en.Wedoany.com Report on Mar 25th, OpenAI recently disclosed an ambitious technology roadmap: it plans to build an "AI researcher" capable of autonomously solving complex problems by 2028—a fully automated multi-agent research system that can independently handle various scientific research tasks ranging from mathematics, physics, biochemistry to policy analysis. This goal, referred to by OpenAI's Chief Scientist Ilya Sutskever as the company's "North Star" direction for the coming years.
According to the plan, the first phase of results will be realized this September—OpenAI will first launch the "Autonomous AI Research Intern" system. This incremental path demonstrates OpenAI's technical approach of attempting to evolve AI research capabilities from a laboratory tool to an autonomous research subject.
In an exclusive interview with MIT Technology Review, Sutskever revealed that this plan marks OpenAI's new attempt to advance AI technology from content generation towards autonomous research, and is also an important strategic move in facing competitors like Anthropic and Google DeepMind. He stated that realizing an "AI researcher" is not only a technical challenge but also a restructuring of the scientific research paradigm itself.
In terms of technical accumulation, OpenAI has been paving the way for this goal in multiple aspects in recent years. In January of this year, the company released the agent application Codex, which can instantly generate code, perform complex computational tasks, analyze documents, generate charts, organize emails, and summarize social media content. Currently, Codex has become a standard tool for OpenAI's internal employees, assisting in code development and problem-solving. Sutskever described Codex as the prototype of the "AI researcher" and stated that it will achieve disruptive innovation in the future.
At the model capability level, OpenAI's technological evolution over the past few years has laid the foundation for an autonomous research system. From GPT-3 to GPT-4, the model's ability to handle complex problems without intervention has achieved a qualitative leap in duration. The "reasoning model" technology launched in 2024, by introducing "chain-of-thought" training, enabled AI to learn to proceed step-by-step and backtrack upon errors like humans. Currently, OpenAI is using challenging problems from mathematics and programming competitions to conduct "intensive training" on the models, aiming to enhance their ability to process ultra-long texts and decompose multiple subtasks, ultimately enabling them to solve real-world scientific research problems.
Sutskever believes the key to automated research lies in the system's ability to operate for extended periods with reduced human intervention. He explained: "Our goal is to develop a research intern system that can take on tasks that would normally require several days of human effort." By training models to gradually solve problems and backtrack from errors, reasoning models can maintain coherent work over relatively long periods, which is a key capability for advancing towards an autonomous AI researcher.
In terms of application exploration, OpenAI is currently more focused on research relevant to the real world. According to reports, researchers have already used the GPT-5 model that powers Codex to discover solutions to several unsolved mathematical problems and have made progress on certain challenges in biology, chemistry, and physics. These achievements validate the potential of AI in assisting research and provide practical groundwork for subsequent fully automated research systems.
Sutskever emphasized that OpenAI is continuously iterating its models, hoping that through sustained technical accumulation, it can demonstrate that the AI researcher possesses scientific reliability even before deeply engaging in real-world research. From "research intern" to "AI researcher," OpenAI is attempting to build an incremental bridge between AI and scientific discovery, bringing fundamental change to the research paradigm.
