The AI race is moving so fast that even high salaries and enormous bonuses are no longer enough to retain some of the industry’s top specialists.

Something unusual is happening in the world of artificial intelligence: experts working on the most promising technologies are beginning to leave leading laboratories, despite the enormous financial opportunities and their involvement in projects capable of changing the future.

On September 9, 2026, 27-year-old Cambridge mathematician Jacob Coxon announced his departure from Anthropic. Before that, he spent three years conducting pretraining research at OpenAI, where he worked on GPT-4o. In early 2026, he moved to Anthropic, taking on a similar role. At both companies, he worked on methods for pretraining large language models.

After resigning, Coxon posted a message on X that received around 90 million views within 24 hours:

“Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.”

The statement publicly expressed concerns that have long been discussed within the AI industry. According to Coxon, OpenAI and Anthropic are accelerating the development of systems capable of improving their own abilities, while competitive pressure prevents them from paying sufficient attention to the risks.

Self-improving AI may be closer than it seems

In an interview with Fox News, Coxon said that artificial intelligence capable of making itself smarter could emerge as early as next year. He also noted that some experts are suggesting even shorter timelines – around six months.

This is not simply about creating a more powerful language model. Self-improvement means that a system could find ways to enhance its own architecture, algorithms, or training process. If such a cycle proves sufficiently effective, AI development could accelerate far beyond what most people expect.

Coxon emphasized that he does not consider the leaders of Anthropic and OpenAI to be villains. The problem, he said, lies in competitive pressure. Companies are trying to outpace one another, attract the best researchers, and gain an advantage in developing technologies that could become the foundation of a new economy.

His position was publicly supported by Evan Hubinger, head of Alignment Science at Anthropic. This is particularly notable because he was still an employee of the company at the time of the statement.

Alignment Science focuses on how to make AI behavior consistent with human goals and intentions. Support from a specialist in this field shows that concerns about the speed of the AI race are not limited to former employees.

Money no longer guarantees loyalty

Coxon’s story is not the only one. In recent months, the industry has seen other cases of specialists leaving technology companies despite extraordinarily attractive financial conditions.

At Meta, researcher Tullock left the company despite an offer reportedly involving $1.5 billion in compensation from Mark Zuckerberg. The story became a symbol of just how high the stakes have become in the race for artificial intelligence talent.

Meanwhile, Chinese DeepSeek demonstrated how quickly new models can reshape the economics of the AI market. Its developments put pressure on token prices and forced investors to reassess the prospects of companies spending tens of billions of dollars on model training.

A high salary alone no longer means that an employee is willing to stay at a company at any cost. For some researchers, the safety of the technology is becoming just as important as money, career growth, or access to computing resources.

The main risk is not evil AI, but the speed of competition

The debate about the future of artificial intelligence often centers on whether AI could become dangerous. However, Coxon’s statements point to a different problem: even if developers are not trying to create a dangerous system, competitive pressure may force them to move faster than their understanding of the risks allows.

OpenAI and Anthropic are competing for computing power, investment, researchers, and commercial customers. Each new model is not only a technological achievement but also a tool in the battle for market share.

Under these conditions, slowing development for additional safety checks may be seen as a strategic advantage for a competitor. This conflict between speed and safety has become one of the central questions facing the modern AI industry.

It is still impossible to say with certainty when self-improving superintelligence will emerge, or whether it will appear within the timelines being discussed. But the departure of researchers from leading laboratories shows that serious disagreements already exist within the industry over how quickly humanity should move toward the next stage of artificial intelligence.

And perhaps the most important signal is not that someone has left a company. It is that even specialists being offered billions are beginning to consider certain technologies too risky to continue working on without fundamental changes to the approach to safety.