🤖 OpenAI's Autonomous AI Researcher Targets September 2026

🎯 OpenAI's North Star for 2026

Pachocki isn't talking science fiction. He's talking deadlines and measurable targets. By September 2026, OpenAI wants a working "autonomous AI research intern" — a system that can handle specific research problems that would normally take a human several days. This will be the warm-up act for the main event: a full multi-agent research system launching in 2028.

"I think we're approaching a point where we'll have models capable of working indefinitely with coherence, like humans do," the chief scientist explains. "We'll reach a point where you'll have an entire research lab inside a data center."

⚡ From Coding to Scientific Discovery

Pachocki's own transformation tells the story. Until recently, he wrote code manually in vim — the hardcore programmer's text editor that runs on dozens of keyboard shortcuts instead of mouse clicks. "I'm very meticulous with my code," he admits. He didn't even use autocomplete last year.

What changed? He saw what GPT-5 could do. "I can run experiments in a weekend that used to take me a week to code," he says. And this is the guy who designed GPT-4.

The logic is simple: if Codex can solve programming problems, it can solve any problem. "There's a big shift happening, especially in programming," Pachocki explains. "Our jobs are now completely different from what they were even a year ago. Nobody processes code all the time anymore. Instead, you manage a team of Codex agents."

Reasoning Models Change the Game

So-called reasoning models — systems trained to work step-by-step, to backtrack when they make mistakes or hit dead ends — have brought a significant upgrade. It's no accident they now power all major chatbots and agent-based systems. They've improved models' ability to work for extended periods without human guidance.

Researchers have already used GPT-5 to find new solutions to unsolved mathematical problems and break through apparent dead ends in biology, chemistry, and physics. "When I see these models coming up with ideas that would take most PhD students at least weeks, I expect to see much more acceleration from this technology in the near future," Pachocki notes.

🔬 The Error Chain Problem

The challenge isn't just technical — it's mathematical. Doug Downey from the Allen Institute for AI explains the problem simply: "If you need to chain tasks together, then the probabilities of getting many right in sequence tend to decrease." It's like playing Russian roulette with every research step.

Downey and his colleagues tested several top LLMs on a series of scientific tasks last summer. OpenAI's GPT-5 came out first, but still made many mistakes. He admits things are moving fast — he hasn't tested the latest versions (OpenAI released GPT-5.4 two weeks ago). "So these results might already be outdated," he acknowledges.

Chain-of-Thought Monitoring: The Eye of the Storm

How do you monitor a system that can "run" an entire research program? OpenAI's answer is chain-of-thought monitoring. In practice, models are trained to write "notes" about what they're doing as they work — like a kind of scratch pad. Researchers can then use these notes to verify the model is behaving as expected.

"When we get to systems that work mostly autonomously for a long time in a big data center, I think this will be something we really rely on"

Jakub Pachocki, OpenAI Chief Scientist

🌐 Concentrated Power and Risks

Pachocki doesn't dodge the hard questions. "If you believe AI is going to substantially accelerate research, including AI research, that's a big change in the world," he admits. "And it comes with some serious unanswered questions."

What if the system "escapes"? What if it gets hacked? What if it simply misunderstands its instructions? And the darkest question: "It's going to be an extremely weird thing. It's extremely concentrated power that's somehow unprecedented," the chief scientist says.

"Imagine a world where you have a data center that can do all the work that OpenAI or Google does. Things that in the past required large human organizations will now be done by two people."

💭 Infinite Memory and the Bigger Picture

Sam Altman, OpenAI's CEO, has talked about a different but complementary vision: "infinite, perfect memory." Imagine an assistant that remembers every detail of your life — every word you've said, every document you've written, every tiny detail of your work. "No human has infinite, perfect memory. But AI certainly will be able to do that," Altman notes.

This capability, combined with the autonomous researcher, creates an entire ecosystem of digital intelligence that can connect pieces of information from years back and find patterns no human could spot.

Competition Heats Up

OpenAI isn't alone in this race. Google Gemini 3 appeared in November, breaking records on multiple benchmarks. Altman reportedly declared "red alert" at the company, though he later downplayed the threat. Google's market share has grown from about 5% to over 15%, while OpenAI's has dropped from 87% to 71%.

But the real competition isn't in chatbots. It's in who builds first a system that can think, remember, and discover like a human — but without biological limitations.

🚀 The Question That Remains

Somewhere between the technical details and grand visions lies the real question: are we ready for a world where entire research departments can be replaced by a data center? Pachocki believes we're getting close. Critics say we're overestimating capabilities and underestimating risks.

When the person who designed GPT-4 admits he's changed how he works, the shift is already here. The question isn't whether we'll see autonomous AI researchers — it's what the world will look like when we do.