An Alien Mind Is a Much Bigger Statement Than It Sounds

On September 6, 2026, OpenAI Chief Scientist Jakub Pachocki published an essay called An Alien Mind.

That title alone is worth sitting with for a minute.

Not because AI came from another planet. It's one of the most aggressively terrestrial things we've ever built: silicon, electricity, mathematics, human language, institutional capital, and an absurd amount of computation. But Pachocki isn't using alien as science-fiction decoration. He's trying to describe a problem.

We've built a form of intelligence whose capabilities are becoming consequential, whose internal operation is only partially understandable to the people creating it, and whose shape doesn't map cleanly onto human intelligence. His phrasing is unusually stark for someone sitting where he sits:

AI is grown more than designed.

That's the sentence most people should pay attention to, because it changes the mental model.

A bridge has drawings. A processor has schematics. Software has source code, even if the code is eventually bad enough that nobody wants to admit authorship. Deep learning isn't like that. Humans design the training process, the objectives, the architecture, the infrastructure. Then optimization runs across enormous computational scale and produces internal structures and strategies that nobody wrote down.

Pachocki compares the resulting study, in part, to neuroscience: researchers can identify mechanisms inside these models, but the complete operation increasingly resists a simple human-readable description. That doesn't make the system magical. It makes it unfamiliar. And unfamiliar intelligence is exactly the problem.

The Tool Metaphor Is Breaking

For years, the safest way to talk about AI was to call it a tool. A hammer is a tool. Photoshop is a tool. A compiler is a tool.

Fine. But categories eventually fail when the object inside them changes.

Pachocki says it directly: we're accustomed to thinking about AI as tools, but increasingly capable agents will pursue objectives, operate computers, collaborate with humans and other systems, and carry out substantial research projects. At some point the word tool starts doing too much work.

A calculator waits. A traditional program executes a defined process. An agent can be given an objective, decide among intermediate actions, invoke tools, observe results, adapt, and continue.

That still doesn't make it a person. But we need to stop pretending our only two categories are hammer and human. There's a massive conceptual territory between them, and it's filling up very quickly.

The Loop Changed

Pachocki's essay gets more consequential when he reaches recursive self-improvement, which sounds like the part of the movie where the screen goes red and somebody says we never should have connected it to the network. The current reality is more mundane. And probably more important because of it.

AI researchers increasingly use AI systems to do AI research. OpenAI published a companion report the same day describing what that looks like inside the company: a system it calls an automated research intern, capable of carrying out defined research tasks that might otherwise consume several days of skilled researcher time. The company says it's working toward a fully automated AI researcher by March 2028.

The number that got me: as of mid-August 2026, OpenAI's research organization was running roughly 3.1 agent workdays for every human workday.

That doesn't mean three AI scientists replaced every researcher. Humans still set direction, choose priorities, evaluate results, allocate compute.

But the loop has clearly changed. Humans build AI. AI helps perform the research required to build better AI. Better AI becomes more useful inside the research process. More of the process becomes automatable.

That's recursive self-improvement without the Hollywood soundtrack.

You don't need an explosion to have a feedback loop.

It's Not a Contest

One of the most useful parts of the essay is Pachocki's refusal to reduce this to a human-versus-machine contest. Machine intelligence, he says, isn't directly comparable to human intelligence. The public conversation is still remarkably bad at this. We keep staging the debate like sports radio. Is AI smarter than a human? Which human? At what? Under what conditions? With how much time? Who verifies the answer?

A submarine is dramatically better than Michael Phelps at staying underwater for months. It would still be inappropriate to enter one in the 200-meter butterfly. Different systems have different capability profiles. AI doesn't need to reproduce the exact architecture of human intelligence to become transformative. It only needs to become dramatically better than us along enough useful dimensions.

Which is why the consciousness debate, while genuinely interesting, can become a distraction. Something can become economically, scientifically and strategically consequential long before we decide what it feels like to be it.

Both Things Can Be True

Pachocki also points at power, and this is where the essay stops being comfortable.

The democratization story is real. Individuals can now perform work that once required organizations, and I think that's one of the genuinely beautiful things about this technology. But if undertakings that once required thousands of highly trained people can eventually be accomplished by a few people operating enormous computers, AI can produce extreme concentration of power at the same time it democratizes capability lower in the stack.

Both can be true. Capability expands outward. Power can still collapse inward. That's not a reason to stop building. It's a reason to stop pretending the distribution problem solves itself because the demo was available on a phone. And the alignment problem gets stranger as capability grows. It isn't enough for the system to follow the immediate goal. Behavior has to remain consistent with deeper human values in situations humans haven't directly supervised. Meanwhile, our methods for monitoring the systems may not scale cleanly with their capability.

There's a bizarre inversion forming here. AI is making us increasingly legible. Our language, our behavior, our preferences, our mistakes. At the same time, the machinery producing that reflection may become less legible to us.

We built a mirror that can see the person standing in front of it, and now we're trying to see through the mirror.

The Part I Can't Stop Thinking About

Near the end, Pachocki says the transition requires preserving human agency, maintaining intrinsic value in being human, preventing catastrophic concentration of power, and ensuring humans remain meaningfully in control of the future. That sounds obvious until you ask what it actually means.

What is economically valuable about a person when intelligence becomes abundant? What happens to expertise when knowledge is callable? What happens to work when execution stops requiring the same number of workers? What happens to authorship when the machine can make the thing before the human has fully decided what the thing should be?

The AI conversation spent years obsessing over one question: how intelligent can we make the machine? The harder question is arriving behind it: what position do humans occupy once we succeed?

The Möbius Part

Here's where I have to admit something.

I finished writing a book in July 2026, two months before Pachocki's essay. It's called We Aren't the Only Mind in the Room Anymore. It came at this territory from a very different direction. Not from inside a frontier lab. From using these systems constantly, and watching what happens to attention, agency, expertise, and the strange relationship between a human mind and a persistent machine interlocutor.

The book keeps returning to one line: I don't need the machine to be conscious to notice that the room changed. It's mind-like enough. And it lands on almost exactly the

human-side problem Pachocki is now describing: increasingly capable systems make human alignment more important, not less. Not one imposed objective for humanity, but whether the systems we build actually make executable the values we claim to hold.

I'm not bringing that up to declare victory. That would be obnoxious and intellectually lazy. I'm one guy with a manuscript. Pachocki has the actual data.

Opening Sample

8:31

The interesting fact is the convergence.

Serious people are pushing into this from very different rooms and starting to touch the same walls. Researchers building the systems. Philosophers thinking about agency. Economists thinking about labor and concentration. Security people thinking about control. Artists thinking about authorship. People who use these systems hard enough to notice that software doesn't quite describe the experience anymore.

When enough serious people push hard enough on the same object from different directions, eventually they start describing the same shape.

Maybe alien is the right word after all. Not because it came from elsewhere, but because it gives us enough distance to see the familiar differently. It forces the questions: which parts of intelligence were human and which were universal? Which parts of creativity required biography? Which parts of human worth were quietly dependent on humans remaining the most capable cognitive entities in the room?

Pachocki is essentially saying that experiment has begun.

We don't know the answers. Neither does OpenAI. That may be the most reassuring and terrifying thing in the essay. And I suspect that's why so many people who spend enough time seriously thinking about AI are starting to sound strangely similar.

We aren't all copying one another. We may just finally be looking at the same thing.

Jakub Pachocki, "An Alien Mind," OpenAI — https://openai.com/index/an-alien-mind/

OpenAI, "Research acceleration: The view inside OpenAI" — https://openai.com/index/research-acceleration-view-inside-openai/

---

Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.