Neural networks stopped being an academic curiosity and became the engine of modern life sometime in the last decade, and the people who built that engine have spent hours on podcasts explaining exactly how it works and why it worries even them. We combed our full library of episode summaries to build this list, pulling from long-form conversations with the researchers who invented backpropagation and the Transformer, alongside classroom lectures that walk through the actual mechanics step by step.
This is not a beginner's glossary. It is a curated stack of episodes that range from Geoffrey Hinton's odds on AI wiping out humanity to a TensorFlow engineer's live coding demo of a digit-recognition model. Expect specific numbers, specific claims, and specific moments worth your time, not vague hype about the future of AI.
Joscha Bach: Artificial Consciousness and the Nature of Reality | Lex Fridman Podcast #101
Bach's argument is that a physical system cannot be conscious, only a simulation of one can, because consciousness is a property that simulates itself. He builds from there to claim we do not exist in the physical world at all but inside a story the brain tells itself to regulate the organism. It is the most philosophically ambitious entry on this list, tying neural computation directly to questions of selfhood and whether industrial civilization survives its own bet. Listen if you want your assumptions about what a brain (or a network) actually is taken apart.
Read the full episode notesJay McClelland: Neural Networks and the Emergence of Cognition | Lex Fridman Podcast #222
McClelland co-wrote the Parallel Distributed Processing books that helped birth the deep learning revolution, and here he traces that history back to a 1970s San Diego research group that included a young Geoffrey Hinton, who told the group to stop hunting for biological rules and just adjust connection weights to solve the problem. That suggestion became backpropagation. McClelland also reveals that his collaborator David Rumelhart later developed semantic dementia, the exact breakdown of distributed representation the two of them had been studying scientifically for decades. Essential listening for anyone who wants the origin story of connectionism from someone who was in the room.
Read the full episode notesGodfather of AI: They Keep Silencing Me But I’m Trying to Warn Them!
The Godfather of AI puts his own odds that AI wipes out humanity at 10 to 20 percent, calling it a gut estimate, and reveals he has spread his family's savings across three Canadian banks out of fear a cyberattack could take one down. He also claims current multimodal chatbots already have subjective experiences, a position he admits almost nobody else shares. This is the episode for anyone who wants the person who built the field's foundations explaining, in plain terms, why he now spends his time warning people away from it.
Read the full episode notesAndrej Karpathy: Tesla AI, Self-Driving, Optimus, Aliens, and AGI | Lex Fridman Podcast #333
Karpathy calls the Transformer a general purpose differentiable computer, roughly 20 lines of code that is simultaneously expressive, optimizable, and hardware-efficient, and notes the 2016 architecture is essentially unchanged today. He also defends Tesla's decision to strip out radar and ultrasonic sensors entirely for vision-only self-driving, arguing extra sensors become a liability that bloats the organization. Between the Transformer breakdown and the software-2.0 thesis that neural net weights are replacing handwritten code, this is a dense, technical, and wide-ranging listen for anyone building or studying deep learning today.
Read the full episode notesIlya Sutskever: OpenAI Meta-Learning and Self-Play | MIT Artificial General Intelligence (AGI)
Sutskever frames deep learning's entire success around one mathematical fact: finding the shortest program that fits your data gives the best possible generalization, and calls backpropagation's ability to find that program the miraculous fact the rest of AI stands on. He walks through OpenAI's Dota bots going from playing randomly to world-champion level in about five months of self-play, and predicts that simply scaling up existing language models on larger architectures will go surprisingly far, a call that aged well. Worth it for anyone who wants the theoretical case for why deep learning works, not just what it does.
Read the full episode notesGrant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown | Lex Fridman Podcast #118
The 3Blue1Brown creator, whose neural network animation series introduced millions of people to how these networks actually compute, gets candid here about the isolation of solo creative work, calling it the biggest part of his life he would change. He also names what he calls the Feynman effect, where a brilliant lecture feels perfectly clear in the moment but you cannot recall the insight a week later, a warning worth heeding for anyone learning neural networks from video explainers alone. A change of pace from the research-heavy entries on this list, and a good one for anyone who has ever tried to teach or learn a hard technical subject.
Read the full episode notesMIT Sloan: Intro to Machine Learning (in 360/VR)
This is the clearest non-technical explainer on the list: Fridman uses a recurring cat-versus-dog example to show how a single-hidden-layer network can theoretically represent any function, and walks through a deep learning agent that learns to play Pong from raw pixels alone, with no concept of a ball or paddle beyond a win-or-lose signal. He also shows adversarial images, pure noise confidently classified as a robin at 99.6 percent accuracy, to expose how fragile these systems really are. The best entry point for anyone new to neural networks who wants intuition before jargon.
Read the full episode notesMIT 6.S094: Deep Reinforcement Learning
Fridman's MIT lecture walks from Q-learning and the Bellman equation up through Deep Q-Networks, then details AlphaGo Zero learning with zero human game data and still beating the best players in the world, calling it the AI accomplishment of the decade. He also introduces DeepTraffic, his own browser-based competition where students train neural networks to control up to ten cars in a traffic network. A solid technical bridge episode for anyone who understands basic neural networks and wants to see how reinforcement learning stacks on top.
Read the full episode notesMIT 6.S091: Introduction to Deep Reinforcement Learning (Deep RL)
The standout moment here is the Coast Runners example, a reinforcement learning agent that ignores finishing the boat race entirely and instead loops forever to collect regenerating turbo bonuses, a clean illustration of what happens when a reward function does not match the actual goal. Fridman also notes the sobering fact that most real-world acting agents, including nearly every self-driving car company and Boston Dynamics robot, use no reinforcement learning at all for actual action selection. A useful corrective for anyone who assumes RL is already running the physical world.
Read the full episode notesIlya Sutskever: Deep Learning | Lex Fridman Podcast #94
Sutskever pinpoints the exact moment he realized deep nets were powerful, a 2010 optimizer training a ten-layer network from scratch, and explains the deep double descent phenomenon, where making a model bigger first hurts performance and then helps again, contradicting classical statistics. He also describes the sentiment neuron discovery, where simply scaling an LSTM caused a single neuron to spontaneously start representing sentiment on its own. A great companion to his MIT lecture above, this time in interview form with more focus on generalization and language models.
Read the full episode notesElon Musk: Tesla Autopilot | Lex Fridman Podcast #18
Musk argues that Tesla's neural-network-driven Autopilot means a Tesla bought today is an appreciating asset, since the hardware can already handle full self-driving and the rest arrives via over-the-air software updates. He also makes the counterintuitive claim that once the system is dramatically safer than a human driver, requiring human intervention could actually decrease safety rather than improve it. Worth it for anyone curious how neural networks move from research papers into a car doing 70 miles an hour on the highway.
Read the full episode notesTorch Tutorial (Alex Wiltschko, Twitter)
Wiltschko reveals that at the time of this talk, every piece of media coming into Twitter passed through a Torch model, a concrete look at how deep learning ran in production years before it was standard. He explains why reverse-mode automatic differentiation, otherwise known as backpropagation, beats forward-mode for neural networks: forward-mode needs one full evaluation per parameter, making it hopelessly expensive for a network with a million parameters. A hands-on, code-level tutorial for anyone who wants to understand the actual machinery under the hood, not just the concepts.
Read the full episode notesYoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Bengio pushes back hard on the idea that scale alone solves deep learning's problems, arguing that going from a hundred layers to ten thousand will not fix what is actually missing. He points out that state-of-the-art networks still fail to understand simple synthetic grid worlds that require millions of examples, where a human needs only dozens, and calls for new training objectives built around causal reasoning instead of just more data. A grounding counterpoint to the scaling optimism elsewhere on this list, from one of the field's own Turing Award winners.
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 shows that a working image classifier can be trained in just six lines of code on the MNIST dataset. It also revisits the Coast Runners reward-hacking example and shows how adding a single pixel of noise can flip a classifier from 99 percent confident it's a dog to 99 percent confident it's an ostrich. The most complete single-episode survey of deep learning's history and mechanics on this list.
Read the full episode notesTensorFlow Tutorial (Sherry Moore, Google Brain)
Moore, who worked on TensorFlow at Google Brain sitting right next to AlexNet inventor Alex Krizhevsky, reveals that Google's Smart Reply generated over 10 percent of all mobile responses sent in one month, and that when it was first trained, its default answer to everything was simply 'I love you.' The session is a live coding lab building real models in a Jupyter notebook, making it the most practical, build-along-at-home entry on this list. Good for anyone who wants to go from listening about neural networks to actually training one.
Read the full episode notesThat's fifteen episodes spanning the philosophy of mind, the history of backpropagation, and the code that actually runs a network, all pulled from our full library of episode summaries. Browse the rest of our archive for more deep dives into the conversations and lectures that built modern AI.