karpathy.ai

Andrej Karpathy is an AI researcher and engineer whose career spans founding roles at OpenAI, directing AI at Tesla, and creating some of the most widely used deep learning educational material online. His personal site, karpathy.ai, is a concise record of that work: it lists his roles, his research focus, and the educational videos he now produces. If you want to understand his influence in AI, the site itself is the primary source, and the short answer is that his impact comes from three things — building Autopilot's vision stack at Tesla, helping found OpenAI, and teaching a generation of practitioners through Stanford's CS 231n and his YouTube series.

What you find on karpathy.ai

The site is a personal homepage rather than a blog or product. Its content is organized around his own summary of his career:

• A one-line description of what he does: training deep neural nets on large datasets.
• A 2024 section pointing to his YouTube channel, split into a technical track (the Zero to Hero playlist) and a general audience track (Deep Dive into LLMs like ChatGPT, How I Use LLMs, and Intro to Large Language Models).
• Pointers to where he spends most of his time: X/Twitter and GitHub.
• A reverse-chronological career timeline from 2011 to 2024.

The page also includes a playful "Order of the Unicorn" passage, which is stylistic rather than informational — don't read it as a factual claim about his work.

Career timeline

The Tesla entry is the one most people encounter first. Per the site, his team owned the full pipeline — labeling, training, and deployment onto Tesla's own inference hardware — and the stated goal was making Full Self-Driving a reality at scale. The site points to the August 2021 Tesla AI Day for more detail.

Research and teaching

His PhD, advised by Fei-Fei Li at the Stanford Vision Lab, focused on convolutional and recurrent neural networks and their applications in computer vision, natural language processing, and the intersection of the two. The site also notes he worked with Daphne Koller, Andrew Ng, Sebastian Thrun, and Vladlen Koltun during Stanford's first-year rotation program.

The teaching side is arguably as influential as the research. He designed and was the primary instructor for Stanford CS 231n: Convolutional Neural Networks for Visual Recognition, which grew from 150 enrolled students in 2015 to 330 in 2016 and 750 in 2017 — one of the largest classes at Stanford. That course became a standard entry point into deep learning for students and practitioners well beyond Stanford.

What he is doing now

As of the 2024 section of his site, his main public output is educational video on his YouTube channel, in two parallel tracks:

• Technical track — the Zero to Hero playlist, aimed at people who want to build and train models.
• General audience track — Deep Dive into LLMs like ChatGPT for under-the-hood fundamentals, and How I Use LLMs for practical usage examples from his own life. A third, older video, Intro to Large Language Models, runs in parallel.

For ongoing updates, the site directs readers to his X/Twitter and GitHub accounts rather than the homepage itself.

How to use this if you're deciding where to start

• If you want to understand how LLMs work — start with Deep Dive into LLMs like ChatGPT on the general audience track.
• If you want to actually build models — go to the Zero to Hero technical playlist.
• If you want his research and professional history — the timeline on karpathy.ai is the authoritative summary, since it is his own account.
• If you want current activity — follow the X/Twitter or GitHub links, as the homepage is a static overview rather than a feed.

One caveat worth keeping in mind: the site is a self-description, so it is reliable for what he claims about his own roles and output, but it is not a neutral third-party evaluation of his impact. For that, you would need to cross-reference external coverage of Tesla Autopilot, OpenAI's early team, and the adoption of CS 231n.

Andrej Karpathy held two distinct roles at OpenAI and one major leadership role at Tesla. According to his personal site, he was a founding member and research scientist at OpenAI from 2015 to 2017, then served as Director of AI at Tesla from 2017 to 2022, and returned to OpenAI from 2023 to 2024 to build a team focused on midtraining and synthetic data generation. The sections below break down each role based on what his site states.

At Tesla (2017–2022): Director of AI

Karpathy led the computer vision team behind Tesla Autopilot. His site describes the scope of that team's work in three parts:

• In-house data labeling — the team handled all of Tesla's own labeling rather than outsourcing it.
• Neural network training — model development ran in-house.
• Deployment on Tesla's custom inference chip — the trained networks were deployed onto Tesla's own hardware.

He also notes he worked on Tesla Optimus "very briefly," so his involvement with the humanoid robot program was limited compared with Autopilot.

On the goal of the work, his site is direct: today Autopilot increases the safety and convenience of driving, but the team's goal was to make Full Self-Driving a reality at scale. He points to the August 2021 Tesla AI Day for more detail.

At OpenAI (2015–2017): Founding Member and Research Scientist

Karpathy was a research scientist and a founding member at OpenAI during its earliest years. His site lists this as a 2015–2017 role, placing him there before he moved to Tesla.

Return to OpenAI (2023–2024): Midtraining and Synthetic Data

He came back to OpenAI for a second period, where his site says he "built a new team working on midtraining and synthetic data generation." This is the most specific description of his second stint — the site does not list a title for this period, only the team's focus areas.

How the Timeline Fits Together

His site presents these as a chronological overview rather than a detailed project log, so for deeper technical context on the Tesla work, the August 2021 Tesla AI Day is the reference he points to.

Andrej Karpathy's AI educational videos are on his YouTube channel, and they are organized into two parallel tracks: a technical track and a general audience track. If you want to build or train models, start with the technical track's Zero to Hero playlist. If you want to understand and use tools like ChatGPT without writing code, start with the general audience videos such as Deep Dive into LLMs like ChatGPT and How I Use LLMs.

The two learning tracks

Karpathy describes the videos as coming in "two parallel tracks," which lets you pick a path based on whether you intend to work with models directly or mainly use them.

Technical track: Zero to Hero

This is the recommended starting point for a hands-on, code-oriented path. Follow the Zero to Hero playlist in order. It is aimed at building up the skills to train deep neural networks, which matches Karpathy's own stated focus: "I like to train deep neural nets on large datasets."

General audience track

Two videos anchor this track:

• Deep Dive into LLMs like ChatGPT — covers the under-the-hood fundamentals of how large language models work.
• How I Use LLMs — a more practical guide with examples of how Karpathy uses LLMs in his own life.

A third, older video sits alongside these: Intro to Large Language Models.

How to choose

• Choose the technical track if your goal is to understand the mechanics well enough to build, train, or fine-tune models. Expect to write and run code.
• Choose the general audience track if your goal is to use LLMs effectively, understand what they can and cannot do, or get practical usage examples without a coding requirement.
• Watch Intro to Large Language Models if you want a shorter, earlier overview before committing to either track.

Where to find updates

For the latest material, Karpathy says he spends most of his time on X/Twitter and GitHub. Those are the places to check for new videos and projects beyond the YouTube channel.

Background that explains the tracks

The split between technical and general audience content reflects Karpathy's range of work. He was a founding member and research scientist at OpenAI (2015–2017), Director of AI at Tesla (2017–2022) leading the Autopilot computer vision team, and returned to OpenAI (2023–2024) to build a team working on midtraining and synthetic data generation. He also designed and was the primary instructor for Stanford's CS 231n: Convolutional Neural Networks for Visual Recognition, which grew from 150 students in 2015 to 750 in 2017. That teaching background is the through-line behind the structured playlists on the site.

karpathy.ai is the personal homepage of Andrej Karpathy, an AI researcher and educator. It is not a commercial product site or a documentation portal — it is a single page that collects his professional history, his AI educational videos, and links to where he is most active online. If you are looking for his courses, his video series on large language models, or a quick overview of his career, this page is the intended starting point.

Who runs the site

The site belongs to Andrej Karpathy himself. The page describes him as someone who likes to "train deep neural nets on large datasets," and it mixes that technical self-description with a deliberately playful personal note about being a member of the "Order of the Unicorn." That tone is a useful signal: the page is a personal introduction, not a corporate or product landing page, so you should read it as a curated self-presentation rather than as official documentation for any tool or service.

What the page actually contains

The content falls into three practical categories.

1. A career timeline

The page lays out his roles in reverse-chronological blocks:

It also notes that he designed and was the primary instructor for Stanford's CS 231n (Convolutional Neural Networks for Visual Recognition), which grew from 150 enrolled students in 2015 to 750 in 2017.

2. AI educational videos

This is the part most visitors come for. The page states that in 2024 he creates educational videos on AI on his YouTube channel, organized into two parallel tracks:

• Technical track — follow the "Zero to Hero" playlist.
• General audience track — "Deep Dive into LLMs like ChatGPT" covers under-the-hood fundamentals, and "How I Use LLMs" is a more practical guide with examples from his own life.
• A third, older video, "Intro to Large Language Models," is listed as a parallel option.

If your goal is to learn how LLMs work, the page itself tells you which playlist matches your background: technical viewers go to Zero to Hero, general viewers start with the Deep Dive video.

3. Where to follow his latest work

The page says that for all the latest updates, he spends most of his time on X/Twitter or GitHub. Those are the two outbound channels it points to, so if the homepage feels static, those links are where ongoing activity lives.

How to use the page

• If you want to learn about LLMs: pick the track that matches your level — Zero to Hero for a technical path, the Deep Dive and practical-use videos for a general audience path.
• If you want background on his work: read the timeline blocks; they cover OpenAI, Tesla Autopilot, and his Stanford teaching in one place.
• If you want current activity: follow the X/Twitter or GitHub links rather than expecting the homepage itself to be frequently updated.

What it is not

There is no pricing, no signup, and no product offering described on the page, so do not treat it as a service you can purchase or an account you can create. It is a personal hub: career record, learning resources, and pointers to social profiles. The playful "Unicorn Magic" language confirms the personal, non-commercial framing — expect a human homepage, not a polished company site.

Andrej Karpathy earned his PhD at Stanford (2011–2015), where his research focused on convolutional and recurrent neural networks and their applications in computer vision, natural language processing, and the intersection of the two. His adviser was Fei-Fei Li at the Stanford Vision Lab, and he also worked with Daphne Koller, Andrew Ng, Sebastian Thrun, and Vladlen Koltun during the first-year rotation program. During the same period he designed and was the primary instructor for Stanford's CS 231n: Convolutional Neural Networks for Visual Recognition, which grew from 150 enrolled students in 2015 to 330 in 2016 and 750 in 2017, becoming one of the largest classes at Stanford.

PhD Research: What He Actually Worked On

Karpathy's doctoral work sat at the meeting point of two model families and two application domains:

• Convolutional neural networks (CNNs) — the architecture behind image recognition, which later became central to his industry work.
• Recurrent neural networks (RNNs) — sequence models used for text and other ordered data.
• Applications — computer vision, natural language processing, and the intersection of the two (for example, tasks that require connecting images with language).

This combination of vision and language is a useful signal for anyone trying to understand his later trajectory: the same pairing shows up in autonomous driving (perception plus reasoning about scenes) and in large language models (sequences of tokens at scale).

The People Around Him

The page names his adviser and collaborators explicitly, which matters if you're evaluating the research environment rather than just the degree:

Fei-Fei Li is closely associated with computer vision and the ImageNet line of work, so a PhD under her supervision places Karpathy squarely in the vision community. The rotation-year names span machine learning, robotics, and vision, which is consistent with a broad first year before specializing.

CS 231n: The Course He Built

Karpathy designed CS 231n and served as its primary instructor. The course title — Convolutional Neural Networks for Visual Recognition — describes its scope directly: how CNNs work and how they are applied to recognizing what is in an image.

Enrollment Growth

The page gives exact figures, and the trend is the clearest evidence of the course's impact:

Going from 150 to 750 students in three years made it one of the largest classes at Stanford. That growth also tracks the broader surge of interest in deep learning over the same period.

Why This Course Matters Beyond Stanford

CS 231n became a reference point for how deep learning is taught, not just at Stanford but well beyond it. If you are trying to learn the material today, the practical takeaway is that the course's structure — convolutional networks applied to visual recognition — is a well-defined prerequisite path: you need linear algebra, basic calculus, and some programming before the CNN material will land.

How to Place This in His Wider Career

The academic phase is one segment of a longer timeline documented on karpathy.ai:

• 2011–2015 — Stanford PhD; designed and taught CS 231n.
• 2015–2017 — research scientist and founding member at OpenAI.
• 2017–2022 — Director of AI at Tesla, leading the computer vision team for Autopilot (and briefly Optimus), covering in-house data labeling, neural network training, and deployment on Tesla's custom inference chip.
• 2023–2024 — returned to OpenAI, building a team working on midtraining and synthetic data generation.
• 2024 onward — creates educational videos on AI via his YouTube channel, in a technical track (the Zero to Hero playlist) and a general-audience track (Deep Dive into LLMs like ChatGPT, How I Use LLMs, and the earlier Intro to Large Language Models).

The through-line from the PhD to the rest: CNNs and sequence models, applied first to vision and language research, then to production-scale perception at Tesla, then to LLM training at OpenAI, and now explained to a general audience.

Quick Answers to Common Questions

Did he get a PhD? Yes — Stanford, 2011–2015, advised by Fei-Fei Li.

What was his thesis area? Convolutional and recurrent neural networks, applied to computer vision, NLP, and their intersection.

Did he create CS 231n? Yes — he designed it and was the primary instructor.

How big did the class get? 150 (2015) → 330 (2016) → 750 (2017).

Where can I verify this? The source is his own site, karpathy.ai, which lists his positions, adviser, collaborators, and the enrollment numbers.