Andrej Karpathy

Who they are

Andrej Karpathy is Founder and CEO of Eureka Labs — co-founded OpenAI in 2015, built Tesla's Autopilot vision stack, and created 'Neural Networks: Zero to Hero', the widely regarded free YouTube series teaching modern LLMs from scratch.

Person

Karpathy did his BSc in Computer Science and Physics at the University of Toronto (2009), then an MSc at UBC (2011), before a Stanford PhD in 2015 under Fei-Fei Li — his dissertation connected images and natural language, which turned out to be a founding thesis of the GPT era. Straight out of Stanford he co-founded OpenAI in 2015 as a research scientist, stayed through 2017, then joined Tesla as Director of AI and Head of Autopilot Vision — spending five years (2017–2022) building the perception stack for self-driving at scale. He returned to OpenAI for a second stint (2023–2024) before departing to found Eureka Labs in July 2024, an AI-native education platform whose flagship product, LLM101n, teaches students to train large language models from scratch. He co-founded two companies: OpenAI (AI research and deployment, 2015) and Eureka Labs (AI-native education, 2024). He also built a string of side projects: ConvNetJS (neural nets in the browser), Hacker's Guide to Neural Networks, Neural Networks: Zero to Hero (YouTube series at karpathy.ai), LLM Wiki (an experimental AI-collaborative knowledge repository), and coined 'vibe coding' — the practice of expressing developer intent in natural language to generate code. His public writing spans karpathy.bearblog.dev and karpathy.medium.com, covering deep learning, LLMs, software evolution (his 'Software 2.0' essay is widely cited), and AGI timelines; he spoke at YC AI Startup School in 2025 on 'Software 3.0' and appeared at Sequoia Ascent 2026. The through-line is research depth converted into maximum public legibility — every move has made hard AI concepts more accessible at wider scale.

Network

Karpathy's most prominent named relationship is Fei-Fei Li, his Stanford PhD advisor and mentor in computer vision and NLP. He is a recurring guest on Dwarkesh Patel's podcast, where they discuss AGI timelines and LLM capabilities, making Patel a consistent intellectual interlocutor. No direct co-worker or investor edges are available from the claims.

Fei-Fei Li· Stanford Professor, PhD advisor and mentor in computer vision and NLP
Dwarkesh Patel· Podcast host, recurring intellectual interlocutor on AGI and LLM capabilities

How they likely show up

Hybrid role-type pattern across his career (researcher → operator → founder) → he likely context-switches between deep technical work and high-level strategic framing without needing a hard boundary between the two.
Built Neural Networks: Zero to Hero and LLM101n as teaching-from-first-principles projects → prefers to understand systems by rebuilding them from scratch, not by abstracting over them.
Public writing signal rated 'thought-leader' with essays like 'Software 2.0' and regular blogging on deep learning and AGI timelines → comfortable putting half-formed theses in public and iterating on them openly.
Medium tenure at each employer (OpenAI 2015–17, Tesla 2017–22, OpenAI 2023–24) → moves when the learning curve flattens, not for money or prestige signaling.
Coined 'vibe coding' and built LLM Wiki as experimental side projects alongside full-time roles → high personal agency; likely pursues parallel intellectual threads regardless of what the day job demands.
Spoke at YC AI Startup School on 'Software 3.0' and at Sequoia Ascent 2026 → willing to engage with the founder and investor community on framework-level ideas, not just technical details.

Conversation tips

→ Reference a specific piece of his writing — 'Software 2.0', the Zero to Hero series, or his bearblog posts — and react to the actual argument, not just the topic. He'll notice if you've read it.
→ Ask about first-principles reasoning: his teaching style (LLM101n, Zero to Hero) is built on rebuilding things from scratch, so questions that start from fundamentals will land better than ones that assume shared abstraction.
→ He has a stated view that AGI is still roughly a decade away — if you want to discuss AI timelines, come with a position, not just curiosity. He engages with people who push back.
→ Don't open with flattery about his career arc — he's heard it. Lead with a specific idea or claim you want to pressure-test.
→ He publicly compared Eureka Labs' model to Khan Academy's free-plus-paid structure — referencing that comparison signals you understand what he's actually building, not just that he 'does AI education'.

Toolbox

Openers

Open on 'vibe coding' — he coined the term and it's become a flashpoint in debates about what software engineering even is; asking whether he thinks it changes who gets to build software is a direct entry into something he cares about.
Reference the LLM Wiki project — he described it as an 'idea file' for the AI community, treating LLMs as collaborative knowledge repositories, which is a distinct bet from what Eureka Labs is doing and worth pulling on.
Bring up his Sequoia Ascent 2026 talk — he wrote a summary on his bearblog, so you can react to a specific claim from it and signal you followed the thread past the headline.

Discovery questions

LLM101n teaches people to train large language models from scratch — how do you think about the right depth of understanding for someone who just wants to build on top of models versus someone who needs to go all the way down?
You've described AGI as roughly a decade away — what's the leading indicator you'd watch to know you're wrong in either direction?
Neural Networks: Zero to Hero was built entirely in the open as a side project while you were employed elsewhere — how does that compare to building Eureka Labs as a company with the same teaching mission?

Avoid

Don't lead with questions about OpenAI's current direction or internal dynamics — he departed twice and has been careful not to editorialise publicly about the organization; it's likely a closed topic.

Make it yours