Deep learning has produced more genuinely strange, world-changing stories than almost any other technical field, and the people who built it turn out to be excellent talkers. We combed through our full library of episode summaries to find the conversations that actually explain how the field works: the failed bets, the accidental discoveries, the arguments the founders still have with each other about whether any of this adds up to real intelligence.
This list skips the hype and goes straight to the specifics: the napkin sketch that convinced Andrew Ng to bet on scale, the patent Yann LeCun spent five years hoping nobody would notice had expired, the single beer-fetching robot demo that was secretly a person in a headset. Fifteen episodes, one clear picture of how deep learning actually got built.
Yann LeCun: Deep Learning, ConvNets, and Self-Supervised Learning | Lex Fridman Podcast #36
LeCun opens by comparing AI objective functions to legal codes and closes with the argument that human intelligence isn't actually general, just highly specialized. In between he explains why neural nets fell out of favor in the mid-90s (bad tooling, easy beginner mistakes, and AT&T lawyers blocking an open-source release) and admits he spent 2002 to 2007 quietly hoping nobody at NCR would notice the LeNet patent had expired. Listen for the clearest case anyone makes for why reasoning needs a world model, not just a bigger network.
Read the full episode notesEric Schmidt — The Promises and Perils of AI, the Future of Warfare, Profound Revolutions, and More
The former Google CEO walks through AlphaGo inventing Go moves unknown in 2,500 years of the game's history, and the discovery of Halicin, the first new broad-spectrum antibiotic candidate in roughly 40 years, found by screening 100 million compounds with AI. He also lays out a war-game scenario where an attack, a countermove, and a counterattack all happen inside 10 milliseconds, breaking every assumption in current military doctrine. For anyone who wants the geopolitical stakes explained without the sci-fi filter.
Read the full episode notesOriol Vinyals: Deep Learning and Artificial General Intelligence | Lex Fridman Podcast #306
DeepMind's Oriol Vinyals explains Gato, the generalist agent that handles text, images, and robot actions with a single one-billion-parameter model, tiny compared to the trillion-parameter models of its era. He details how Flamingo froze a 70-billion-parameter language model and bolted on new parameters to give it sight, and how that resulting system understood a joke image that Andrej Karpathy once said no computer vision system could grasp. Essential for anyone trying to understand why one shared architecture is swallowing every AI subfield.
Read the full episode notesJeff Hawkins: Thousand Brains Theory of Intelligence | Lex Fridman Podcast #25
Hawkins makes the case that deep learning's point-neuron model is far too simple, and that real neurons act as predictive engines using thousands of synapses. His Thousand Brains Theory argues there is no single model of an object in your head at all: thousands of cortical columns each build a complete model and vote on what you're looking at. He even explains how the neocortex repurposed ancient navigation circuitry (grid cells and place cells) to model abstract ideas. For anyone who thinks the path to real AI runs through neuroscience, not more GPUs.
Read the full episode notesAndrew Ng: Deep Learning, Education, and Real-World AI | Lex Fridman Podcast #73
Ng recounts filming early Coursera lectures alone at 10pm with a webcam and a Wacom tablet, then admits the field's early mistake: overweighting unsupervised learning when supervised learning and raw scale were what actually worked. The turning point was a single chart from Adam Coates showing bigger models perform better, which gave Ng the conviction to pitch Google Brain even as well-meaning friends warned him scale was a bad career bet. Worth it for the honest look at how research bets actually get made, and his argument that near-term bias and inequality deserve more attention than distant AGI fears.
Read the full episode notesSergey Levine: Robotics and Machine Learning | Lex Fridman Podcast #108
Levine argues the human-robot gap is a mind gap, not a hardware gap, and that with enough money and engineering the hardware problem nearly disappears. He opens by debunking a famous 2004 Stanford video of a robot tidying a living room and fetching a beer: it was entirely teleoperated by a person. His real argument is that robotics is the best vehicle for understanding intelligence itself, because it forces a system to build common sense by actually interacting with the world. Recommended for anyone who thinks reinforcement learning research is too abstract to matter.
Read the full episode notesFrançois Chollet: Keras, Deep Learning, and the Progress of AI | Lex Fridman Podcast #38
The creator of Keras pushes back hard on the intelligence-explosion narrative, arguing that recursively self-improving systems hit exponential friction, not takeoff, citing a study that found scientific progress has stayed flat for 150 years even as paper output exploded. He also tells the origin story of Keras, built in February 2015 mainly as a reusable LSTM implementation, before TensorFlow's Rajat Monga recruited him to fold it into Google's framework. A sharp corrective for anyone who takes AGI timelines as settled fact.
Read the full episode notesDavid Silver: AlphaGo, AlphaZero, and Deep Reinforcement Learning | Lex Fridman Podcast #86
Silver traces AlphaGo from a childhood BBC Micro to the 2016 win over Lee Sedol, revealing he predicted the exact 4-1 scoreline based on data showing AlphaGo developed a 'delusion' in roughly one of every five games, the same flaw Sedol exploited in his lone win. He also covers the leap to AlphaZero, which learned purely from self-play with zero human data, and MuZero, which learns without even being told the rules. The story behind Move 37, the play that broke every convention Go players are taught, is the clearest case anyone has made for machine creativity.
Read the full episode notesMax Tegmark: Life 3.0 | Lex Fridman Podcast #1
In the inaugural episode of the Lex Fridman Podcast, MIT physicist Max Tegmark argues intelligence is substrate-independent information processing with no biological secret sauce, and that finding no life on Mars would actually be encouraging news for humanity's future. He walks through why an AGI doesn't necessarily need self-preservation instincts, since the space of possible minds is far larger than what evolution produced, and pushes for 'kindergarten ethics' as a starting point for AI alignment rather than waiting for perfect philosophical consensus. A foundational listen for the alignment debate.
Read the full episode notesSequence to Sequence Deep Learning (Quoc Le, Google)
Google's Quoc Le builds sequence-to-sequence learning from scratch, starting from a yes/no email auto-reply classifier he needed after 508 emails piled up on vacation, all the way to the encoder-decoder architecture behind translation and image captioning. He reveals that Smart Reply in Gmail is actually two models working together, one deciding whether a reply is warranted and a second generating candidate replies via beam search. A rare inside look at how a lab toy becomes a feature used by hundreds of millions of people.
Read the full episode notesIlya Sutskever: Deep Learning | Lex Fridman Podcast #94
OpenAI's chief scientist pinpoints the exact moment he realized deep nets were powerful: watching a 2010 Hessian-free optimizer train a 10-layer network from scratch. He walks through the deep double descent phenomenon, where making a model bigger first hurts performance and then helps again, directly contradicting classical statistics, and calls deep learning 'the geometric mean of biology and physics.' Essential listening for understanding why the field's biggest names still can't fully explain why their own systems work.
Read the full episode notesYoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Bengio argues that simply stacking more layers won't fix deep learning's core problems, pointing out that state-of-the-art networks still fail at simple synthetic grid-world tasks that humans solve in dozens of examples instead of millions. He dismisses AI existential risk as very unlikely while flagging the dangers he actually loses sleep over: surveillance, autonomous weapons, and concentration of power. Worth hearing for a Turing Award winner's blunt take on catastrophic forgetting and why causal reasoning, not scale, is the field's real missing piece.
Read the full episode notesJohn Hopfield: Physics View of the Mind and Neurobiology | Lex Fridman Podcast #76
The physicist behind Hopfield networks argues evolution turns biological 'glitches,' like neurons syncing up the way pedestrians unconsciously fell into step and made London's Millennium Bridge sway, into features that artificial neural networks completely ignore. He claims the brain is not actually as deep as today's deepest computer-science networks, and that depth may just be an artifact of how machines are easiest to train. A genuinely strange, physics-first perspective on where AI's biological metaphors break down.
Read the full episode notesDeep Learning for Computer Vision (Andrej Karpathy, OpenAI)
This lecture traces computer vision from Hubel and Wiesel's 1960s cat experiments through the 2012 AlexNet breakthrough, and reveals that in roughly 20 years the two biggest algorithmic advances over the original LeNet were dropout and ReLU, both essentially one-line changes. Karpathy also built a web tool to test his own accuracy against a CNN on ImageNet and lost points mostly on dog-breed identification, a detail that says a lot about how granular these benchmarks really are. The best single primer on how CNNs actually work, mechanically.
Read the full episode notesJeremy Howard: fast.ai Deep Learning Courses and Research | Lex Fridman Podcast #35
The fast.ai founder describes how a team of about four of his students beat Google and Intel on the Stanford DAWNBench competition using a single-GPU setup, training on tiny 64x64 images first before fine-tuning on full-size ones. He argues most deep learning research is wasted effort compared to giving domain experts practical tools, and notes that outside South Africa the entire African continent has only five pediatric radiologists, the kind of gap he thinks transfer learning should be closing. Recommended for anyone who thinks state-of-the-art results require a data center.
Read the full episode notesThat's fifteen conversations spanning the people who built convolutional nets, the ones betting against the current AGI hype, and the ones convinced the brain still has more to teach us than any dataset. Browse the full library of episode summaries on Episode Notes for hundreds more reveals like these, organized by show and guest.