OpenAI has just set two specific dates: September 2026 for an intern-level AI research assistant, March 2028 for a completely autonomous researcher. The announcement came during a live broadcast in which Sam altman e Jakub Pachocki, the company's chief scientist, described the system as capable of “autonomously completing large-scale research projects.”
This is not an incremental update to ChatGPT, but a paradigm shift: from a passive tool that answers questions to an active entity that formulates hypotheses, designs experiments and draws conclusions. He whistles.
From Chatbot to Researcher: A Two-Step Roadmap
The first phase It involves a system that behaves like a laboratory intern: it is capable of following complex instructions, analyzing data, and conducting experiments under supervision. The second phase eliminate that oversight. Pachocki he clarified that this is not a human studying artificial intelligence, but aAI conducting scientific researchThe distinction is important.
The artificial researcher will be able (I rephrase: should be able) to identify gaps in scientific knowledge, formulate relevant research questions, design methodologies to answer them, run experiments, and finally interpret the results to produce new knowledge.. A monumental thing.
The technical strategy is based on two pillars. The first is continuous algorithmic innovation, building on the progress already demonstrated by GPT-4 models and reasoning systems like o3. The second, more radical, is the massive expansion of test time compute: giving models much more time to think about problems. How?
Current models already demonstrate impressive capabilities. They can handle tasks with time horizons of around five hours and compete with top human performers in contests such as International Mathematical Olympiad.
According to Pachocki, however, for the most ambitious scientific discoveries “it will be worth dedicating entire data centers of computational power to a single problem”.
The renovation that makes the impossible possible
The roadmap announcement comes on the same day that OpenAI completed its transition from a nonprofit to a public benefit corporationThis is no coincidence. The new structure removes the limitations imposed by the original statute and opens up new opportunities for raising capital.
According to, the nonprofit OpenAI Foundation will retain a 26% stake in the new for-profit entity and continue to lead the direction of research, with a $25 billion commitment dedicated to using AI to fight disease and manage security initiatives.
Altman explained that the restructuring creates a framework to support the ambitious timeline for AI researchers while maintaining a commitment to responsible development. The for-profit arm's ability to raise more funding means it can scale the infrastructure needed to achieve these scientific advances. OpenAI has committed to building 30 gigawatts of AI infrastructure. an investment estimated at 1,4 trillion dollars over the next few years.
AGI redefined: no longer a goal, but a process
Altman directly addressed the issue ofArtificial General Intelligence (AGI), arguing that the term has become “enormously overloaded” and that it will be “a process that will unfold over the course of several years.” Instead of trying to accommodate all possible definitions of AGI (impossible how to uniquely define consciousness), OpenAI has concrete goals. The one for March 2028 is “to have a real researcher Automated AI and defining what that means, rather than satisfying everyone with one definition of AGI.”
It is a pragmatic approach that shifts the focus from theoretical debates to practical skills. An autonomous AI researcher would still represent a huge, perhaps decisive, qualitative leap: a system capable not only of processing information but of genuinely generating new knowledge, just like a human scientist does.
Precedents: When AI is already doing research
The idea is not entirely new. A project called “The AI Scientist”, developed in collaboration with the Foerster Lab for AI Research ofOxford University, has already demonstrated a system for fully automated scientific discovery. The system generates research ideas, writes code, runs experiments, visualizes results, and produces complete scientific manuscripts. Each paper costs about $15, and the process can be repeated iteratively to develop ideas in an open-ended manner.
Even national laboratories are experimenting. Lawrence Berkeley National Laboratory uses robotic systems as Autobots investigate new materials for energy applications and quantum computing, dramatically reducing material validation times for batteries and electronics.
OpenAI says these goals align with the company's overall drive to advance scientific research and enable AI to make discoveries faster than human researchers, tackle complex problems beyond current human capabilities, and dramatically accelerate technological innovation across multiple fields such as medicine, physics, and technology development.
If the "hundred-gun" question is: when will AI become autonomous enough to not need us? OpenAI has just provided an answer: three years. Perhaps less, if progress accelerates as it has so far.
In the meantime, someone should start thinking about what all those PhD students will do when their main competitor doesn't need either sleep or grants.
