The AI Talent War Has No Handcuffs Left

Andrej Karpathy joined Anthropic today.

Not quietly. Not after a gardening leave cool-down period. Not after a year on the bench while lawyers negotiated. He announced it on X this morning and started this week.

Quick résumé if you somehow need one: founding member of OpenAI in 2015, ran Autopilot at Tesla for five years, came back to OpenAI in 2023, left again. Started Eureka Labs, an AI-native education company. PhD at Stanford under Fei-Fei Li. Taught CS231n, the convolutional neural networks course responsible for training roughly half the people now building these systems. His YouTube series building GPT from scratch in a Jupyter notebook is the closest thing the field has to a Feynman lecture. He coined "vibe coding" last year, which, love it or hate it, landed.

He could have done anything. Gone back to OpenAI. Joined Meta or xAI for a check with enough zeros to make the decision easy. Stayed at Eureka full time. Started his own lab. He picked Anthropic.

He is not the first. Jan Leike walked out of OpenAI's superalignment team in 2024 and went straight there. Mike Krieger, the Instagram co-founder, took the CPO role. Now Karpathy. At some point you stop calling it a coincidence and start asking what Anthropic is doing that nobody else can replicate.

But that question has a legal backstory worth understanding first.

The Legal Infrastructure Designed to Prevent This Has Been Collapsing for Two Years

In April 2024, the FTC voted to ban most non-compete agreements nationwide. The rule never took effect. Federal courts in Texas and Florida blocked it before the September deadline, and by September 2025 the FTC under new leadership formally dropped its appeal. Dead on arrival.

But the old playbook was already wrecked.

California has banned non-competes since 1872, which is a fact that never stops being wild to people outside the state. In 2024, two new laws (SB 699 and AB 1076) expanded that ban to cover agreements signed in other states and required employers to notify workers that any non-competes they had signed were void. The notification deadline was February 2024. Every AI researcher in California got a letter. The letter said: you are free.

OpenAI learned what that meant the hard way. In May 2024, Vox reported that OpenAI's exit agreements contained provisions that could claw back vested equity, potentially worth millions, if former employees criticized the company. The agreements banned disparagement forever. Even acknowledging the NDA existed was a violation. CEO Sam Altman said he was "genuinely embarrassed" and claimed he did not know the provisions existed, though leaked documents showed the language had been signed by Altman himself. OpenAI reversed course within days, releasing former employees from the agreements and pledging never to cancel vested equity.

That scandal broke the same week Jan Leike resigned and publicly said OpenAI's "safety culture and processes have taken a backseat to shiny products." He went to Anthropic. So did the narrative.

Garden Leave Only Works Where Non-Competes Are Enforceable. In California, They Are Not.

Google's DeepMind tried a different approach, and it tells you something about why legal mechanisms alone do not solve this problem.

UK-based DeepMind researchers face six-month to one-year garden leave clauses. They get paid to sit at home. They cannot work for competitors. Sounds generous until you remember that six months in AI is a geological era. A former DeepMind employee told Business Insider: "Who wants to sign you for starting in a year? That's forever in AI." Former DeepMind director Nando de Freitas posted publicly urging people not to sign these contracts at all.

That is not the kind of endorsement your retention strategy is supposed to generate.

Garden leave only works in jurisdictions where non-competes are enforceable. In California, they are not. And California is where the talent keeps ending up.

The Quieter Cousin Nobody Thinks About

Non-solicitation agreements are the other tool in the box, and most people outside of HR and legal never consider them.

A non-compete says you cannot work for a rival. A non-solicit says you cannot recruit your former colleagues to come with you. In theory, non-solicits survive even in California because they protect against active poaching rather than restricting individual movement. Clean distinction. Elegant, even.

In practice the line is blurry. When Jan Leike leaves OpenAI and joins Anthropic, and then six months later another OpenAI researcher follows, was that solicitation or gravity? When a SignalFire report finds that OpenAI engineers are eight times more likely to leave for Anthropic than the reverse, and DeepMind engineers eleven times more likely, is that a non-solicit problem or a culture problem?

Those are very different diagnoses, and they lead to very different responses. One you fix with lawyers. The other you fix by looking in the mirror.

Non-competes are dead in California. The FTC's nationwide ban failed, but four states have full bans and 33 more have restrictions. Non-solicits still hold up in most jurisdictions, but they only prevent active recruitment. They cannot stop someone from walking out the door on their own and joining a competitor who never called them.

Which means the real retention tool was never a contract. It was always the work.

Anthropic Didn't Need to Win This With a Legal Maneuver

Anthropic did not poach Karpathy. They did not navigate a non-compete or sidestep a non-solicit. Karpathy left OpenAI in 2024. He has been independent since. He chose Anthropic because, in his own words, "the next few years at the frontier of LLMs will be especially formative" and he wanted to get back to R&D. Anthropic confirmed he will join the pre-training team and build a new group focused on using Claude to accelerate pre-training research.

Anthropic's interpretability work is the strongest in the field. Karpathy's entire craft is stripping neural networks down to their essentials and explaining what is actually happening inside them. The fit is so obvious it barely needs stating, and yet nobody else closed it.

SignalFire's 2025 State of Talent Report puts Anthropic's retention rate at 80% for employees hired in the last two years. Former employees describe a culture of intellectual discourse and researcher autonomy. Claude has become a developer favorite. None of that is a legal strategy. All of it is brand, built through the product, the research, and a culture that treats researchers like adults rather than flight risks to be contractually contained.

OpenAI, meanwhile. Ilya Sutskever left and started SSI. Mira Murati quietly assembled a team before even announcing her new venture, pulling roughly 20 staffers from OpenAI in the process. Leike. Krueger. Schulman. Now Karpathy. These people did not leave because a competitor's legal team outmaneuvered OpenAI's. They left because a competitor's mission was more convincing than OpenAI's.

The Musk v. Altman trial ended yesterday with the jury ruling in OpenAI's favor. But the talent verdict has been coming in for months, and it is not going the same way.

What Brand Leaders Should Take From This

Three things.

First, non-competes are a shrinking tool and the direction is clear even without a federal ban. Four states ban them outright. Thirty-three more restrict them. The FTC still has enforcement authority on a case-by-case basis and launched a task force specifically targeting anti-competitive labor practices in 2025. If your talent strategy depends on contracts that prevent people from leaving, you are building on a foundation that is actively eroding underneath you.

Second, non-solicits are more durable but they only solve half the problem. They prevent active recruitment. They do not prevent someone from wanting to leave. They do not prevent a competitor's reputation from doing the recruiting for them. When engineers are leaving your company for a rival at eight-to-one ratios, the problem is not that someone called them. The problem is that something pulled them.

Third, the companies winning the AI talent war are not winning it with contracts. They are winning it with clarity of mission, quality of work, and a culture that treats researchers like adults. Anthropic's hiring streak (Leike, Krieger, Ross Nordeen from xAI, now Karpathy) is not a legal strategy, but a brand strategy executed through the product, the research, and the culture. And it is working better than any contract clause ever could.

The handcuffs are now off. Every tool that companies used to keep top talent locked in place is either illegal, unenforceable, or irrelevant in the jurisdictions where it matters most. What remains is the work itself. The mission. The culture. The thing you are actually building and whether it is good enough that people choose to stay.

Right now, in this particular window of AI history, Anthropic is building that. Most of the competition is not. The race is about the work now. It always should have been.