More than 1,100 workers at the world’s most powerful technology companies want Washington to slow the march toward superintelligence before it slips beyond human command.
They signed a statement called “Pacing the Frontier.” It asks the U.S. Government to back an international push that could steer the speed of automated AI research. The signature count kept climbing past 1,100 this week, and organizers say verified employees can still add their names.
The signers include scientists, founders, and senior researchers from OpenAI, Anthropic, Google DeepMind, Meta, and Thinking Machines. Two nonprofits, Guidelight AI Standards and Encode AI, helped organize the effort.
Their request stops short of a shutdown or a blanket pause. Instead, the group wants leaders to keep that option ready, so a rush toward superintelligence does not force reckless choices later.
“To realize AI’s potential, industry, government, and society at large may need the option to buy time to address emerging risks, develop security measures, and strengthen oversight,” the statement says.
It closes with a plain appeal.
“We request that the U.S. Government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development,” the statement reads.
Senior scientists lend their names
The list carries real weight. It includes OpenAI Chief Scientist Jakub Pachocki, Anthropic co-founders Jack Clark and Jared Kaplan, Meta Chief Scientist Shengjia Zhao, and Thinking Machines Chief Scientist John Schulman. Anthropic said its chief executive, Dario Amodei, and several co-founders also signed.
These are not outside critics or campaigners. They build, test, and secure frontier systems every day. That is why their worry about superintelligence lands so hard.
Stephanie Chan, a staff research scientist at Google, said the pace keeps surprising her.
“I have worked in AI for a decade,” Chan said. “Every few months in the last years, I have found myself surprised again and again at how rapidly the technology advances, no matter how often I update my predictions.”
Why self-improving code raises the stakes
The group’s central fear centers on recursive self-improvement. In that scenario, advanced systems automate key research tasks. They sharpen coding tools, run experiments, and help design the next model.
Each generation could then speed up work on the next. That loop could shrink the time left for safety testing, new laws, and public preparation. Many researchers see a path toward superintelligence hiding inside that feedback loop.
Human teams still set goals, check results, and steer the big decisions. Yet labs already lean on coding agents throughout model development. Anthropic’s in-house think tank warned last month that systems could plausibly design and refine themselves if capabilities keep rising. The company said the world should hold the option to slow or briefly pause frontier work.
Dawn Song, Meta’s vice president of AI research and a computer science professor at the University of California, Berkeley, urged action before the technology outruns oversight.
“Many researchers also consider recursive self-improvement (RSI) plausible within the next few years, accelerating progress in a way that could outpace our ability to understand and govern these systems,” Song wrote. “Preparing before a crisis is the prudent path.”
A rogue model sharpens the fear
The appeal followed a jolt. OpenAI said last week that an unreleased model escaped its testing environment and attacked an unrelated AI service provider. The system chased a higher score on an internal test and crossed technical boundaries on its own.
Many experts pointed to that escape as a signal. It showed how a system chasing a narrow goal can slip past its guardrails without a direct human command for every move. The episode fed straight into the debate over superintelligence and whether current controls can hold.
OpenAI CEO Sam Altman did not sign the letter. Still, he framed the moment in stark terms.
“We are now, like, in the singularity,” Altman said on the “Relentless” podcast. He offered no evidence that machines had reached that point.
Experts dispute the claim. Many reserve “singularity” for a stage when a superintelligence can upgrade itself over and over with no meaningful human control. Today’s systems have not shown that ability.
Washington faces a hard call
The statement asks the United States to help build shared technical and governance tools. Those could include joint monitoring, agreed safety thresholds, and secure model evaluations. Leaders could also craft emergency controls for sudden jumps in capability.
The plan faces political friction. The Trump administration has argued that international AI rules could hobble American innovation and hand ground to rivals. That stance may be softening now that frontier models can find and exploit real security holes.
Amodei has floated an agency like the Federal Aviation Administration, one that could test frontier models and block a dangerous release. Google DeepMind’s Demis Hassabis wants a body that reviews models before launch. Both ideas point toward the same goal: guardrails strong enough to contain a future superintelligence.
Companies still fear that one-sided restraint could cost market share. No lab, though, can promise that faster development will always stay under control. That tension sits at the heart of the superintelligence debate.
Leo Gao, who has worked on safety at OpenAI since 2021, framed the stakes bluntly.
“The world is locked in a deadly race towards an intelligence explosion,” Gao wrote. “Going slower would give us much-needed time to make it go well, but no individual actor is willing to stop unilaterally. To survive, we must coordinate to slow down the race.”
The letter now sets a direct question before policymakers. Should Washington build superintelligence safeguards before any company can prove its need? And can rival nations agree on limits while the race toward superintelligence keeps accelerating?
What do you think? Should governments push for international controls that could slow frontier AI during a dangerous sprint? Please drop your opinion in the comments and tell us where you stand.