The World's Best AI Labs Came to Him.

BUILDERS SERIES  ·  ISSUE 002

Andrej Karpathy didn't land a role at OpenAI or Tesla by applying. He built in public — publishing research, teaching for free, releasing code anyone could use. The output made the case no CV ever could.

Born in Slovakia. Raised in Canada. Drawn to the Hardest Problems.

Andrej Karpathy was born in Bratislava in 1986 and moved with his family to Toronto when he was 15. He studied Computer Science and Physics at the University of Toronto — where, as an undergraduate, he learned from Geoffrey Hinton, one of the founding fathers of modern deep learning. Most students pass through famous professors. Some absorb them completely.

After completing a Master's degree at the University of British Columbia, Karpathy arrived at Stanford for his PhD under Fei-Fei Li — one of the most respected computer vision researchers in the world. He was entering the field at exactly the right moment. Deep learning was beginning to show results nobody had expected.

Neural networks were starting to see.

His PhD focused on connecting images and natural language — teaching machines to look at a picture and describe what they saw. It was foundational work at the intersection of computer vision and natural language processing, at a time when most researchers treated the two fields as entirely separate.

The Class That Changed Everything — Published for Free

In 2015, while still completing his PhD, Karpathy designed and taught Stanford's first-ever deep learning course: CS231n — Convolutional Neural Networks for Visual Recognition. It started with 150 students.

What made it different wasn't just the content — it was how he shared it. The lectures, notes, and course materials were published openly online. For free. Anyone in the world could learn from them. No enrolment. No tuition. No gatekeeping.

The course became one of the most widely used deep learning resources on the planet. Developers in São Paulo, engineers in Nairobi, researchers in Seoul — thousands of people who had no connection to Stanford learned to build neural networks from Andrej Karpathy's publicly released notes. By 2017, the course had grown to 750 students on campus alone. Online, the reach was incalculable.

This is important to understand. Karpathy didn't have to do this. He could have taught the class, kept the materials internal, and moved on. Instead, he put his thinking on the internet. Permanently. Publicly. Accessible to anyone who wanted it.

Building in Public Before Anyone Called It That

Alongside the Stanford course, Karpathy was relentlessly building and releasing things publicly. He wrote ConvNetJS — a deep learning library written entirely in JavaScript, so anyone with a browser could train neural networks without setting up a development environment. He released char-rnn, a character-level language model that let people train small language models on their own text. He built arxiv-sanity, a tool to make the flood of AI research papers navigable.

His blog posts spread through the AI community like textbooks. 'The Unreasonable Effectiveness of Recurrent Neural Networks' — published in 2015 — became one of the most shared pieces of writing in the history of AI education. It was a technically rigorous but genuinely readable explanation of how language models work. It wasn't a paper. It wasn't behind a paywall. It was just a post on his personal website.

He also published 'Software 2.0' — an essay arguing that neural networks weren't just a tool for AI research but a new paradigm for writing software itself. Instead of coding explicit rules, you curate data and let the network learn the behaviour. It reframed how the entire industry thought about building intelligent systems. It's been referenced thousands of times.

None of this was part of a strategy. It was just how he worked — thinking out loud, building in public, sharing what he understood.

OpenAI Didn't Post a Job Ad. They Already Knew His Work.

In 2015, OpenAI was founded — a new kind of AI research organisation, backed by some of the most significant names in technology. They needed to hire the best researchers in the world. They needed people who understood deep learning at a level that barely existed outside a handful of labs.

Karpathy became one of OpenAI's founding research scientists. He didn't submit a CV to a job board. He didn't go through a standard interview pipeline. He was known. His research, his course, his open-source code, his writing — all of it had been visible to the people who mattered for years. When OpenAI was assembling its founding team, there was no question he belonged in it.

The work was the application. The body of public output was a more complete and persuasive case for hiring him than any document could ever be.

Tesla. Autopilot. The Real World.

In 2017, Karpathy left OpenAI to join Tesla as Director of AI, reporting directly to Elon Musk. His brief was one of the most ambitious in the history of applied AI: take neural networks out of research papers and deploy them in cars, at scale, in the real world.

He led the team responsible for Tesla's Autopilot and Full Self-Driving systems — overseeing all neural network training, data labelling, and deployment on Tesla's custom inference chips. This was his 'Software 2.0' thesis made physical: instead of writing rules for how a car should navigate a junction, you train a network on millions of hours of real driving data and let it learn.

The scale was unlike anything in academic AI. Tesla's fleet represented a live data collection machine of hundreds of millions of miles. Karpathy's team was training and deploying neural networks that shipped directly to vehicles on public roads. The feedback loop between model, data, and real-world performance was operating in near real time.

He stayed at Tesla for five years. He left in 2022, describing a desire to return to individual work and open-source projects.

YouTube. nanoGPT. Teaching the World Again.

After leaving Tesla, Karpathy did what he'd always done when he had freedom: he built things in public and gave them away.

His YouTube channel — which he'd started building before joining OpenAI — became something else entirely. His 'Zero to Hero' series walks learners through building neural networks from scratch, in plain Python, with no shortcuts. His video on building a language model from first principles has been watched by millions of people. His one-hour deep dive into large language models has over three million views.

He released nanoGPT — a minimal, clean implementation of a GPT-style model that anyone could read, understand, and train themselves. It became one of the most starred AI repositories on GitHub. Not because it was the most powerful. Because it was the most legible. He made the most significant architecture in modern AI understandable to a generation of builders who didn't have a PhD.

In 2024, he founded Eureka Labs — an AI-native education company built on the belief that AI can be the greatest educational tool ever created. His first course: LLM101n — teaching people to build a language model from the ground up.

There Are Thousands of Andrej Karpathys. Most Don't Have His Network.

The reason Karpathy was found is straightforward: he built in an environment where the people with power to hire him were already watching. Stanford put him in the same rooms as the people founding OpenAI. His PhD supervisor was one of the most connected researchers in the field. The forums and conferences where he shared his work were exactly where the right people gathered.

He didn't need a CV because his network made the CV irrelevant. The work reached the people who mattered because the infrastructure around him made sure it did.

Most builders don't have that infrastructure.

Right now there are developers, designers, and engineers producing work at the same depth as Karpathy — publishing code nobody's found, writing on platforms no recruiter visits, building projects that solve real problems for real users without the institutional scaffolding that gets work in front of the right eyes.

Their output is real. Their evidence is there. The companies that need them just have no mechanism to find it.

That's what Vaultt is built for. A hiring platform where the work is the profile — not a two-page summary of job titles, but real projects, real output, real evidence of what someone can actually do. A place where the builder without the Stanford network gets seen the same way as the one with it.

The work has always been the proof. Vaultt is where it gets found.s been the proof. The industry just needed a better way to see it.