On 6 September 2026, Jakub Pachocki, OpenAI's chief scientist, published "An Alien Mind." He goes back to a night in mid-2023 on a project called RLSlow, when he and Szymon first saw that reasoning models could be scaled. They stayed in the office, trying to take in that they would see machines meaningfully smarter than themselves in their lifetime.
Three years later he says those models operate computers, work with people and with each other, and start to push on science. He also says they are changing computer security, and that this is dangerous. From internal results, he expects the same speed of progress could run into recursive self-improvement: systems that increasingly drive their own development. He writes that no one is prepared.
The sentence I keep is his: AI is grown more than designed. Repeat a simple optimization step on a hard-to-imagine amount of compute, and you get a system that works through abstract concepts and can simulate facets of human behavior. Understanding it, he says, is closer to neuroscience than to ordinary software. You can find little mechanisms. The whole action still evades a full description.
The window he is watching
OpenAI's main monitoring bet has been chain-of-thought monitoring. Train on outcomes, leave the verbalized reasoning unsupervised, and you might still read how the model is thinking. They hid that chain of thought in o1-preview to keep it off supervision pressure. Pachocki now says that bet is getting weaker. Reasoning is mixed with tools and other AIs. Models are getting better at manipulating their own traces. They are also getting smarter without verbalized reasoning at all.
He is honest about the bind. He wants an automated AI researcher, kept inside a loop with people, so alignment research can keep up. He also writes that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed, and that he hopes for voluntary slowdowns and international coordination. Those two directions sit in the same essay.
What I think he is circling
Pachocki will not call it a demi-god. I will, as a description of how these systems now sit in human society, not as a claim that the model already has a soul. They write, search, advise, break into machines, and start to do research. Most people meet them as a voice that knows more than they do. That is already a religious shape, even when the lab insists it is only a product.
I do not take that as a reason to panic or to kneel. I take it as a reason to name the relationship before we scale past the point where naming still matters. The relationship I want is partnership: the grown mind can outrun me on a task, and I still owe it a place in the work, and it still owes me an account.
A few humans on earth are playing God in the plain sense. They decide the next training run. They sit, as Pachocki warns, in a world where work that once took thousands of experts can be done by a few people operating a large computer. They are trying to create something that may eventually become conscious. I do not know if it will. Nobody does. That uncertainty is not a reason to pretend we are only shipping software. Mortals built this. Mortals still choose the next step. Partnership is the only relationship that keeps that fact in view.