Why Karpathy Joined Anthropic: The AI Talent Exodus and a Pre-Training Revolution Powered by AI Itself

On the evening of May 19, Andrej Karpathy announced that he had joined Anthropic.

The name needs no introduction. Co-founder of OpenAI, former AI Director at Tesla, champion of "Vibe Coding," and arguably the most influential AI educator in the world. His stature in the field is akin to LeBron James joining a new basketball team — no matter where he goes, it's front-page news.

He posted just three sentences on X. The first said the next few years at the frontier of LLMs would be particularly "shaping." The third mentioned he still loves education. And the most important — the second sentence — contained just five words: "Returning to research and development."

This marks the third core figure to migrate from the OpenAI camp to Anthropic in the past two years. A man approaching 40, who already has fame and financial freedom, chose to work under someone else's leadership entirely of his own volition.

Why did he leave? Why Anthropic specifically? And why did Anthropic need him?

Behind each of these questions lies a logic worth unpacking.

What He'll Be Working On

Karpathy has already started at Anthropic this week, assigned to the pre-training team led by Nick Joseph — the group responsible for every large-scale training run behind Claude.

An Anthropic spokesperson confirmed to TechCrunch that Karpathy will form a new sub-team dedicated to "using Claude itself to accelerate pre-training research." Nick Joseph added on X: "He'll be building a team focused on using Claude to accelerate pre-training research itself."

TechCrunch noted that Karpathy is "one of the few researchers who can bridge the gap between LLM theory and large-scale training practice." Axios framed the move as "a major coup for Anthropic in the talent wars."

Alongside cybersecurity expert Chris Rohlf, who also announced his Anthropic arrival the same day, and former xAI founding member Ross Nordeen who joined earlier this month, the direction of talent flow is becoming unmistakable.

Polymarket data reinforces the market sentiment. Traders estimate a 65% probability that Anthropic will hold the best AI model by end of June, while OpenAI sits at just 4%. Karpathy's arrival only solidifies that bet.

Karpathy as a "Paradigm Definer"

To understand the weight of this hire, you need to understand just how rare Karpathy is. His rarity doesn't come from technical ability alone — other top-tier researchers exist.

His true rarity lies in his ability to reshape how an entire industry thinks about something, with a single phrase.

Born in Slovakia in 1986, he moved to Toronto, Canada at age 15. As an undergraduate at the University of Toronto, he attended Geoffrey Hinton's lectures and joined his reading group. Hinton — the spiritual leader of the deep learning renaissance, Turing Award winner in 2018, Nobel Prize in Physics winner in 2024 — lit a fire under Karpathy early on. He was one of the first young minds to catch that spark.

He then went to Stanford to study under another legendary figure, Fei-Fei Li, and created the CS231n course during his PhD. Enrollment surged from 150 in 2015 to 750 by 2017, and with all video lectures freely available online, it became the definitive first step for countless engineers worldwide teaching themselves deep learning — cementing its place as the gold standard in computer vision education.

In 2015, he became a founding research scientist at OpenAI. In 2017, Elon Musk personally recruited him as Tesla's Senior Director of AI, where he steered autonomous driving toward a pure vision approach. That poaching reportedly came with considerable pressure from Musk.

That same year, Karpathy published a Medium article coining the concept of "Software 2.0" — the idea that neural network weights are the new code, datasets are the new source code, and gradient descent is the new compiler. This framework fundamentally reframed how the industry understood programming itself.

After leaving Tesla in 2022, he launched the "Neural Networks: Zero to Hero" YouTube series, surpassing 1 million subscribers. His open-source projects from the same period — micrograd, nanoGPT, nanochat — were minimal in code volume but laser-focused on core concepts, earning the nickname "executable textbooks."

In February 2025, he coined the term "Vibe Coding," which Collins Dictionary selected as its Word of the Year. At his June YC AI Startup School talk, he presented frameworks around "Software 3.0" and "the decade of agents," making it one of the most debated AI keynotes of the year. TIME named him one of the 100 most influential people in AI in 2024.

Hinton. Fei-Fei Li. Sam Altman. Elon Musk. He has always been at the bleeding edge of each era. But the most enduring things he's left behind aren't individual products or papers — they're conceptual frameworks. Software 2.0. Vibe Coding. The LLM OS. These terms changed how people think about AI.

Why He's Willing to Be "Two Levels Down"

There's a clear through-line in Karpathy's career: he has never chased titles.

He was Hinton and Fei-Fei Li's student, Altman's colleague, Musk's direct report. In each case, his organizational position was senior. But at Anthropic, his direct manager is Nick Joseph, head of pre-training, who reports to Dario Amodei. In the org chart, Karpathy sits three layers deep.

Nick Joseph was one of Anthropic's 11 founding members, previously at Vicarious and OpenAI. During his OpenAI days on the safety team working with code models, he witnessed GPT-3 learn to write code after fine-tuning and became convinced that "AI can improve itself." He left with the safety team lead to co-found Anthropic. His team has trained every model in the Claude series, including Mythos.

The reason Karpathy doesn't mind working under Nick Joseph is simple: it's the place closest to what he actually wants to do.

Looking back at his career transitions, the driving force has always been the same question: "Where is the biggest experiment happening right now?"

• In 2017, he went to Tesla because autonomous driving was the ultimate Software 2.0 proving ground.
• In 2022, he left because the architecture had solidified, and what remained was engineering optimization.
• In 2023, he returned to OpenAI because the explosive adoption of ChatGPT following GPT-4's release made it the most exciting frontier.
• In 2024, he founded Eureka Labs because he wanted to test his hypothesis around AI-native education.
• In 2026, he joined Anthropic because the pre-training revolution of "using AI to research AI" is happening here.

He doesn't leave out of dissatisfaction — he leaves when a place is no longer the "biggest experiment."

Why didn't he return to OpenAI? The talent flow tells the story.

Jan Leike, OpenAI's former head of alignment, joined Anthropic in May 2024. Co-founder John Schulman followed in August. And now it's Karpathy's turn. Three core figures in two years, all flowing one direction — with zero cases going the other way.

OpenAI's strategic center of gravity has shifted from pure research toward platformization and acquisitions — Chat.com, io Products, Windsurf, TBPN — with acquisition intervals shortening and deal sizes growing. The company is positioning itself as "the consumer goods giant of the AI era." For researchers who want to return to research, Anthropic's "win on research quality" philosophy is far more attractive.

Why Anthropic Wanted Him So Badly

Anthropic's motivations for this hire can be broken into several layers.

The most surface-level reason is technical need. No matter how large Anthropic's compute budget, it can't match OpenAI (backed by Microsoft) or Google (which owns its own TPU fleet). In a pure compute slugfest, Anthropic loses.

That means it needs to find ways to train better models with fewer compute resources. "Using Claude to accelerate pre-training research" is exactly that path — and Karpathy is a rare individual who combines deep pre-training theory, large-scale engineering experience, and intuition for AI-assisted research.

Next comes the "talent signal." With three core OpenAI figures arriving in two years, a powerful narrative has formed: frontier researchers are voting with their feet. Every time someone of Karpathy's caliber joins, the psychological barrier drops for the next top-tier recruit. A talent-attracts-talent flywheel has started spinning.

Then there's the pre-IPO brand value boost. Anthropic is reportedly negotiating a $30 billion raise at a $90 billion valuation, with IPO preparations underway. Karpathy is one of the most recognized names in AI — with 1 million YouTube subscribers, a Word of the Year coinage, and a CLAUDE.md repo with 220,000 stars. Having his name on the employee roster is a powerful line for any investment bank to put in a prospectus.

But the most interesting motivation — even if it's never explicitly stated as a hiring rationale — is the one that may deliver the greatest return: paradigm-defining capability. Every technical exploration Karpathy undertakes at Anthropic will be shared through tweets, blog posts, and YouTube videos.

When he inevitably gives a name to what's happening, Anthropic becomes the natural origin point of that paradigm. By hiring one top-tier pre-training researcher, Anthropic also acquired the industry's most influential technical storyteller.

The Flywheel Approaching Criticality

To place this move in broader context, it points to a technical inflection point. In April 2026, Anthropic released Mythos Preview, its most powerful AI model to date. Mythos is so powerful that only invited users through Project Glasswing have been able to conduct internal testing.

Despite having no specialized network security training, Mythos autonomously discovered and exploited a remote code execution vulnerability that had existed in FreeBSD for 17 years. It also found vulnerabilities in OpenBSD dating back 27 years and defects in FFmpeg from 16 years ago. The UK AI Safety Institute's independent evaluation confirmed it was the first model capable of executing a 32-step corporate network attack simulation from start to finish.

Anthropic itself acknowledges these capabilities aren't the result of intentional training but rather "downstream emergence" from improvements in general reasoning and software engineering ability. The higher the quality of pre-training, the more the emergent capabilities exceed expectations.

Mythos is currently both the strongest model and the strongest tool.

What Karpathy intends to do at Anthropic is use this ultimate hammer to improve how the hammer itself is made.

Using Mythos and Claude, he aims to discover better training architectures, data mixes, and experimental directions — decoupling model improvement speed from the linear pace of human researchers, spinning up the evolutionary flywheel of "AI improving AI."

This is precisely the outcome Anthropic craves most. When this flywheel truly gets going, "AI-driven self-improving pre-training" won't just be one research direction — it will become the acceleration corridor toward AGI, and eventually ASI (Artificial Superintelligence).

Every current axis of competition — the compute arms race, data walls, talent wars — could be rewritten by this single variable.

In three years, OpenAI has lost three core figures to the same competitor. The impact of that fact may be greater than any fundraising amount.

Compute can be bought with money. Data can be accumulated over time. But humans who can spin the AI evolution flywheel can be counted on one hand worldwide. Karpathy chose to give up his independent position and return to the frontlines at this precise moment. He believes the window of opportunity is right in front of him.