It is no longer just a billion-dollar war to snap up the finest minds in artificial intelligence. Something is shifting inside the major labs building the world’s most powerful AI systems. Researchers, engineers and specialists who, until a few months ago, were working directly on frontier models are leaving companies such as Anthropic, OpenAI, Google DeepMind and xAI. Some move to rival firms, others launch their own startups. But there are also those who choose to step out of the race entirely, saying they no longer feel comfortable with the speed at which the technology is advancing.
The latest case comes from Google DeepMind itself. Josh Engels, a researcher on the team dedicated to AGI safety, has announced he is leaving the company after working on alignment issues for its most advanced systems. Engels explained that the capabilities of artificial intelligence are growing faster than our ability to understand and control its behaviour. Despite having received offers from OpenAI and Anthropic as well, he chose a different path: joining METR, an independent organisation that subjects the most advanced AI models to safety testing and evaluation.
His departure comes just days after one that made far more noise. Jacob Coxon, a British researcher who previously worked at OpenAI and then at Anthropic, quit his job while levelling a scathing accusation at both companies. According to Coxon, the labs are locked in an ever-faster race towards systems capable not only of carrying out complex tasks, but eventually of autonomously contributing to the development of even more powerful AI. The researcher argued that competition between the firms risks pushing necessary precautions into the background.
His exit drew particular attention for one financial detail. Coxon reportedly left Anthropic before a portion of the equity he stood to receive had vested, effectively forfeiting a significant sum. The researcher himself has pointed to this choice to rebut accusations that he raised the alarm out of self-interest or to influence the company’s valuation.
Coxon’s language was stark. He argued that some of the people directly building artificial intelligence genuinely believe extremely advanced systems could pose an existential risk. This does not mean there is industry-wide consensus that such a catastrophe will actually happen, nor that these predictions have been proven. But the more interesting point is this: the concern is not coming solely from outside observers or anti-AI activists. It is coming from the people who built these systems themselves.
And Coxon is not alone. Evan Hubinger, head of alignment research at Anthropic, has publicly shared a significant part of these concerns while remaining at the company, assigning a probability of over 10% to extremely catastrophic AI-driven scenarios in the near future. It’s an estimate disputed by many other scholars, but one that shows just how heated the debate has become even within the tech companies themselves.
The dilemma: leave, or try to change the companies from within
This is precisely where a divide opens up among the researchers themselves. Leave the lab, or stay and try to make it safer?
Some specialists argue that abandoning these companies means removing exactly the people most attuned to safety from the rooms where the most sensitive decisions are made. Others reach the opposite conclusion: continuing to work on ever more powerful models, in their view, means directly fuelling the very race they wish to slow down.
It’s a dilemma now openly debated within the AI community. Some researchers are therefore moving from commercial companies to independent institutes dedicated exclusively to model evaluation, while others are heading into academia or organisations focused on AI governance.
But it would be wrong to interpret all these departures in the same way. The exodus from the major labs has at least two entirely distinct faces.
The first is about safety: people leaving because they believe development is moving too fast.
The second, and by far the larger one, is the global war for talent.
Researchers are now worth millions
Specialists capable of designing, training and improving large models have become one of Silicon Valley’s most prized assets. Giant data centres and tens of thousands of processors are not enough: building a competitive model requires just a few hundred people with extremely rare skills.
That is why OpenAI, Meta, Anthropic, Google and xAI are constantly competing for researchers and engineers.
According to an independent database tracking staff movements in the sector, hundreds of senior figures have switched companies, founded startups or left the major labs in recent years. OpenAI stands out as one of the main sources of this new generation of entrepreneurs, with former employees having helped launch companies that have themselves become major players in the AI race.
In some cases the movement is almost circular: a researcher leaves Google for OpenAI, another leaves OpenAI for Anthropic, yet another leaves xAI for Meta. The labs are constantly trying to poach from rivals the people who best understand the techniques used to build the most advanced models.
In recent months, Google DeepMind has also had to deal with several significant departures. The market has closely watched the movements of researchers involved in developing Gemini and other strategic projects. In June, news of certain exits from Google’s AI division even helped stoke nervousness in the financial markets, precisely because investors regard the ability to retain top researchers as a key indicator of a company’s future competitiveness.
This is one of the most distinctive features of today’s AI industry. In traditional software, losing a few programmers rarely puts a tech giant’s future at risk on its own. In frontier AI, it can be different. A single researcher may hold knowledge accumulated over years of experiments costing hundreds of millions of dollars, know a model’s architecture in intimate detail, and already have in mind the approach needed to build the next generation.
The real issue is speed
Behind the safety-related departures, one word keeps surfacing: speed.
Models are being updated within a matter of months. AI agents are increasingly capable of writing code, using software tools, browsing the web and carrying out sequences of operations with reduced human oversight. And it is precisely this growing autonomy of the systems that has reignited the debate.
In the United States, certain incidents during tests and experiments with AI agents have caught the attention of researchers and authorities, fuelling discussion over the possibility that systems designed to operate in controlled environments might accidentally interact with external infrastructure. This isn’t the science-fiction scenario of “AI taking over the world,” but far more concrete problems: automated cyberattacks, autonomously exploited vulnerabilities, the production of malicious code, and agents’ capacity to carry out actions their programmers never intended.
It is precisely these warning signs that have turned what was until recently a largely theoretical discussion into a much more heated confrontation.
Even company leadership is starting to adopt more cautious tones. Dario Amodei, CEO of Anthropic, has called for greater coordination between the leading developers and independent model-verification systems. Sam Altman of OpenAI and Demis Hassabis of Google DeepMind have also voiced support for the need for more robust standards, while the political debate in the United States is rapidly shifting towards how much regulation should be imposed on the labs.
A paradox that’s hard to ignore
And this is where AI’s great paradox emerges.
Companies keep investing staggering sums to build ever more powerful systems. At the same time, some of the very scientists helping to develop them are calling for a slowdown.
Even companies that have built part of their reputation on safety, such as Anthropic, find themselves caught up in the same commercial competition as their rivals. They must raise capital, grow their customer base, train new models and, above all, make sure OpenAI, Google or Meta don’t get to the next generation of technology first.
The result is what many researchers describe as a classic arms-race problem: even a company convinced of the need for caution can be pushed to accelerate simply because it fears a competitor will do so regardless.
Stopping unilaterally could mean giving up billions in investment and years of technological advantage.
Pressing on, meanwhile, means accepting risks that not even the developers themselves claim to fully understand.
Not the end of the AI industry, but an important signal
Talk of a mass exodus of researchers would still be an exaggeration. The big companies continue to hire thousands of engineers and specialists, and demand for AI professionals has never been higher. For every scientist who leaves a major lab, there are many more trying to get in.
But the resignations of those working directly on the most advanced models carry a different kind of weight.
Not because they automatically prove that the most extreme scenarios will come to pass, but because they reveal just how deep the uncertainty runs, even among those who understand this technology best.
The question now hanging over Silicon Valley is no longer simply who will build the most powerful artificial intelligence.
It has become another one as well: how quickly can we build it before our ability to control it falls behind?
And the fact that some researchers have chosen to walk away from salaries, equity stakes and careers at the wealthiest tech companies on the planet just to raise this question is, perhaps, the most telling signal of all.
