OpenAI sets sights on fully autonomous AI researchers

If current coding tools feel like fast-typing assistants, OpenAI now wants to build something with an actual PhD.

OpenAI Chief Scientist Jakub Pachocki revealed that the company is shifting its primary research focus toward a massive new challenge: building a fully automated "AI researcher" capable of solving complex problems completely on its own.

OpenAI is treating this as its north star for the coming years. The goal is to bring together reasoning models, agentic workflows, and interpretability research until an entire research lab can run straight out of a data center.

From intern to full lab by 2028

The roadmap has two distinct milestones. By September, OpenAI plans to deploy an "autonomous AI research intern" designed to tackle assignments that usually take a human researcher a few days to finish.

That intern will pave the way for a fully automated multi-agent research system slated for 2028. OpenAI expects this future setup to take on frontier problems across mathematics, physics, biology, chemistry, and even complex policy dilemmas.

The shift is already changing how OpenAI works internally. Pachocki, a self-described Vim purist who used to avoid generative autocompletion, admitted he now spends weekends orchestrating groups of Codex agents to run experiments that previously took him a full week of manual coding.

The compounding error problem

Pulling this off in the real world is going to be significantly harder than generating lines of code.

Doug Downey, a research scientist at the Allen Institute for AI, pointed out that complex research requires chaining many tasks together. When you chain operations over days, even minor inaccuracies compound rapidly, quickly derailing the entire project.

Then comes the governance nightmare. If an autonomous agent runs a multi-day scientific program without check-ins, catching mistakes or malicious behavior after the fact is dangerous.

OpenAI's current defensive bet is chain-of-thought monitoring. The models jot down their internal reasoning on a digital scratchpad as they work, while separate supervisory models monitor those notes in real time. For the foreseeable future, Pachocki says these high-powered systems will remain tightly sandboxed.

Why it matters

We think this pivot marks a major inflection point in how AI labs frame progress.

For the past few years, the race was about human productivity boosts—answering emails, writing boilerplate, or generating charts. By aiming for an autonomous researcher, OpenAI is shifting the focus toward recursive self-improvement, where models help invent the next wave of science and technology.

For researchers and engineers, this means the nature of work will shift from hands-on execution to high-level delegation and evaluation. The real hurdle is reliability: until someone solves the error-compounding problem across multi-day trajectories, trusting an AI intern with genuine scientific discovery remains a risky bet.