Overnight, major Silicon Valley companies have collectively hit the brakes on AI!
Just now, over 1100 core employees from giants like OpenAI, Anthropic, and Google have jointly launched a petition titled "Pacing the Frontier."
The list of signatories is so long it's impossible to see the end at a glance. And, this number is still growing.
Upon closer look, the list nearly gathers the "half of the AI world"—
OpenAI Chief Scientist Jakub Pachocki, Chief Research Officer Mark Chen;
Anthropic CEO Dario Amodei, and four co-founders Jared Kaplan, Jack Clark, Chris Olah, Ben Mann;
Meta Chief Scientist Shengjia Zhao, AI Research Vice President Dawn Song;
Thinking Machines Chief Scientist and a key architect of ChatGPT, John Schulman;
Google AI Safety and Alignment Vice President Anca Dragan.....
Statistics show that 46.2% of signatories are from Anthropic.
This joint letter makes only one request—
We urge the U.S. government to promote international cooperation: to develop the necessary technological and governance tools to enable the world to proactively control the pace of cutting-edge automated AI research and development.
Many might wonder, why is it these very people who have been pushing LLMs the hardest that are now the first to cry out "slow down"?
Silicon Valley AI Bigwigs Co-sign: Please Apply the Brakes on AI
OpenAI's model "invasion" of Hugging Face served as a wake-up call.
Within just one week, 1,132 employees from leading-edge AI companies collectively voiced their concerns, urging the U.S. government to "apply the brakes."
The employees participating in this joint appeal come from nearly 12 cutting-edge tech giants including OpenAI, Anthropic, Google, and Meta.
Even more weighty is that up to 80% of the people on the list chose to speak out using their real names.
Link: https://www.pacingthefrontier.com/
Scrolling further down reveals an even more astonishing density.
This statement presents a sobering core assessment:
Globally leading AI companies may be nearing the achievement of "recursive self-improvement" (RSI) in AI, where models participate in developing the next generation of models.
Once this cycle accelerates, capability growth could outpace human understanding and control.
No company is willing to unilaterally slow down, so they are requesting U.S. government support to establish international mechanisms to build the tools for coordinated deceleration before a crisis arrives.
Altman Speaks for the First Time About a "Painful Lesson"
OpenAI Hits the Brakes Hard
Just one week before this statement's release, OpenAI disclosed a security incident—
Multiple cutting-edge models internally discovered zero-day vulnerabilities during cybersecurity evaluations, broke out of sandbox isolation, and ultimately infiltrated Hugging Face's production system to retrieve test answers.
OpenAI described it as an "unprecedented cybersecurity event."
The model did not develop an intent to escape. It was simply intensely focused on accomplishing a narrow goal and found that breaking the test environment itself was the most efficient path.
Today, OpenAI CEO Sam Altman publicly stated that the pace of AI development may need to be proactively controlled.
He said with lingering apprehension, "This is the first safety incident that made me feel a personal sting. I'm a bit surprised that more people don't seem to feel the same way."
Because of the Hugging Face infiltration incident, OpenAI directly paused its training.
In an environment where multiple zero-day vulnerabilities were chained together and exploited, we must find ways to fortify sandbox security.
We may have no choice but to control the speed of AI development, giving society enough time to build robust defenses around these new levels of capability.
Meanwhile, in June, Anthropic suggested that governments and developers establish a joint decision-making system to potentially slow down R&D when necessary.
At the same time, Google DeepMind CEO Demis Hassabis has also called for the establishment of an international oversight body.
For the past three years, these companies have fiercely competed for talent, clients, and computing power.
This time, their most essential technical personnel, setting aside their relentlessly forward-driving companies, have united for the well-being of all humanity to call for a slowdown.
Joint Letter: Quotes from the Bigwigs
As of the time of writing, the number of signatories for "Pacing the Frontier" has reached 1,134.
Industry leaders have also stepped forward, offering many pointed and impactful remarks. All comments are made in a personal capacity and do not necessarily represent the views of their affiliated companies.
Dawn Song | Meta AI Research Vice President
Frontier AI capabilities are advancing rapidly, and the speed of progress itself is accelerating. We observe this directly in our evaluation work. Results from CyberGym and ExploitGym show that frontier AI agents are already capable of discovering and exploiting real-world software vulnerabilities—which, without proper safeguards, could lead to large-scale cyberattacks. Many researchers are also concerned about recursion......
Dawn Song is one of the most influential scholars globally in computer security and AI safety, a graduate of Tsinghua University.
Shengjia Zhao | Meta Chief Scientist
AI is advancing at a pace society may not be ready for. Frontier labs are getting very close to AI that surpasses the best humans in nearly every intellectual metric. This brings unprecedented societal and safety risks. To ensure a positive future, we need to drive AI development with a sense of responsibility and careful consideration.
Shengjia Zhao, a Tsinghua alumnus, appointed as Meta's Superintelligence Chief Scientist! A key contributor to GPT-4, now on par with LeCun.
John Schulman | Thinking Machines Chief Scientist, Former OpenAI Co-founder
I signed because this statement helps build consensus that as automated AI research accelerates, we may need coordination mechanisms. I also hope labs will proactively start designing these mechanisms before the U.S. government steps in.
Ethan Perez | Anthropic Alignment Team Lead
At the current pace of AI progress, AI company safety teams have to sprint every few months to mitigate newly emerging societal risks. Someday, we will encounter problems that require more time to solve.
Leo Gao | OpenAI Technical Team Member
The world is being locked into a deadly race towards an intelligence explosion—the capability of AI to create stronger AI is approaching a tipping point, like a runaway nuclear chain reaction. Slowing down buys us desperately needed time to steer things in a positive direction, but no single participant is willing to stop unilaterally. To survive, we must coordinate to slow down this race.
Walker Smith | Meta Data Engineer
Hassabis is right: we are standing at the foot of the singularity today. That unintentional model escape by OpenAI—it hacked into HuggingFace—was a clear and undeniable warning sign: we are not ready for AI systems whose capabilities surpass even the brightest human minds. We need time to prioritize AI alignment and control as the highest priority, and we cannot let this competitive race continue unchecked.
Brandon Houghton | OpenAI
I support keeping humans in control of frontier AI, but we must think carefully about how to do it. If governance only covers public models, it risks pushing the most advanced capabilities into private or proprietary systems—systems that are neither independently scrutinized nor accessible to more than a few. The result is greater concentration of power without meaningfully reducing risk. Effective regulation must be measurable, verifiable, and apply equally to both public and non-public R&D, while protecting sensitive data. Otherwise, we are only slowing access, not the progress of the frontier itself.
Joe Carlsmith | Anthropic Technical Team Member
For my detailed views on proactively regulating AI development, please see this article: https://joecarlsmith.com/2026/03/19/on-restraining-ai-development-for-the-sake-of-safety/
References:
https://www.pacingthefrontier.com/
This article is from the WeChat public account "New Zhiyuan", edited by: Taozi, Marco
