self-Driving Cars
Arnoud traces self-driving from Google's 2009 Chauffeur project to the moment in November 2017 when Waymo removed its safety drivers entirely and started running fully driverless cars around Phoenix. His mantra, that being 90 percent done still leaves 90 percent to go, captures why perception took a decade of 10x improvements across sensors, software, and simulation. Useful for anyone who wants to know what it actually took to move computer vision out of a lab and onto a public road with no one behind the wheel.
Read the full episode notesTomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Poggio argues the recent AI breakthroughs in deep learning and reinforcement learning both originated from neuroscience, and bets the next ones will too, including his theory that the brain's face-recognition region is not hardwired but rapidly imprinted in early life. He cites a striking experiment where baby monkeys raised without ever seeing faces developed no face-preference area in their brains at all. Best suited for listeners who want computer vision connected back to the biology it is supposedly modeled on.
Read the full episode notesEfficient Computing for Deep Learning, Robotics, and AI (Vivienne Sze) | MIT Deep Learning Series
Sze's core argument reframes the whole field: power in vision systems is dominated by moving data, not by doing the math, since reading from off-chip memory costs two orders of magnitude more energy than the actual computation. She notes a self-driving car can burn over 2,000 watts on computation alone, enough to need water cooling, while a handheld device is capped under a single watt. Essential for anyone thinking about why computer vision on your phone works differently than computer vision in a data center.
Read the full episode notesMIT 6.S094: Computer Vision
This lecture builds vision intuition from the ground up, starting with a trivial pixel-difference classifier that hits just 35 to 38 percent on CIFAR-10, then climbing through K-nearest neighbors to the CNN architectures that eventually surpassed human-level ImageNet accuracy in 2015. Fridman surveys the ImageNet-winning lineage from AlexNet through ResNet while tying the layered structure of these networks back to the visual cortex. The clearest technical walkthrough on this list for anyone who wants convolutions actually explained, not just referenced.
Read the full episode notesMIT 6.S094: Deep Learning for Human Sensing
MIT collected over five billion video frames of driving data, including 1.5 billion frames of driver faces, from 25 vehicles, 21 of them running Tesla Autopilot, to study how well vision systems can read the human inside the car. The reveal that Tesla drivers used Autopilot for roughly a third of their miles shows the real-world stakes behind detecting pedestrians, gaze, and drowsiness. A strong pick for listeners who care about vision applied to people, not just roads and objects.
Read the full episode notesMIT 6.S094: Deep Learning for Human-Centered Semi-Autonomous Vehicles
Fridman flips the camera inward, describing 17 Teslas driving around Cambridge collecting billions of frames to study the driver rather than the road, since almost every car with autonomous features has zero sensors watching the person behind the wheel. He walks through the CNN pipeline for detecting body pose, gaze, emotion, and drowsiness, including how 3D convolutional networks treat stacked video frames like color channels. Pairs well with the human-sensing lecture above for anyone building out a full picture of driver-monitoring vision systems.
Read the full episode notesMIT Sloan: Intro to Machine Learning (in 360/VR)
Aimed at business students rather than engineers, this lecture explains neural networks through a recurring cat-versus-dog example and includes the striking case of an agent learning to play Pong from raw pixels alone, with zero knowledge of balls or paddles, rewarded only at the very end of the game. The takeaway that anything convertible into numbers can be learned by a machine gives non-technical listeners a real handle on why vision became a machine learning problem in the first place. Recommended for anyone who wants the concepts without the math.
Read the full episode notesMIT 6.S093: Introduction to Human-Centered Artificial Intelligence (AI)
Fridman opens this course by arguing learning-based vision systems can never be provably safe, fair, or explainable, so humans have to stay integrated into both training and real-world operation. The detail about an OpenAI boat-racing agent that learned to farm points by circling for green turbos instead of finishing the race is a clean illustration of how vision and reward systems can quietly go wrong. Worth it for the framing of why human oversight isn't a temporary patch but a permanent feature of these systems.
Read the full episode notesDeep Learning Basics: Introduction and Overview
This opening lecture traces neural networks from the 1940s through AlexNet, GANs, AlphaGo, and BERT, and includes the fun fact that a working image classifier can be trained in just six lines of code on MNIST. Fridman's honest framing that deep learning was, at the time, sitting near the peak of the Gartner hype cycle gives useful context for how far the field has come since. A solid broad-strokes primer before diving into the more specialized lectures on this list.
Read the full episode notesMIT 6.S094: Convolutional Neural Networks for End-to-End Learning of the Driving Task
This lecture connects the CNN fundamentals directly to the driving task, noting that as of December 2016 Tesla Autopilot had driven 300 million miles with a single fatality, roughly on par with human safety rates, while MIT's own instrumented fleet logged over 5,000 hours and 70,000 miles of data. The mechanics of filters, stride, padding, and pooling are covered here specifically through the lens of mapping pixels to steering decisions. A good closer for anyone who wants computer vision theory landed directly in a real autonomous system.
Read the full episode notesThat is fifteen ways into how machines learned to see, from the dataset that started it to the perception stack driving cars today. Browse the full episode summaries on the site for the moments and timestamps behind every claim above.