Andrej Karpathy

Who they are

Andrej Karpathy is a researcher at Anthropic — co-founded OpenAI in 2015, ran Tesla's Autopilot vision team, and then founded Eureka Labs, an AI-native education startup, before joining Anthropic's pre-training team in 2026.

Person

Karpathy did his BSc in Computer Science and Physics at the University of Toronto (2009), his MSc at UBC (2011), then his PhD at Stanford under Fei-Fei Li (2015), focusing on connecting images and natural language — foundational work for what became the multimodal era. From Stanford he went straight to OpenAI as a co-founding research scientist in 2015, then moved to Tesla where he rose to Senior Director of AI and Autopilot Vision, building one of the most closely watched computer-vision pipelines in industry. He left Tesla, returned briefly to OpenAI, then in July 2024 launched Eureka Labs from scratch — an AI-native education platform, still in early hiring mode at launch — with flagship product LLM101n and the 'Zero to Hero' series on LLM fundamentals. Alongside his employed track he's built and released ConvNetJS (neural networks running in the browser), the Hacker's Guide to Neural Networks, and coined the term 'vibe coding' for intent-driven code generation. His public writing spans karpathy.ai, karpathy.bearblog.dev, and Medium essays including the widely-cited 'Software 2.0'; he spoke at YC AI Startup School in 2025 on 'Software 3.0,' at MIT CSAIL on AI and education, and appeared on the Dwarkesh Patel podcast arguing AGI is still a decade away. The through-line is moving between the sharpest possible technical frontiers — vision, autonomy, language models, now pre-training fundamentals — while simultaneously explaining those frontiers to everyone else.

Network

No direct network edges are available in the claims. The most clearly established relationship is with Fei-Fei Li, his Stanford PhD advisor and a formative influence in computer vision. His public interlocutors include Dwarkesh Patel (podcast, AGI timelines) and Lex Fridman (podcast, Tesla AI and self-driving).

Fei-Fei Li· PhD advisor, Stanford; AI researcher and entrepreneur
Dwarkesh Patel· Podcast host; public intellectual on AI and science
Lex Fridman· Podcast host; AI researcher

How they likely show up

Career pattern of moving between the hardest open problems (multimodal vision → autonomous driving → LLMs → pre-training) → suggests someone who self-selects into difficulty and gets bored once a problem is solved.
Consistent public writing and teaching (blog, Medium, 'Zero to Hero' series, YC talk) → he processes ideas by explaining them; meetings where he can teach or frame a concept will get more energy from him than status updates.
Built ConvNetJS, nanochat, and LLM101n as standalone releases alongside full-time roles → high-agency, ships independently, doesn't wait for organizational permission to create.
Coined 'vibe coding' and spoke at YC AI Startup School on 'Software 3.0' → he names and frames paradigm shifts, then builds into them early; he is likely impatient with incremental thinking.
Dwarkesh podcast appearance arguing 'AGI is still a decade away' → willing to take a public contrarian position against prevailing hype; values precision over narrative.
Tenure shape is mixed (OpenAI → Tesla → OpenAI → Eureka Labs → Anthropic) → follows the problem, not the institution; institutional loyalty is not a strong motivator.

Conversation tips

→ Engage him on the 'Software 3.0' framing from his YC talk — he put real thought into that thesis and will have a lot to say if you push on where it breaks down.
→ Cite a specific post or essay (e.g. 'Software 2.0' on Medium, or his bearblog year-in-review) rather than a general compliment — he'll notice you actually read it.
→ Ask about AGI timelines with a specific technical angle; he gave a concrete 'decade away' view on Dwarkesh and will engage seriously with pushback grounded in capability benchmarks.
→ Don't try to impress with hype — he is publicly precise about the gap between marketing language and what models actually do.
→ If the conversation touches on education and AI, reference the Khan Academy comparison he drew publicly; it signals you followed the Eureka Labs launch closely.

Toolbox

Openers

Open on the Anthropic pre-training move — he left his own company (Eureka Labs) to join Anthropic's pre-training team in 2026, a very specific bet on where the core technical leverage still sits; that's a striking choice worth understanding.
Reference 'Software 3.0' from his 2025 YC AI Startup School talk — he laid out a specific thesis about LLMs as a new programming layer, and Sequoia Ascent 2026 suggests he's still actively shaping that frame.
Bring up ConvNetJS or the Hacker's Guide to Neural Networks — he built tools specifically to make neural nets legible to practitioners before it was fashionable, which is the same instinct behind Eureka Labs; it shows you know the longer arc.

Discovery questions

You argued on Dwarkesh that AGI is still a decade away — has anything in pre-training since then shifted that view, or hardened it?
With Eureka Labs you were explicit about a Khan Academy model (free content, paid services) — what did you learn about AI-native education that made you step away from running it yourself?
You coined 'vibe coding' and then almost immediately started pushing back on its limits — where exactly does intent-driven code generation break down at production scale, in your view?

Avoid

Don't open with generic praise about his 'impact on AI' or his 'visionary career' — he engages with specific technical claims and framings, and broad flattery will signal you haven't done the reading.

Make it yours