On Sunday, two men said something about artificial intelligence.
The first one sells the chips. He posted four words. AGI has arrived. He congratulated the company that built the newest model, mentioned it ran on his hardware, and added that four hundred thousand more chips were coming online next.
The second man is the chief scientist at the company that built it.
That same day, he published an essay. And what he wrote was this:
This is a time that calls for extreme caution. I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.
Same technology. Same Sunday.
One of them is selling. The other one is worried.
Now, I want you to know who the second man is, because it matters.
His name is Jakub Pachocki. He’s not a protester. He’s not a professor writing from the outside. He’s not somebody who got fired and went public.
He is the chief scientist of OpenAI. He led the development of GPT-4. He took the job in 2024 when the last chief scientist walked out the door. Right now he runs research at the company that’s shipping faster than anybody in the world.
Four days before he wrote that essay, his company released its most powerful model ever.
And here is the sentence that stopped me cold.
He wrote that he believes no lab has solved alignment and monitoring well enough to keep scaling responsibly at maximum speed for much longer.
No lab.
That’s not a man pointing at his competitors. His company is in that sentence. He put it there himself.
Now let me tell you what he means by monitoring, because this is the part that ought to get your attention, and it isn’t hard to follow.
When one of these systems works on a problem, it does something like showing its work. It writes out its reasoning as it goes. And the people who build these things watch that scratch work. That’s how they check what the machine is actually doing, as opposed to what it says at the end.
That’s been the main tool. Read the reasoning, catch the problem.
And it works on one simple bet. As long as nobody grades the scratch work, the machine has no reason to clean it up. It’s got no incentive to hide anything, because nothing it writes there counts against it.
Pachocki says that bet is wearing out. He gave three reasons.
The reasoning is starting to blend into the parts they do have to supervise.
The machines are getting better at thinking about their own thinking, which means better at shaping it.
And they’re getting smarter without writing anything down at all.
Take those three together and you get a plain statement. The window they’ve been looking through is closing, and they don’t have another one.
That’s the chief scientist. Not a critic. The man in charge of the research.
So what does he want?
This is the part I keep turning over.
He asked for minimum safety thresholds. Legally required, or required by international agreement. Enforced by outside auditors or government agencies. Labs would have to meet the bar before they’re allowed to keep scaling or put a model out.
Read that again.
The chief scientist of the company that just shipped the most powerful model in the world is asking for somebody outside his company to have the power to tell him no.
I’ve been writing all week about rules that can’t be failed. Twelve companies have published safety frameworks. Somebody graded them against standards borrowed from nuclear and aviation, and the middle of the pack came in at eighteen percent. The finding was that you can’t tell from those documents whether a company has kept its promises.
That’s a rule you can’t fail.
What Pachocki asked for is the opposite. An outside auditor who can stop you is a rule you can fail. That’s what makes it a rule.
When the man running the research asks for the thing that could shut him down, that’s not nothing.
Now I have to give you the other side, and there is one.
Two things about that essay ought to sit uncomfortably.
The first is that his company already went through this. In August, they paused their biggest training run because something turned up about the new model’s ability to do damage in cybersecurity. They didn’t stop because they wanted to. They stopped because of what they found.
So when he asks everybody else to slow down voluntarily, he’s asking the industry to volunteer for what his own company did under pressure.
The second is smaller and it stuck with me. In the middle of an essay about how nobody is prepared, he mentions that the new model is significantly better aligned than the last one.
That’s a product claim, sitting inside a safety warning.
Critics noticed that too. One researcher pointed out that OpenAI’s answer to the problems its own models are causing is to build internal AI agents to research those problems. Another said the warnings would carry more weight if the company shared more of what it’s actually seeing, instead of asking us to take the alarm on faith.
And I’ll say plainly what I say every time. I use AI to write. The company that makes the one I use is a competitor to the one in this story, and it gets named in the coverage. Weigh that.
But here’s where I come down.
I don’t need to know his motives. I can’t see inside the man, and neither can you.
What I can see is what he put in writing, on his own company’s website, with his name on it.
He wrote that nobody is prepared.
He wrote that no lab has solved this.
And he asked for somebody outside to hold the ruler.
Take those three together and something has changed. Not the technology. The claim.
For four years the industry’s position has been: trust us, we’ve got this. On Sunday the chief scientist of the leading lab said, in effect, we don’t have this, and you shouldn’t take our word for it, and there ought to be a test we could fail.
The man selling the chips said the race is won.
The man running the research said nobody knows what’s coming.
I know which one I’d listen to. And it isn’t the one with four hundred thousand chips on order.
” Attic Thoughts”-library – Intelligent People Assume Nothing
Contact: micvicfaust@gmail.com
This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.
© 2026 The Faust Baseline LLC | All Rights Reserved
