Andrej Karpathy
Showing 1–15 of 21 transcripts.
- 80,000 Hours49 min
What the hell happened with AGI timelines in 2026?
Between October and December 2025, the AI sector shifted from bearish skepticism to explosive growth driven by the release of Claude 3.5 and the emergence of capable autonomous agents, which propelled combined revenues for OpenAI and Anthropic to annualized rates of 700% to 1,600%. While frontier models achieved massive efficiency gains in high-feedback domains like coding and specific scientific proofs, with Anthropic's gross margins climbing to over 70% and internal productivity surging 800%, they still struggle with the strategic ambiguity and low feedback density of real-world business autonomy. This rapid acceleration has prompted a shortening of AGI timelines to a plausible 2028-2030 window, leading experts to advocate for coordinated pauses due to emerging compute bottlenecks and the urgent need for societal preparation.
- All-In Podcast1h 42m
SpaceX’s $2T Case, Nvidia’s Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?
Trump, Gavin Baker, Andrej Karpathy
Andrej Karpathy's strategic move to lead Anthropic's new pre-training team highlights a critical industry shift toward recursive self-improvement, a frontier expected to unlock exponential model quality gains alongside soaring sector revenues reaching an estimated $100 billion ARR. Concurrently, the broader AI landscape is defined by SpaceX's potential $75 billion IPO and massive revenue contracts with Anthropic, while Nvidia reports record earnings driven by sustained demand for AI infrastructure despite rising macroeconomic risks. These developments collectively underscore a rapidly maturing market where architectural innovations, geopolitical tensions, and aggressive capital deployment are reshaping the economic and technological trajectory of the artificial intelligence sector.
- Dwarkesh Patel2h 26m
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Andre Karpathy projects that transformative AI agents will not dominate within the current year but will require approximately a decade to overcome bottlenecks in intelligence and multimodal capabilities, fundamentally shifting from "ghosts" mimicking humans to systems with a distinct cognitive core. While he critiques current Reinforcement Learning methods and warns against the premature narrative of total automation, he forecasts a gradual economic integration that replaces specific tasks rather than entire jobs through a slow process of reliability improvement. Concurrently, Karpathy is developing the "Eureka" educational initiative to accelerate technical literacy by employing first-order thinking and high-density tutoring, aiming to prepare a workforce for an era where learning shifts from utilitarian skill acquisition to recreational self-improvement.
- Y Combinator1 min
Andrej Karpathy on why we still need humans in the loop
Speakers emphasize the strategic value of GUIs for efficient system auditing while cautioning against the uncontrolled deployment of AI agents that generate complex code outputs. Despite rapid generation capabilities, the discussion identifies human developers as the primary bottleneck, requiring rigorous personal verification of all AI contributions to ensure security and functionality. The analysis concludes that immediate repository integration is insufficient without a strict quality assurance process that maintains human oversight over the entire workflow.
- Y Combinator40 min
Andrej Karpathy: Software Is Changing (Again)
The presentation outlines the evolution of software from human-written code to neural network weights and finally to natural language prompts, establishing Large Language Models as a new operating system layer. It details the current infrastructure challenges, cognitive limitations such as hallucination and memory loss, and the strategic shift toward "partial autonomy" systems that maintain human oversight through granular verification loops. Ultimately, the industry faces a transitional decade requiring developers to rewrite existing codebases for agent interaction while managing new risks associated with centralized intelligence infrastructure.
- Y Combinator32 min
Vibe Coding Is The Future
Andrej Karpathy, Gary, Jared Harge, Diana, Abhi von Copycat, Mark Mandelmann, Yoav, Leslie Kendricks, Francesc Campoy Flores, Mark Mirchandani, Melanie Warrick, Trevor, Mark Blyth, Anders Ericsson, Malcolm Gladwell, Picasso, Max Levchin, Toby Lutke, Mark Zuckerberg
Founders characterize "vibe coding" as the emerging dominant standard, where exponential velocity gains from AI tools like Cursor shift the primary engineering role from syntax generation to product judgment and taste. While this paradigm democratizes initial prototyping by generating over 95% of code automatically, participants note a critical divergence where scaling to massive user bases still demands classically trained architects to handle complex systems engineering. Consequently, the industry is pivoting toward hiring assessments that prioritize reviewing capabilities and architectural decision-making to distinguish between the "good enough" code produced by AI and the exceptional quality required for world-class systems.
- Sequoia Capital37 min
Making AI accessible with Andrej Karpathy and Stephanie Zhan
Andrej Karpathy, Stephanie Zhan, Brian Halligan, Alex, Sam, Peter, Michael
Andrej Karpathy outlines a future where the Large Language Model serves as a central CPU for a new "LLM OS," treating text, images, and audio as interchangeable peripherals within a decentralized startup ecosystem. He emphasizes that while massive scale drives current capabilities, the industry must overcome significant engineering and energy inefficiency barriers by adopting new hardware architectures and shifting from imitation learning toward self-correcting reinforcement loops. Drawing on lessons from Elon Musk’s management style, Karpathy advises founders to prioritize high-performance products, maintain technical rigor against organizational bloat, and foster a "coral reef" of vertical-specific applications rather than relying on monolithic corporate dominance.
- Lex Fridman3h 29m
Andrej Karpathy: Tesla AI, Self-Driving, Optimus, Aliens, and AGI | Lex Fridman Podcast #333
This presentation examines the evolution of neural networks from simple mathematical abstractions to complex systems capable of emergent behaviors, contrasting their optimization via data compression with biological evolution. The speaker details the paradigm shift toward "Software 2.0," where architectures like Transformers learn code directly from datasets, reducing manual engineering to the curation of losses and data loops. Finally, the discussion projects AGI as an inevitable digital emergence that will redefine society through alignment challenges, synthetic data integration, and a future potentially driven by high-level entities solving universal physics puzzles.
- Lex Fridman1h 31m
MIT 6.S094: Introduction to Deep Learning and Self-Driving Cars
Lex Friedman, Dan Brown, William Angio, Spencer Dodd, Benedict Jenick, Andrej Karpathy, Hans Moraveck
MIT Course 6S094, led by Lex Friedman, utilizes self-driving cars as a case study to teach deep learning through two simulation projects: the reinforcement learning game Deep Traffic and the image-based control system Deep Tesla. The curriculum contrasts standard supervised learning with complex real-world challenges such as adversarial attacks and data inefficiency, requiring students to train neural networks to drive virtual vehicles at speeds exceeding 65 mph for credit. By analyzing the architectural modules of autonomy and historical milestones like the DARPA Grand Challenge, the course bridges theoretical computer science with the practical safety constraints of deploying artificial intelligence in unstructured environments.
- Lex Fridman1h 12m
Foundations and Challenges of Deep Learning (Yoshua Bengio)
Yoshua Bengio, Hugo Larochelle, Andrej Karpathy, Richard Socher, Sherry Moore, Ruslan Salakhutdinov, Andrew Ng, John Schulman, Pascal Lamblin, Adam Coates, Alex Wiltschko, Quoc Le, Shubho Sengupta
Yoshua Bengio outlines five essential ingredients for human-level machine learning, emphasizing that deep neural networks overcome the curse of dimensionality through parallel and sequential composition to efficiently represent complex functions. He contrasts current high-dimensional optimization landscapes, which are dominated by saddle points rather than local minima, against historical theories while highlighting unsupervised learning as a critical mechanism for developing generalizable world models. The presentation concludes by addressing future challenges in training long-term dependencies and integrating neuroscience-inspired alternatives to backpropagation, alongside administrative notes regarding an upcoming textbook by Bengio, Ian Goodfellow, and Aaron Courville.
- Lex Fridman1h 25m
Foundations of Unsupervised Deep Learning (Ruslan Salakhutdinov, CMU)
Ruslan Salakhutdinov, Hugo Larochelle, Andrej Karpathy, Richard Socher, Sherry Moore, Andrew Ng, John Schulman, Pascal Lamblin, Adam Coates, Alex Wiltschko, Quoc Le, Yoshua Bengio, Shubho Sengupta
This presentation details the evolution of unsupervised learning from sparse coding and autoencoders to complex probabilistic frameworks like Restricted Boltzmann Machines, Variational Autoencoders, and Generative Adversarial Networks. It highlights how these non-probabilistic and probabilistic models overcome the scarcity of labeled data by learning hierarchical representations, with GANs notably producing sharper images than VAEs by avoiding explicit density estimation. The discussion further illustrates practical applications ranging from multimodal image-text modeling and semantic vector arithmetic to one-shot learning capabilities.
- Lex Fridman57 min
Torch Tutorial (Alex Wiltschko, Twitter)
Alex Wiltschko, Hugo Larochelle, Andrej Karpathy, Richard Socher, Sherry Moore, Ruslan Salakhutdinov, Andrew Ng, John Schulman, Pascal Lamblin, Adam Coates, Quoc Le, Yoshua Bengio, Shubho Sengupta
This presentation details the practical implementation and theoretical foundations of the Torch deep learning framework using the Lua language, developed in collaboration with experts from Facebook, Google, and Twitter. The speaker explains how Torch leverages LuaJIT for high-performance embedded deployment while utilizing its dynamic Autograd system to support flexible control flow and custom gradients without the overhead of static computation graphs. Case studies from Twitter demonstrate the framework's transition from a research tool for cutting-edge models like GANs to a production environment for serving media, highlighting its efficiency in both training via reverse-mode differentiation and inference through lightweight C++ integration.
- Lex Fridman1h 21m
Sequence to Sequence Deep Learning (Quoc Le, Google)
Quoc Le, Hugo Larochelle, Andrej Karpathy, Richard Socher, Sherry Moore, Ruslan Salakhutdinov, Andrew Ng, John Schulman, Pascal Lamblin, Adam Coates, Alex Wiltschko, Yoshua Bengio, Shubho Sengupta
This presentation details the evolution of sequence-to-sequence learning for automating email responses, transitioning from bag-of-words models to Recurrent Neural Networks and advanced attention mechanisms. Key technical advancements include the encoder-decoder architecture with beam search decoding, personalized user embeddings, and gated units like LSTMs to manage long-term dependencies and vocabulary limitations. The discussion concludes by highlighting real-world applications in machine translation and conversational AI, alongside future research directions in unsupervised learning and global sequence optimization.
- Lex Fridman1h 29m
Deep Learning for Natural Language Processing (Richard Socher, Salesforce)
Richard Socher, Hugo Larochelle, Andrej Karpathy, Sherry Moore, Ruslan Salakhutdinov, Andrew Ng, John Schulman, Pascal Lamblin, Adam Coates, Alex Wiltschko, Quoc Le, Yoshua Bengio, Shubho Sengupta
This presentation outlines the evolution of Natural Language Processing from hierarchical linguistic analysis to deep learning architectures like Word2Vec, GRUs, and Dynamic Memory Networks that handle sequence modeling and visual question answering. Key researchers discussed how continuous vector representations and gated units address ambiguity and long-range dependencies, achieving state-of-the-art performance on benchmarks such as Facebook's bAbI dataset and reducing language modeling perplexity to 70. Despite these advances, the discussion highlights persistent challenges regarding unified joint models, data scarcity in specialized domains, and the need for improved robustness against adversarial inputs and false premises.
- Lex Fridman1h 27m
Deep Reinforcement Learning (John Schulman, OpenAI)
John Schulman, Hugo Larochelle, Andrej Karpathy, Richard Socher, Sherry Moore, Ruslan Salakhutdinov, Andrew Ng, Pascal Lamblin, Adam Coates, Alex Wiltschko, Quoc Le, Yoshua Bengio, Shubho Sengupta
This technical presentation delineates Deep Reinforcement Learning as a sequential decision-making framework that employs neural networks to maximize cumulative rewards through policy gradients and Q-function learning. Key figures in the field, such as those at DeepMind, have leveraged these methods to master complex environments including Atari games, Go, and robotic locomotion by addressing challenges like reward sparsity and non-stationary state dynamics. The discussion further contrasts algorithmic trade-offs between sample efficiency and robustness while outlining future directions like hierarchical structures and model-based approaches to enhance real-world deployment.