OpenAI Says AI Getting Closer to Research Intern Level Capabilities

OpenAI is making significant progress toward one of its major milestone goals: developing artificial intelligence systems capable of functioning at the level of research interns. During an appearance on the "Unsupervised Learning" podcast on Thursday, OpenAI's chief scientist Jakub Pachocki discussed recent advances that suggest the company is on the right track.

Pachocki highlighted several key areas of progress, including breakthroughs in coding capabilities, advances in mathematical research, and improvements in physics applications. These developments indicate that AI systems are becoming increasingly capable of handling complex, multi-step technical work with minimal human intervention. "I definitely see this as a signal that something here is on track," Pachocki stated during the podcast discussion.

The critical distinction between a research intern and a fully autonomous researcher, according to Pachocki, lies in how long a model can work independently. "The way I would distinguish a research intern from a full automated researcher is the span of time that we would have it work mostly autonomously," he explained, emphasizing that longer task horizons represent the key metric for measuring progress in this area.

At an October company livestream, Pachocki outlined OpenAI's ambitious internal targets: building an "AI research intern" by September 2026, with a fully autonomous AI researcher to follow by March 2028. OpenAI CEO Sam Altman subsequently posted on X acknowledging that the company "may totally fail" at these goals, but stressed the importance of transparency given the potential impact of such technology.

Pachocki pointed to concrete examples of progress already underway. Coding agents like Codex are now handling a substantial portion of the company's programming work, demonstrating that AI tools are increasingly capable of autonomous code generation. Additionally, Pachocki identified math benchmarks as a "north star" for improving model reasoning, since these provide easy-to-verify performance metrics. "We've seen this explosive growth of coding tools," he noted. "For most people, the act of programming has changed quite a bit."

Looking ahead, Pachocki identified the near-term challenge as developing systems capable of tackling specific technical tasks with greater autonomy, utilizing more computational resources, and operating for extended periods without interruption. "For more specific technical ideas, like I have this particular idea how to improve the models, how to run this evaluation differently, I think we have the pieces that we mostly just need to put together," he stated.

However, Pachocki was careful to clarify that current AI systems are not yet ready to operate completely independently at the level of a full researcher. "I don't expect we'll have systems where you just tell them, 'go improve your model capability, go solve alignment,' and they will do it, not this year," he cautioned, setting realistic expectations for the timeline of autonomous AI researcher capabilities.