Andrej Karpathy

wiki
, , , ,
1

Slovak-Canadian AI researcher, educator, and open-source builder; a founding member of OpenAI, former Director of AI at Tesla, founder of Eureka Labs, and Member of Technical Staff at Anthropic.

Basic Information / Profile

Field Details
Full Name Andrej Karpathy
Born October 23, 1986, Bratislava, Czechoslovakia (now Slovakia)
Nationality Slovak-Canadian
Current Institution Anthropic
Research Fields Deep learning, computer vision, natural language processing, large language model pre-training
PhD Advisor Fei-Fei Li
PhD Thesis Connecting Images and Natural Language (Stanford University, 2016)
Personal Website karpathy.ai
X / Twitter @karpathy
GitHub github.com/karpathy
Google Scholar scholar.google.com/citations?user=l8WuQJgAAAAJ

Overview

Andrej Karpathy is a Slovak-Canadian AI researcher and educator widely regarded as one of the most influential figures in modern deep learning. A founding member of OpenAI, he later served as Director of AI at Tesla, where he oversaw the computer vision systems underpinning the Autopilot and Full Self-Driving programs. He is the architect of CS 231n, Stanford’s first deep learning course, which grew into one of the university’s most-attended classes and became a reference curriculum worldwide. In May 2026, he joined Anthropic as a Member of Technical Staff focused on pre-training research, tasked with building a team that uses Claude to accelerate the science of large-scale training runs. Beyond his research roles, Karpathy is one of the most-followed AI educators on YouTube and the coiner of the term vibe coding, which Collins Dictionary selected as its Word of the Year.

Early Life & Education

Childhood and Move to Canada

Karpathy was born on October 23, 1986, in Bratislava, then part of Czechoslovakia (present-day Slovakia). At the age of 15 his family relocated to Toronto, Canada, where he completed secondary school. Even before university, he had developed an interest in computation and competitive hobbies: beginning in 2006 he ran a YouTube channel under the handle badmephisto, posting Rubik’s cube speed-solving tutorials that accumulated over 9 million views and influenced prominent speedcubers including world-record holder Feliks Zemdegs.

University of Toronto (2005–2009)

Karpathy completed a Bachelor of Science with a double major in Computer Science and Physics and a minor in Mathematics at the University of Toronto in 2009. During this period he first encountered deep learning by attending Geoffrey Hinton’s lectures and reading groups, an experience he cites as foundational to his research trajectory.

University of British Columbia (2009–2011)

He pursued a Master of Science at the University of British Columbia under the supervision of Michiel van de Panne, focusing on learning controllers for physically simulated figures — in effect, applying machine learning to agile character animation and robotics in simulation. This work included a SIGGRAPH 2011 co-authored paper on locomotion skills for simulated quadrupeds.

Stanford University PhD (2011–2015)

Karpathy joined Stanford’s Vision Lab to pursue a PhD under Fei-Fei Li, completing his degree in 2015 (with the thesis formally submitted in 2016). His dissertation, Connecting Images and Natural Language, explored the intersection of convolutional and recurrent neural networks for tasks such as image captioning and visual-semantic alignment. During the PhD he completed three research internships: Google Brain (2011, unsupervised learning from video), Google Research (2013, large-scale supervised learning on YouTube), and DeepMind (2015, deep reinforcement learning with Koray Kavukcuoglu and Vlad Mnih). He also designed and taught CS 231n: Convolutional Neural Networks for Visual Recognition, Stanford’s inaugural deep learning course, which enrolled 150 students in 2015 and grew to 750 by 2017.

Career

OpenAI — Founding Researcher (2015–2017)

Karpathy was among the original research scientists named when OpenAI was publicly introduced in December 2015, joining as part of the nonprofit’s founding team. During this period his work centred on deep learning and computer vision. He left in mid-2017 when Tesla recruited him directly.

Tesla — Director of AI (2017–2022)

In June 2017 Karpathy became Tesla’s Director of Artificial Intelligence and Autopilot Vision, reporting to Elon Musk. In this role he led the full computer vision stack for Tesla Autopilot and Full Self-Driving (FSD): in-house data labeling pipelines, neural network training, and deployment on Tesla’s custom inference chip. He presented the team’s architecture publicly at Tesla Autonomy Day 2019 and Tesla AI Day 2021. He briefly also oversaw early work on the Tesla Optimus humanoid robot before departing. After a months-long sabbatical, he announced his resignation in July 2022.

Independent — YouTube & Open-Source (2022–2023)

Between leaving Tesla and rejoining OpenAI, Karpathy published a series of highly influential open-source projects and educational YouTube lectures. This period produced nanoGPT (54k+ GitHub stars), a minimal GPT training framework; llm.c (29k+ stars), a pure C/CUDA implementation of LLM training; and the “Neural Networks: Zero to Hero” video series, a step-by-step curriculum building neural networks from first principles. These materials became widely adopted in university courses and self-study programmes globally.

OpenAI — Return (2023–2024)

On February 9, 2023, Karpathy announced his return to OpenAI, where he joined a team focused on mid-training and synthetic data generation. He departed again on February 13, 2024, with the move confirmed by an OpenAI spokesperson. During this second tenure he delivered the widely-shared talk State of GPT at Microsoft Build 2023.

Eureka Labs — Founder (2024–2026)

On July 16, 2024, Karpathy announced the founding of Eureka Labs, an AI-native education company. The core concept was to pair human instructors with AI teaching assistants capable of personalising instruction at scale. The company’s flagship product was LLM101n: Let’s Build A Storyteller, a course walking students through building a large language model end-to-end. In February 2025 he coined the term vibe coding to describe a programming workflow in which the developer fully delegates code generation to an AI model via natural language prompts; Collins Dictionary named it Word of the Year for 2025.

Anthropic — Member of Technical Staff (2026–present)

On May 19, 2026, Karpathy announced he had joined Anthropic as Member of Technical Staff. He works on the pre-training team under team lead Nick Joseph, with a mandate to build a new research group focused on using Claude to accelerate the science of pre-training itself — a form of AI-assisted research aimed at making future model training more efficient and automated.

Key Contributions

  • CS 231n: Convolutional Neural Networks for Visual Recognition — Designed and taught Stanford’s first dedicated deep learning course in 2015; it became one of the university’s largest classes (150 → 750 students over three years) and its freely available lecture videos and notes remain a primary entry point for practitioners worldwide.

  • “Deep Visual-Semantic Alignments for Generating Image Descriptions” (CVPR 2015, Oral) — An early landmark in vision-language modelling, jointly aligning image regions with sentence fragments using a multimodal recurrent neural network.

  • “The Unreasonable Effectiveness of Recurrent Neural Networks” (blog post, 2015) — A widely read demonstration of character-level language models that became one of the canonical introductions to sequence modelling for a generation of practitioners.

  • char-rnn — An early open-source character-level language model in Lua/Torch that popularised hands-on experimentation with sequence models before the transformer era.

  • Tesla Autopilot Vision Stack — Led the end-to-end redesign of Tesla’s Autopilot perception system away from radar-and-map fusion toward a pure-vision, data-centric approach deployed on custom silicon at scale.

  • “Software 2.0” (Medium, 2017) — An essay framing neural networks as a new programming paradigm in which model weights replace hand-written code, widely cited in discussions of the industry’s structural shift.

  • nanoGPT — A minimal, hackable GPT training repository (54k+ GitHub stars) that made reproducing GPT-class models accessible to researchers and students without large compute budgets.

  • llm.c — A from-scratch implementation of LLM training in C/CUDA (29k+ stars) demonstrating the full algorithm without framework abstractions.

  • micrograd — A 100-line scalar-valued autograd engine with a PyTorch-like API (15k+ stars) widely used as a teaching tool for backpropagation.

  • “Zero to Hero” YouTube series — A free, technically rigorous video curriculum building neural networks and language models from first principles; among the most-viewed technical AI education content available online.

  • “Deep Dive into LLMs like ChatGPT” (YouTube, 2024) — A general-audience lecture on the internals of large language models that reached broad viewership beyond the ML community.

  • Vibe coding — Coined in February 2025 to describe AI-delegated programming via natural language prompts; the term was adopted industry-wide and named Collins Dictionary Word of the Year 2025.

  • arxiv-sanity — A web application for filtering and discovering arXiv preprints by relevance and similarity, used by many researchers to manage the volume of ML publications.

Awards & Recognition

  • MIT Technology Review Innovators Under 35 (2020) — Recognised for his work leading Tesla’s AI and Autopilot vision team.

  • TIME100 Most Influential People in AI (2024) — Named to Time magazine’s inaugural list of the 100 most influential figures in artificial intelligence.

  • Collins Dictionary Word of the Year — “vibe coding” (2025) — A term Karpathy coined in February 2025; Collins selected it as the defining new word of the year.

Key Relationships

  • Fei-Fei Li — PhD advisor at Stanford’s Vision Lab; co-author on several foundational vision-language papers including the CVPR 2015 image captioning work.

  • Geoffrey Hinton — Undergraduate influence; Karpathy credits attending Hinton’s University of Toronto reading groups as the origin of his interest in deep learning.

  • Michiel van de Panne — Master’s supervisor at UBC; introduced Karpathy to physics-based simulation and reinforcement learning for motor control.

  • Elon Musk — Direct reporting relationship during five years as Tesla’s Director of AI; Musk publicly acknowledged Karpathy’s departure in 2022.

  • Sam Altman / OpenAI leadership — Worked under at both periods of OpenAI employment (2015–2017, 2023–2024); Karpathy was part of the original founding cohort.

  • Justin Johnson — Stanford PhD colleague and frequent collaborator; co-authored DenseCap and the Visualizing and Understanding Recurrent Networks paper; co-hosted a deep learning reading group on Clubhouse.

  • Andrew Ng — Stanford colleague whose presence alongside Daphne Koller, Sebastian Thrun, and Vladlen Koltun shaped Karpathy’s PhD environment; later cited Karpathy in discussions of AI education.

  • Nick Joseph — Current team lead at Anthropic overseeing the pre-training department under which Karpathy now operates.

Personal Style

Karpathy combines a strong preference for minimalism and legibility with pedagogical instinct: his most influential open-source projects deliberately strip away framework abstractions to expose the underlying algorithm, a philosophy he articulates explicitly in repositories like micrograd and llm.c. His public writing and lectures favour concrete, runnable examples over theoretical exposition, and he consistently prioritises making frontier concepts accessible to practitioners without specialised hardware. On social media his posts oscillate between technical observations, conceptual coinages (vibe coding, tokenmaxxing), and candid reflections on the pace and direction of the field, giving him a voice that spans both specialist and general audiences. He is known for intellectual directness and a degree of irreverence — his personal website, written in pure HTML and CSS with no frameworks, is itself a statement about technological minimalism.

References