AI is on track to become much more than just an assistant. OpenAI’s chief scientist Jakub Pachocki suggested that in the near future, AI models will be capable of carrying out original scientific research on their own — a leap beyond simply helping humans with existing knowledge, according to a report by Nature.com.
Pachocki, who took over as OpenAI’s chief scientist in 2024, says models are gradually moving towards being able to work with minimal human input. While tools like ChatGPT still depend on prompts and constant direction, OpenAI has already developed systems that can work independently for short durations and produce useful results. He says this could evolve rapidly if more computing power is applied to open-ended scientific challenges.
He points to Deep Research, a system designed by OpenAI, as an early example of this. Even with limited compute, the tool can already process and analyse large volumes of information unsupervised for several minutes. With more advanced models and greater resources, Pachocki says AI could soon contribute in fields like automated software development, hardware engineering, and even novel scientific discoveries.
At the heart of these capabilities is reinforcement learning — a process that teaches models through feedback and repetition. Pachocki explains that while AI models initially learn by absorbing massive datasets during pre-training, the real progress happens when reinforcement learning helps them develop strategies to solve complex tasks.
He says recent improvements in reasoning models are largely driven by how reinforcement learning is being used — not just to polish the model, but to help it develop its own way of thinking and decision-making. This is a change from earlier models that simply imitated human-like output based on patterns in data.
However, Pachocki also acknowledges that AI models do not think the way humans do. They don’t remember how or when they learned something, and they lack an understanding of time and experience. Still, their ability to simulate logical steps and solve problems makes them useful tools for tasks that require structured thinking.
As OpenAI works on more powerful models and prepares to release an open-weight version soon, Pachocki is focused on pushing the boundaries of what AI can achieve in science and technology. While the debate around artificial general intelligence (AGI) continues, he believes the real test will be how well these systems can operate independently and tackle problems we haven’t solved yet.