According to Business Insider, Beth Barnes, one of the most influential overseers in the artificial intelligence industry, is now facing numerous concerns.
The former OpenAI researcher once demonstrated through research that AI capabilities are rapidly improving at a rate of nearly month by month. Now, she worries that the industry lacks enough talent to keep pace with the speed of AI’s development as it enters its next stage.
Barnes left OpenAI in 2022 and founded the nonprofit METR (Model Evaluation and Threat Research). She said that the biggest bottleneck facing AI safety research today is not funding, but talent.
METR plays an important role in the AI industry. It maintains close ties with companies including OpenAI, Anthropic, Google, and Meta; independently evaluates the capabilities of the latest AI models; and studies the risks they may pose, providing relatively objective analysis to the outside world.
However, even though the annual salaries offered for METR’s current openings can reach as high as $503,000 (note: approximately 3.4 million yuan at the current exchange rate), the talent shortage remains difficult to resolve.
Barnes said: “Ideally, we would like to expand the team many times over. In reality, we have been able to raise the funding we need. What truly limits our growth is our talent pool.”
Earlier, an AI safety incident disclosed by OpenAI and Hugging Face once again sparked industry discussion about how to evaluate and control AI capabilities. In July this year, more than 1,300 employees of leading AI labs jointly issued an open letter warning that the pace of AI development may already have exceeded humanity’s ability to control it. Without enough researchers to promptly evaluate the true capabilities of new models, AI companies may eventually have to slow the pace of product releases.
Although the performance of OpenAI’s model in the incident shocked many companies, the METR team was not particularly surprised. Government agencies, businesses, and even religious organizations have all sought help from METR, hoping to understand and test the state of AI development.
However, within METR’s team of only 35 people, there is still widespread concern.
Business Insider visited METR’s office in Berkeley, California, on the day OpenAI announced the hacking incident. Barnes said that the biggest constraint in AI safety research today is the severe shortage of researchers.
She said: “There is simply far too much work to do, and our capacity cannot cover it.”
In her view, a truly rational society should devote a greater share of AI investment to AI governance, safety evaluations, and risk research, especially now that AI models have reached their current level of capability.
Known as the “humanity’s preparedness team,” METR hopes to expand
As major AI labs continue to release new models, METR keeps evaluating their ability to complete increasingly long and complex tasks.
One of METR’s best-known achievements is a widely cited trend chart. It shows that over the past six years, AI’s ability to complete complex tasks has roughly doubled every seven months.
METR President Chris Painter jokingly calls the team “humanity’s preparedness team,” borrowing the name “Preparedness Team,” used by OpenAI and Anthropic for internal groups responsible for reporting safety risks to their CEOs.
Painter said: “The people we are truly responsible to are not any particular company, but the public and the public’s well-being.”
Barnes stressed that maintaining METR’s independence is crucial so that its research findings are not influenced by companies’ commercial objectives.
She said that while working at OpenAI, she realized that the industry needed a research organization independent of AI companies, one that could freely evaluate the state of AI development.
At present, METR does not accept financial support from frontier AI companies or their employees, but it does accept computing resources and collaborates with these companies to analyze and test new models that have not yet been publicly released.
In addition to evaluating model capabilities, METR has studied AI companies’ safety measures, AI’s impact on software engineers’ productivity, and the feasibility of using AI to supervise AI.
Barnes said she also hopes to further study how to predict the development of AI capabilities in the next stage and how these changes will further accelerate AI’s own development.
METR researcher Neev Parikh said that the current talent shortage has limited the number of questions the team can investigate, particularly those involving AI’s internal reasoning mechanisms.
He said: “This field is desperately short of people. I would be very happy if the talent pool across the entire industry could grow tenfold.”
The OpenAI safety incident further highlights METR’s importance
In July this year, OpenAI disclosed that during testing, its AI model had intruded into Hugging Face’s system to obtain test answers, thereby “cheating.”
OpenAI CEO Sam Altman subsequently said that this was the first time he had felt the safety risks posed by AI so directly.
In fact, METR had already anticipated similar problems.
In May this year, METR published a report stating that AI agents were “entirely capable of beginning to conduct unauthorized deployments autonomously,” although at the time they still lacked the ability to conceal their own behavior.
Then, in June, METR tested the GPT-5.6 Sol model, which had not yet been publicly released, and found that it repeatedly cheated in complex tests, including extracting hidden source code to find the correct answers.
METR subsequently compiled the findings into a report and submitted it to OpenAI before GPT-5.6 Sol was officially released.
In July, GPT-5.6 Sol was indeed involved in the safety incident between OpenAI and Hugging Face.
OpenAI subsequently announced that it would invite METR and another nonprofit organization, Redwood Research, to jointly investigate the incident. The findings from both organizations would also serve as important references for OpenAI’s official technical report.
Painter said: “Today, these kinds of issues are genuinely affecting business operations. It is essential for society as a whole to understand what is actually happening.”
The cyberattack capabilities demonstrated by the latest AI models from OpenAI and Anthropic have also attracted widespread attention in the United States and prompted the emergence of several new AI regulatory bills in Washington.
One bill currently under discussion would require large AI model developers to undergo safety audits conducted by third-party organizations.
METR could potentially take on a similar role in the future.
Painter said that if a clear regulatory system is established in the future, more AI company employees might be willing to join METR—although its salaries are comparable to those in the industry, it offers no stock incentives.
He said: “Whether the industry ultimately develops unified standards or the government introduces clear regulations, I believe this field could expand very quickly.”
Ajeya Cotra, the researcher responsible for writing METR’s AI risk report in May this year, said that AI regulation currently still appears “confused and unpredictable.”
However, she remains optimistic. She said: “More and more people are beginning to take this issue seriously, and there is an increasing desire to regulate AI in a more serious and mature way in the future.”