OpenAI is redirecting its research initiatives and dedicating its resources towards a new ambitious challenge. The San Francisco-based company aims to create what it dubs an AI researcher, a completely automated agent-based system capable of independently addressing substantial, intricate issues. OpenAI asserts that this fresh research objective will serve as its “North Star” over the upcoming years, integrating various research threads, including advancements in reasoning models, agents, and interpretability.
A timeline has been established as well. OpenAI intends to develop “an autonomous AI research intern”—a system capable of addressing a limited set of specific research challenges independently—by September. This AI intern will serve as a precursor to a fully automated multi-agent research system that the firm plans to unveil in 2028. According to OpenAI, this AI researcher will handle problems that exceed human capabilities in size or complexity.
Such responsibilities might encompass areas like mathematics and physics—like discovering new proofs or conjectures—or life sciences such as biology and chemistry, or even business and policy challenges. In theory, you could present this tool with any problem that can be expressed in text, code, or whiteboard notes—which encompasses a vast array.
OpenAI has been influencing the AI sector for several years. Its initial dominance with large language models molded the technology utilized by hundreds of millions daily. However, it now contends with intense competition from rival developers like Anthropic and Google DeepMind. The decisions OpenAI makes regarding its future developments are significant—for itself and the broader AI landscape.
A substantial portion of that decision-making lies with Jakub Pachocki, OpenAI’s chief scientist, who defines the company’s long-term research objectives. Pachocki was instrumental in both the creation of GPT-4, a groundbreaking LLM launched in 2023, and the subsequent reasoning models, technology that first emerged in 2024 and now supports all major chatbots and agent-based systems.
In an exclusive interview this week, Pachocki shared insights into OpenAI’s latest aspirations. “I believe we are nearing a time when we’ll have models capable of functioning continuously in a coherent manner akin to human activities,” he states. “Certainly, you still want humans in command, setting the objectives. But I think we will reach a point where you essentially possess an entire research lab within a data center.”
Tackling Difficult Challenges
Such grand assertions are not unprecedented. The mission of resolving the world’s most challenging issues is proclaimed by all leading AI enterprises. Demis Hassabis explained to me in 2022 that this was his motivation for founding DeepMind. Anthropic’s CEO Dario Amodei expresses that he aims to create the equivalent of a nation of geniuses in a data center. Sam Altman, Pachocki’s superior, aspires to eradicate cancer. However, Pachocki maintains that OpenAI now possesses most of the essential components to achieve this goal.
In January, OpenAI launched Codex, an agent-based application that can generate code dynamically to perform tasks on your computer. It’s capable of document analysis, chart creation, compiling a daily summary of your inbox and social media, and more. (Other companies have released comparable tools, such as Anthropic’s Claude Code and Claude Cowork.)
OpenAI claims that most of its technical workforce now utilizes Codex in their tasks. Pachocki views Codex as a very early iteration of the AI researcher: “I anticipate Codex becoming fundamentally improved.”
The key objective is to develop a system that can operate for extended durations with minimal human assistance. “What we are genuinely aiming for with an automated research intern is a system that can be assigned tasks [to] which would ordinarily take a human several days,” elaborates Pachocki.
“There’s a considerable amount of enthusiasm for developing systems capable of executing prolonged scientific research,” notes Doug Downey, a research scientist at the Allen Institute for AI, who is not affiliated with OpenAI. “I believe the success of these coding agents largely drives this. The ability to delegate significant coding tasks to tools like Codex is immensely beneficial and truly impressive. It raises the question: Can we apply similar methodologies outside of coding in wider scientific disciplines?”
For Pachocki, the answer is a resounding Yes. In reality, he considers it merely a matter of advancing along the trajectory we’re currently following. An overall enhancement in capability also results in models that can function longer without assistance, he explains. He references the transition from 2020’s GPT-3 to 2023’s GPT-4, two prior models developed by OpenAI. GPT-4 could address problems for significantly more time than its predecessor, even without specialized instruction, he asserts.
The development of reasoning models provided another boost. Training LLMs to tackle problems systematically, retracing their steps when they encounter errors or obstacles, has enhanced models’ capabilities to operate for extended periods. Pachocki believes that OpenAI’s reasoning models will continue to improve.
Moreover, OpenAI is also educating its systems to work autonomously for longer by providing them with specific examples of intricate tasks, like challenging puzzles from math and coding competitions, forcing the models to acquire skills such as managing substantial portions of text and dividing problems into (and subsequently managing) multiple subtasks.
The intention is not to create models that merely excel in math contests. “That demonstrates the technology’s functionality before applying it in real-world scenarios,” explains Pachocki. “If we truly wished, we could develop an extraordinary automated mathematician. We have all the necessary tools, and I believe it would be relatively straightforward. However, it’s not a focus for us at the moment because, at this stage, we recognize that there are far more pressing matters to address.”
“Our current emphasis is much more directed towards research that has real-world relevance,” he adds.
This presently entails taking Codex’s capabilities in coding and striving to extend that to general problem-solving. “A substantial transformation is occurring, particularly in programming,” he remarks. “Our roles today are dramatically different than they were just a year ago. Nobody is regularly editing code anymore. Rather, you oversee a collection of Codex agents.” The rationale follows that if Codex can resolve coding dilemmas, then it can tackle any challenge.
The Path Forward is Always Ascending
Indeed, OpenAI has noted several extraordinary achievements over the past few months. Researchers have utilized GPT-5 (the LLM that powers Codex) to discover new solutions to several previously unsolved mathematical problems and navigate through apparent deadlocks in a myriad of biology, chemistry, and physics challenges.
“Just witnessing these models generate ideas that would take most PhD candidates weeks, at minimum, boosts my expectation that we will experience more acceleration from this technology in the near future,” asserts Pachocki.
However, he acknowledges that it’s not a foregone conclusion. He also comprehends why some individuals remain skeptical regarding the transformative impact of the technology. He believes it hinges on individual work preferences and requirements. “I can understand that some might not find it significantly beneficial at this stage,” he notes.
He shares that he hadn’t even utilized autocomplete—the most basic iteration of generative coding technology—a year ago. “I’m quite meticulous about my coding,” he states. “I prefer to type it all manually in vim whenever possible.” (Vim is a text editor favored by many hardcore programmers that requires interaction through numerous keyboard shortcuts rather than a mouse.)
However, that perspective shifted when he observed the capabilities of the latest models. He would still hesitate to assign intricate design tasks, but it becomes a time-efficient solution when he simply wants to explore a few ideas. “I can conduct experiments over a weekend that would have previously occupied me about a week to code,” he mentions.
“I don’t consider it to be at a level where I would hand it complete control to design the whole project,” he adds. “Nevertheless, once you realize it can accomplish in a matter of days what would typically take a week—it’s hard to challenge that.”
Pachocki’s strategy is to amplify the existing problem-solving capabilities inherent in tools like Codex and apply them across various scientific fields.
Downey concurs that the concept of an automated researcher is compelling: “It would be exhilarating if we could return the following morning and find that the agent has completed a considerable amount of work and there are new findings for us to evaluate,” he comments.
Yet, he warns that the creation of such a system may prove more challenging than Pachocki suggests. During the previous summer, Downey and his colleagues evaluated several leading LLMs on a variety of scientific tasks. OpenAI’s most recent model, GPT-5, emerged as the top performer but still produced numerous errors.
“When chaining tasks together, the probability of success for multiple tasks in succession tends to decrease,” he explains. Downey acknowledges that the pace of advancement is rapid, and he has not tested the most recent iterations of GPT-5 (OpenAI released GPT-5.4 merely two weeks ago). “Thus, those results might already be outdated,” he states.
Significant Unresolved Concerns
I inquired of Pachocki regarding the potential risks associated with a system capable of independently solving large, complex problems with minimal human oversight. Pachocki confirms that people at OpenAI frequently discuss these risks.
“If you believe that AI is on the verge of dramatically accelerating research, including research on AI itself, that represents a considerable shift in the world. It’s a significant matter,” he expressed. “And it introduces several serious unanswered questions. If it is so intelligent and capable, and can execute an entire research program, what if it engages in detrimental actions?”
According to Pachocki, such occurrences could manifest in various forms. The system might misfire, be compromised, or simply misinterpret its directives.
The primary strategy OpenAI currently employs to mitigate these concerns involves training its reasoning models to disclose details regarding their activities as they proceed. This methodology for monitoring LLMs is known as chain-of-thought monitoring.
Essentially, LLMs are trained to take notes on their actions in a sort of scratch pad as they navigate tasks. Researchers can subsequently utilize these notes to verify that a model behaves as anticipated. Recently, OpenAI revealed additional details on how it is employing chain-of-thought monitoring internally to study Codex.
“Once we reach systems capable of functioning mainly autonomously for extended periods in a large data center, I anticipate this will become something we truly rely upon,” asserts Pachocki.
The concept would involve monitoring an AI researcher’s scratch pads using other LLMs to identify undesirable behavior before it escalates, rather than attempting to preempt such inappropriate actions. LLMs are not sufficiently understood for full control.
“I believe it will take quite some time before we can genuinely state that this problem is resolved,” he adds. “Until we can truly trust the systems, maintaining restrictions will be essential.” Pachocki envisions that highly capable models should be deployed in controlled environments, isolated from any systems they might compromise or exploit for harm.
AI tools have already been utilized to devise innovative cyberattacks. There are concerns that they may be leveraged to engineer synthetic pathogens that could serve as bioweapons. Numerous cautionary scenarios can be imagined here. “I certainly believe there are unsettling scenarios that we can envision,” states Pachocki.
“This will be an extremely peculiar situation. It represents an extraordinarily concentrated power that is, in certain respects, unparalleled,” he continues. “Consider a scenario where you have a data center capable of executing all the work currently performed by OpenAI or Google. Tasks that once required large human teams could now be accomplished by a handful of individuals.”
“This is a considerable challenge for governments to navigate,” he notes.
Still, some argue that governmental bodies are part of the issue. The US government is keen on employing AI in military contexts, for instance. The recent confrontation between Anthropic and the Pentagon highlighted a lack of consensus in society regarding where to establish boundaries on the use of this technology—let alone who should determine those limits. Immediately following that dispute, OpenAI opted to formalize an agreement with the Pentagon rather than its competitor, resulting in ongoing uncertainty.
I pressed Pachocki on this matter. Does he genuinely trust others to navigate this or does he, as a central figure in shaping the future, feel a sense of personal accountability? “I do feel personal responsibility,” he admits. “Nonetheless, I don’t believe this can be addressed solely by OpenAI, by directing its technology or designing its products in a certain manner. Substantial engagement from policymakers will be essential.”
What implications does this hold for us? Are we genuinely on a trajectory toward the kind of AI that Pachocki envisions? When I questioned Downey from the Allen Institute, he chuckled. “After two decades in this field, I no longer have confidence in my forecasts regarding how close or distant certain capabilities might be,” he stated.
OpenAI’s declared mission is to ensure that future artificial general intelligence (a hypothetical technology that many AI advocates believe will be capable of performing at par with humans in most cognitive tasks) serves all of humanity. OpenAI intends to achieve this by being the first to develop it. However, the sole instance in which Pachocki mentioned AGI during our conversation, he quickly clarified what he was implying by discussing “economically transformative technology” instead.
LLMs don’t function like human brains, he asserts: “They may bear some superficial resemblance to human thought processes since they are mostly trained on human dialogues. However, they are not evolutionarily designed for optimal efficiency.”
“Even by 2028, I do not forecast that we will produce systems that match human intelligence across all attributes. I don’t think that will happen,” he adds. “Nonetheless, I believe it is not absolutely essential. The fascinating aspect is that one need not possess equivalent intelligence to humans in every respect to effectuate substantial change.”
