Andrej Karpathy is an artificial intelligence researcher, educator, and entrepreneur known for contributions to deep learning, computer vision, and large language models. He has held leadership roles at OpenAI and Tesla, developed widely used educational resources, and created open-source tools that have influenced both research and industry practice. His work spans foundational models, autonomous systems, and AI-native education platforms.

Early life and education

Karpathy was born in Bratislava, Slovakia, and moved to Toronto at age 15.^[S25] He completed undergraduate studies in computer science and physics at the University of Toronto.^[S25] He earned a PhD in computer science from Stanford University in 2015, advised by Fei-Fei Li, with a dissertation titled Connecting Images and Natural Language focused on image captioning and multimodal modeling.^[S25]

Career

Karpathy was a founding member of OpenAI in December 2015, joining the original research team alongside Ilya Sutskever and Wojciech Zaremba.^{[S9, S26]} He contributed to early reinforcement learning and generative model research, including work that led to GPT-1.^[S26] He left OpenAI in 2017 to become Director of AI at Tesla, where he led the Autopilot vision stack for five years.^[S27] During his tenure at Tesla, he architected HydraNet, a multi-task vision network, and drove the transition toward end-to-end neural-network-only driving systems.^[S36] He returned to OpenAI in 2023 to lead a small team focused on midtraining and synthetic data generation, followed by a brief sabbatical in 2024 before resuming his role.^{[S2, S8]}

Notable projects and contributions

Karpathy has created and maintained several influential open-source projects and educational resources. nanoGPT is a minimal implementation for training and fine-tuning medium-sized GPT models, designed to be accessible and reproducible.^[S39] micrograd is a tiny autograd engine in pure Python that illustrates neural network fundamentals.^[S41] char-rnn introduced character-level recurrent neural networks for language modeling and became a popular entry point for learning about generative models.^[S42] ConvNetJS enabled browser-based training of convolutional neural networks and helped popularize deep learning among web developers.^[S44]

His educational initiatives include Neural Networks: Zero to Hero, a YouTube course that builds deep learning systems from scratch, and CS231n: Convolutional Neural Networks for Visual Recognition, Stanford’s foundational deep learning course which he taught and whose materials remain widely used.^{[S40, S45]} He also developed arxiv-sanity, a tool for filtering and ranking research papers by similarity to a user’s interests.^[S43]

Karpathy has written extensively on conceptual and practical aspects of AI. His 2017 essay Software 2.0 argued that large-scale training of neural networks is replacing hand-written code for many tasks, and his 2016 post Yes you should understand backprop emphasized the importance of deriving backpropagation manually.^{[S34, S29]} He coined the terms Software 2.0 and vibe coding, the latter referring to programming primarily through natural language prompts.^[S1]

AI-native education and Eureka Labs

Karpathy founded Eureka Labs, an AI-native education startup focused on building teaching assistants that can interact with students in real time.^[S38] The company’s first offering is LLM101n, a course that guides learners through building a large language model end to end, from tokenization to deployment.^{[S12, S3]} The course is designed to be accessible to those with basic Python skills and emphasizes hands-on coding over theoretical prerequisites.^[S3]

Public profile and online presence

Karpathy maintains a public profile across multiple platforms. His personal website, karpathy.ai, serves as a hub for his writing, projects, and updates.^[S47] He is active on Twitter/X, where he shares frequent updates and engages with the AI community.^[S46] His YouTube channel hosts the Neural Networks: Zero to Hero series and other technical content.^[S49] He is also active on GitHub, where his repositories have accumulated hundreds of thousands of stars.^[S48]

Views on AI and technology

Karpathy has expressed cautious optimism about the progress and deployment of AI technologies. He emphasizes the importance of compute and data over architectural cleverness, citing Richard Sutton’s The Bitter Lesson as a guiding principle in AI development.^[S15] He distinguishes between pretraining, midtraining, and deployment, arguing that midtraining—teaching base models new skills and behaviors without full retraining—is a key frontier.^[S18] He cautions that while vibe coding is useful for prototyping, it is not yet suitable for production-critical systems.^[S16]

He has also commented on the pace of capability gains, noting that advances from 2020 to 2025 exceeded many predictions, though the gap between capability and societal deployment remains significant.^{[S7, S10]} He advocates for integrating research and engineering, arguing that modern ML labs benefit when researchers write production code and engineers read papers.^[S37]

Personal life

Karpathy is married and resides in the San Francisco Bay Area.^[S25]