By 2028, OpenAI intends to have a multi-agent AI system that functions as an entire research laboratory - one that can take on scientific problems in math, physics, biology, chemistry, and even business and policy. The system would work on anything expressible in text, code, or whiteboard diagrams, and it would do so largely without human guidance.

"I think we are getting close to a point where we'll have models capable of working indefinitely in a coherent way just like people do," chief scientist Jakub Pachocki told MIT Technology Review in an exclusive conversation, as reported by "Hvylya". "Of course, you still want people in charge and setting the goals. But I think we will get to a point where you kind of have a whole research lab in a data center."

OpenAI is not alone in this race. Anthropic's co-founder Jared Kaplan has said that fully automated AI research could be as little as a year away, while Anthropic CEO Dario Amodei has described building "the equivalent of a country of geniuses in a data center." Google DeepMind founder Demis Hassabis has voiced a similar vision since at least 2022.

Pachocki claimed that OpenAI already has most of the pieces in place. He pointed to GPT-5, the model powering Codex, which researchers have used to find new solutions to unsolved math problems and push through dead ends in biology, chemistry, and physics. "Just looking at these models coming up with ideas that would take most PhD weeks, at least, makes me expect that we'll see much more acceleration coming from this technology in the near future," he said.

The ambition raises urgent questions about oversight. Pachocki himself acknowledged that the safety challenges grow as systems become more autonomous. "Until you can really trust the systems, you definitely want to have restrictions in place," he said, adding that powerful models should run in sandboxes and that chain-of-thought monitoring - where models document their reasoning as they work - would become the primary safeguard.

Doug Downey of the Allen Institute for AI called the idea exciting but cautioned that current models still make frequent errors when chaining tasks together. OpenAI released GPT-5.4 two weeks ago, and Downey noted he has not yet tested it against his earlier benchmarks.

Also read: "Anti-Social Media": Why ChatGPT and Claude Push Public Opinion Where Populists Don't Want It.