OpenAI's Chief Scientist: AI interns are not far from us.

In September 2026, OpenAI plans to launch its first "AI research intern".

This is the timeline given by Jakub Pachocki, the Chief Scientist at OpenAI, in a recent interview. There are only 5 months left until then.

In his view, the positioning of AI has shifted from a mere tool to a "human employee": how long it can operate independently, whether it can handle ambiguous tasks, and how much human intervention it requires have become new standards for measuring models.

This means that the role of AI is starting to change from a daily assistant to an intelligent agent for task completion.

When this transformation occurs, what is reshaped is not only the form of AI but also the way humans work.

Section 1 | Capability Evolution: From Programming to Scientific Research

This wave of AI technology is easily misconstrued by the outside world as simply "more powerful programming tools".

However, within OpenAI, the changes have gone far beyond code - writing and are advancing towards a more complete workflow.

Jakub Pachocki revealed that they have been using Codex in a large number of actual programming tasks. The real change is not that code can be written faster, but that many processes that originally required step - by - step human operation can now be directly executed by the model. You no longer need to type every line of code yourself. You just need to set a goal and let it complete an entire task, and humans are only responsible for the final check and fine - tuning.

In addition to the improvement in work efficiency, the model's capabilities in the fields of mathematics and science have also been greatly enhanced in the past few months.

AI can now handle more complex problems. It not only provides the final answer but also offers the intermediate reasoning process and even valuable problem - solving ideas. In some tests, it has been able to approach or even reach the difficulty level of international math competitions. More importantly, in research - type challenges such as "first proof", the model has found the problem - solving path in a very short time, which would have taken humans days or even longer to figure out.

What does this indicate? It shows that AI can not only execute instructions but also start to have the ability to explore the unknown.

Programming has become a breakthrough because it naturally has a clear feedback mechanism, and it is easy to verify whether it is right or wrong. The same logic applies to mathematics. Because the quality of results can be accurately judged in these fields, the model has the opportunity to continuously correct its process and gradually improve its capabilities to a higher level. Once this ability stabilizes, it begins to transfer to more complex and real - world problems, such as hypothesis verification, experimental design in scientific research, and even the generation of new ideas.

Jakub emphasized in the interview that what they are increasingly concerned about is no longer the model's score on a certain benchmark test, but whether it can really help people solve problems in the real world. Whether it is research, engineering, or other work with actual commercial value, the measurement standard is shifting from "how many questions are answered correctly" to "whether a task can be carried through to the end".

On the surface, the transition from code - writing to research seems to involve a huge industry span, but in fact, there is a very clear evolutionary path at the bottom: first, establish basic capabilities in verifiable fields, then extend this ability to more complex real - world scenarios, and finally deeply participate in high - level work that originally required long - term human investment.

Section 2 | Two Gaps and a Timeline

Precisely positioning the current AI as an "intern" is the well - thought - out conclusion of Jakub Pachocki.

Because there are two most fundamental gaps between an "intern" and an "independent researcher".

First, it is the duration of independent work.

If you assign a task to an experienced human researcher, you only need to point out a general direction, such as "improve the model's capabilities further". The researcher can independently break down the problem, find a path, keep trying and making mistakes, and even continue to advance the work for weeks or months. However, the current AI cannot do this. Although its continuous operation time is getting longer and it can complete a complete task, it still needs clear guidance and correction from humans at key nodes.

Second, it is the ability to handle ambiguous problems.

In the real world, most work does not have a standard answer at the beginning. There is often only a rough direction, a preliminary idea, or even just a vague goal, which needs to be clarified and defined continuously during the process. This is exactly where human researchers are irreplaceable: they are not only executors but also pathfinders, who can decide "what to do next".

At this stage, AI is better at handling tasks that have been clearly defined. For example, verifying a specific idea, running an experiment with a new method, or analyzing data of known results. Within these clear boundaries, it is doing better and better, and its efficiency even surpasses that of humans. However, once it is required to define problems and choose exploration directions on its own, its performance is still unstable.

This is why it is called an "AI intern".

Jakub provided a very clear timeline: before September 2026, achieve an AI system at the level of a "research intern"; before March 2028, advance it to an "AI researcher" with higher autonomy.

In his words, most of the technical elements for achieving this leap already exist, and the current core work is to integrate them effectively.

We are standing on the eve of a revolution.

Section 3 | Outsource Execution, Leave Decision - Making to Humans

When AI crosses the boundary of simply "answering questions" and starts to be able to complete a task in a closed - loop manner, this change is no longer limited to the level of tool upgrade but touches on a more fundamental proposition: how humans should do things.

In the interview, Jakub Pachocki pointed out that the rapid improvement of the model's short - term capabilities is directly changing the speed of research progress. This means that many long and arduous tasks that originally required a lot of human effort are now being accelerated or even completely reset.

In the past, the standard rhythm of advancing a complex research or project was often: humans put forward an idea → verify it personally → keep trying and adjusting → make slow progress. A large amount of time was consumed in tedious execution details, and there were very few moments for deciding the direction.

However, as AI can take over most of the execution work, humans no longer need to complete every step personally and can focus more on three things: determining the direction, breaking down tasks, and evaluating results. "Execution" itself is gradually being outsourced to tools.

When a person is no longer occupied by specific operations, their core value is no longer "how fast they can do things" but "what to do" and "how to coordinate". Facing the same tools, some people can only use them to speed up piece - rate work, while others can use them as a lever to greatly expand the business boundary. The gap between them will become larger and larger.

This reorganization of the workflow has become a reality in top - tier laboratories. Jakub admitted that his team now needs to focus more on choosing "which directions are worth investing in" rather than distributing resources evenly as before. The reason is that when the cost of execution drops significantly, what is truly scarce is no longer manpower but the right choice.

Similar changes are also affecting more work scenarios.

For example, for tasks such as writing code, doing analysis, and writing proposals, which originally required a complete accumulation of experience, the threshold is decreasing. A person does not need to master all the details to complete a considerable part of the work. However, at the same time, the requirement for overall control is higher. You need to know what the goal is, which results are effective, and which only seem good on the surface.

Therefore, if we put the "AI intern" back into the real world, the shock it brings is not just adding a capable assistant. It is forcing a shift in the way humans work: from "doing things personally" to "designing and managing AI to do things".

At this time, the question for every white - collar worker becomes: Can you decide what is worth doing and how to let AI do it well?

Conclusion | Only One Thing Left

The "AI intern" is not a vision of the future but a reality that is happening.

The boundary of AI's capabilities is shifting from "helping you do a little" to "completing a task for you".

What we need to adapt to is a complete transformation of our professional roles: when AI takes over more and more execution work, humans are left with only one thing:

Decide what is worth doing.

Reference materials:

https://www.youtube.com/watch?v=vK1qEF3a3WM

Source: Official media/Online news

This article is from the WeChat official account "AI Deep Researcher". Author: AI Deep Researcher, Editor: Shen Si. Republished by 36Kr with permission.