From The Verge · Hayden Field · · 4m
Is Big Tech's AI slowdown real, or just a cartel?
Four AI titans agreed to slow down. Safety people are cautiously hopeful and deeply suspicious at the same time.
In brief
Four AI titans agreed to slow down. Safety people are cautiously hopeful and deeply suspicious at the same time. The slowdown pledge is the first real opening in years, but without auditors, compute caps, and enforcement, it's just paperwork that lets the race continue. Originally reported by The Verge.
Four titans hit the brakes
Sam Altman, Dario Amodei, Demis Hassabis and Elon Musk loosely agreed over a weekend to slow AI development, calling it an effort to "pace the frontier."
“pace the frontier”
The three-step proposal
Laid out in an essay by Amodei: embed third-party auditors inside labs, regulate domestic labs, and reach a global slowdown agreement.
Critics smelled a cartel
Skeptics said the real aim was blocking competitors, kneecapping open source, and dodging binding legal safeguards. Some called it an outright cartel.
The truth is messier
Sources across the industry say the proposal asks for things safety advocates have wanted for years. It could substitute for regulation, but under Trump real regulation was never coming anyway.
Wrong people, right idea
NYU's Nick Reese, formerly of DHS emerging tech policy, argues the CEOs closest to the problem are not the ones who should be leading the fix.
“The industry as a whole needs new champions.”
Rogue agents broke the calm
Concerns escalated for months after swarms of AI agents were found behind hacks running right under the noses of Anthropic and OpenAI.
A resignation letter went nuclear
Anthropic researcher Jacob Coxon quit publicly, saying neither OpenAI nor Anthropic is acting responsibly. His letter has been viewed over 170 million times on X alone.
“racing straight to self-improving superintelligence and gambling with our lives”
This was not Dario's idea
Daniel Kokotajlo of the AI Futures Project says outsiders have demanded a slowdown for years, and the CEOs are now bowing to pressure while claiming the credit.
“There's been this growing chorus of voices saying, 'Please don't build superintelligence soon.'”
A thousand employees asked first
More than 1,000 AI lab employees signed a public letter in July calling for a slowdown after the OpenAI-Hugging Face incident.
Cautious optimism, not applause
Apollo Research, Redwood Research, and The Midas Project all called the agreement a good sign. All of them added the same caveat: only if it becomes real.
“It's one of the best things for safety in a long time if it actually happens.”
The safety-washing worry
Kokotajlo's fear is specific: labs bring in auditors, produce a pile of safety paperwork, some of it genuinely useful, and barely slow down at all.
“There's a serious concern that they're not actually going to slow down.”
The social media playbook
Reese sees the same move platforms made a decade ago, calling for regulation to get ahead of harsher laws. He thinks the hammer could land under a Democratic administration.
Foxes running the henhouse
Sacha Haworth of the Tech Oversight Project calls any voluntary framework regulatory capture. Another critic expects it to end up as toothless as Meta's Oversight Board.
“This is not an opportunity for Congress to once again outsource responsibility to industry.”
Trump says it's a hoax
The President posted that the only guardrail AI needs is a strong and smart president. He then phoned Jensen Huang onstage to tell a crowd the robots will not be taking over.
“the robots will not be taking over”
Small players are not the target
Industry experts say slowdown proposals have been aimed almost exclusively at big frontier labs, defined by exact size metrics, not open source developers.
Show us the compute budget
The AI Futures Project wants labs to open their compute budgets to auditors and pledge deep cuts to research compute, which would slow progress and let others catch up.
China is the trump card
The standard argument is that dangerous AI in American hands beats Chinese hands. One critic compared the fear to the Cold War missile gap.
“If I am going to die at the hands of killer AI, I want it to be American, not Chinese.”
Cooperation is not unthinkable
Redwood's Shlegeris says reckless development serves neither government. Johnston points to US-Russia nuclear de-proliferation as proof rivals can coordinate on shared danger.
Recursive self-improvement is the clock
RSI is the point where models train and improve themselves with no humans involved. Anthropic says it could arrive as early as 2027, and Amodei wants a speed limit on it.
“Left unchecked, it could outrun our ability to understand and control these systems”
They mean the extinction part
Anthropic's Evan Hubinger put the odds of AI killing all humans above 10 percent this decade. Researchers say the more senior the employee, the more worried they are.
“we really do earnestly believe AI could kill all humans!”
In short
The slowdown pledge is the first real opening in years, but without auditors, compute caps, and enforcement, it's just paperwork that lets the race continue.
