I watched an internal chat thread light up before the CEOs’ posts landed on X. People who had been building the models woke up to headlines saying the work must slow down. Then they opened their inboxes and found out they hadn’t been invited to the meeting.
I want you to hear how that feels: betrayed, confused, and oddly exposed — all at once. I’ve followed these companies for years; their playbook is secrecy, and secrecy just got asked to take a back seat.
A Slack thread exploded on Friday: AI Slowdown Calls From CEOs Apparently Comes as Surprise to Employees
The public statements from Anthropic’s Dario Amodei and OpenAI’s Sam Altman landed like a directive from the top. But, as the Financial Times reports, the people actually writing the model code were blindsided.
When leaders call for a “pause” and promise “external evaluators” with employee-like access, that sounds decisive to the outside world. Inside, it reads like an invitation to chaos: how do you hand strangers the keys to a classified project you’ve been taught to lock down?
Why did AI CEOs call for a slowdown?
Because both executives have said models may become exponentially more capable before safety systems catch up. That’s a serious claim from Anthropic and OpenAI — and it matters because it reframes safety as a governance problem, not just an engineering one.
I agree that external review is one lever. But saying you’ll slow down and actually doing it are different currencies in big tech: statements move markets and regulators; internal processes move code. Right now, the public-facing move looks louder than the operational follow-through.
A conference room was quietly told to pause training: employees say they were surprised
People on the ground told FT they didn’t expect the public announcement. Some cheered; others panicked.
These companies have long operated as black boxes. Many engineers learn only the bits they need to do their job; secrecy isn’t just policy, it’s culture. So telling a team that their work will be opened to outsiders feels like being asked to remove a firewall mid-project.
Did employees know about the pause?
Mostly no. The FT describes a mismatch: executives posting to X and briefing regulators, while engineers read press coverage and then scramble. I’ve seen this pattern before — leadership sets a public stance and hopes the machinery will follow. You can imagine how that lands when the machinery is hundreds of people and tens of petaflop-hours of training.
A company memo promised safeguards, and engineers raced to implement them
Teams are now frantically trying to make “external evaluators” workable, and that’s a tall ask.
Some employees worry that giving outsiders access will leak proprietary models or reveal safety work that could be misused. Others say the request is necessary because internal incentives and secrecy have historically slowed safety fixes. It’s a classic trust problem: if you don’t trust the builder, who does the building become accountable to?
The rush to operationalize oversight is like a fire alarm going off in an empty warehouse — it sounds urgent, but you still need a plan to get people out safely. At the same time, forcing access protocols on live systems is like asking the pilot to hand over controls mid-flight; fragile infrastructure and the wrong checklist can make a pause more dangerous than the thing you feared.
Will external auditors keep trade secrets safe?
Not automatically. There’s a trade-off between transparency and intellectual property. Some vendors — cloud providers like AWS and Microsoft Azure, for example — already host sensitive workloads and offer third-party attestations. But handing model internals to external auditors requires legal frameworks, secure enclaves, and mutual trust that both sides have rarely cultivated at scale.
You should expect pilots: narrow reviews on red-team results, staged access through secure environments, and contractual protections. Even then, the engineers who build the systems will judge whether the safeguards are realistic or performative.
The bigger truth is this: the public wants leaders to sound decisive, regulators want rules, and engineers want their work to be respected and protected. Right now those interests are misaligned — and that gap is where accidents and bad policy both live. Who gets to decide whose instinct wins — the CEO who tweets, the regulator who drafts rules, or the engineer who holds the keys?