Internships are valuable, but they should complement doctoral training rather than replace the deeper work of learning how to define problems, organize resources, and lead AI systems from ideas to real-world impact.
Many PhD students admire internships, especially internships at large technology companies. I understand this view, and I also believe internships have real value. They expose students to industrial constraints, large-scale engineering systems, user feedback, product deadlines, and organizational realities that are difficult to reproduce inside a university.
This essay is therefore not an argument against internships. It is an argument for asking what kind of growth an internship should serve.
In ecosystems such as the Shenzhen Loop Area Institute, there are also mechanisms to support students in local industry internships, including opportunities with companies such as Tencent and Huawei. I believe this kind of collaboration will become more flexible and more inclusive over time.
But we should still ask a harder question: if doctoral training is effectively replaced by four continuous years of big-company internships, is that really the best path for PhD students?
Four years of internships may give students a decent lower bound. They may become familiar with corporate workflows, learn to ship under pressure, and acquire useful engineering habits. But it may also sacrifice the upper bound of what a PhD can become in the AI era. In the worst case, it simply becomes an early, lower-paid version of full-time employment: students enter the industrial assembly line four years earlier, but do not necessarily become the kind of people who can define new directions, lead new systems, or create new value.
The point is not to reject internships. The point is to place internships inside a larger educational design. A good PhD should not merely train students to complete tasks that others have already defined. It should help them build a deeper capability: to define problems where no clear problem exists, find paths where no ready path exists, organize resources when resources are incomplete, and push a system from idea to real-world deployment.
An internship can help a student see industry. It can show how models are integrated into products, how teams coordinate across functions, how infrastructure constraints shape research, and how user needs differ from benchmark scores. For many students, this is an important form of reality testing.
However, industry exposure is not the same as doctoral training. A company usually has its own goals, business priorities, product roadmaps, and resource boundaries. A student may be assigned to a well-defined task, a submodule of a larger system, or a short-term delivery target. This can be valuable, but it may not force the student to ask the most important research questions.
A PhD is not only a job preparation program. It is a long-form training process for independent judgment. It should give students enough space to explore, fail, revise, and build their own taste for important problems.
This distinction matters more in AI than in many previous fields. In the AI era, implementation is becoming cheaper. Code generation, prototyping, data processing, and even parts of experimentation are increasingly assisted by AI. What becomes more valuable is not merely the ability to execute a given instruction. What becomes more valuable is knowing which problem should be solved, why it matters, what resources it requires, and how to turn it into a functioning system.
PhD students need to learn how to define their own problems.
Every new scenario, especially every concrete real-world scenario, may contain new opportunities. When we worked on medical AI, some people thought the direction was not general enough and might not be ideal for career development. But as large models rapidly advance in relatively general capability tracks such as coding and mathematics, concrete domains such as medicine, embodied intelligence, AI for science, finance, economics, and social science are becoming increasingly important.
These domains are not easy. Their data flows are often fragmented. Their problems are not naturally well-defined. Their feedback loops may be incomplete, delayed, expensive, or difficult to measure. That is exactly why they are valuable.
The most important problems rarely appear as clean questions on an exam sheet. They are usually hidden inside the disorder of the real world. The ability to abstract a clear problem from that disorder, explain why it matters, and organize people and resources to solve it is one of the most important capabilities in the AI era.
New environments also create new training opportunities. For example, when working with domestic AI chips, students may encounter constraints and failures that rarely appear on mature NVIDIA-centered stacks. These difficulties are not merely obstacles. They can become the ladder through which students learn systems thinking, hardware-software co-design, debugging under uncertainty, and resource-aware innovation.
The old research pattern is becoming less sufficient. In the past, a student might receive a standard dataset, enough GPUs, and a clear benchmark. The task was to design a clever architecture, loss function, or algorithm, then push the state of the art by a measurable margin.
That era is not completely gone, but it is no longer enough.
In the AI era, students increasingly need to assemble the full production chain themselves. They may need to:
The most competitive people in the future will not only be those who can train a model well. They will be those who can organize the whole loop: problem, data, experts, engineering, scenario, feedback, and deployment.
This is also why a PhD should not be reduced to executing isolated corporate tasks. Real AI leadership requires the ability to build a new loop when the loop does not yet exist. Students need the time and responsibility to practice this.
AI-era PhD training should not emphasize coding ability alone. Empathy and social responsibility are equally important.
We can threaten to unplug an AI system, cancel its coding plan, or shut down its tool access. The AI itself does not care in a human sense. But people are different. Human beings have finite lives. They care about family, health, dignity, recognition, purpose, and the meaning of their work.
This difference matters for product design.
As AI makes technical implementation easier, product realization will become cheaper. But genuinely good product judgment will still depend on human understanding: understanding society, understanding concrete scenarios, understanding users, and understanding the consequences of technology.
Empathy is not sentimental decoration. It is a source of product power. A student who understands human needs can choose better problems, design better workflows, notice hidden failure modes, and build systems that people actually want to use. A student with social responsibility can resist building technically impressive but socially harmful systems.
In the AI era, technical implementation will become less scarce. Value judgment, problem selection, and responsibility will become more precious.
The students we should train for the new era are not merely junior data cleaners, annotation workers, or task executors for enterprises. They should become people with technical leadership.
This leadership should begin while they are still in school. A strong PhD student should gradually learn to lead a small team of three to five people on a serious research project. They should learn to set a clear goal, decompose a complex task, coordinate contributors, manage uncertainty, deliver results, and expand the influence of the work.
The goal is not only to publish a paper. The goal is to connect research, industrial application, and product deployment into a closed-loop engineering capability.
Internships can help students see industry, but internships alone cannot replace the formation of judgment. A four-year internship path may not leave enough space to cultivate these traits. AI-era leadership is more likely to emerge from collaboration between university training and industrial practice, not from replacing one with the other.
After students enter companies full time, many of them will face short-term pressure, unreasonable expectations, and endless demand queues. It is easy to become a “tool person” or a foot soldier inside a large system. True leadership appears precisely in such moments: the ability to see through confusion, open a path for the team, let sunlight enter the room, and solve problems one by one with others.
Such people do not appear from nowhere. They are often shaped by real responsibility during school: leading projects, making decisions under uncertainty, working with teammates, and learning to carry consequences.
The value of a university is that it provides a growth space that companies cannot fully replace.
In a university, students can communicate freely with people from different backgrounds. They can attend international conferences, encounter frontier ideas, and meet global peers. They can collaborate with different companies, institutions, hospitals, and domain experts, rather than being limited to the current goals and resource boundaries of one employer.
This open space is not a luxury. It is part of how judgment is formed.
Industry internships help students see the real world. University training helps students develop the ability to judge the real world. These two functions are different, and both are important.
A good PhD program should therefore not isolate students from industry. It should help them enter industry with stronger questions, return from industry with sharper judgment, and connect industrial reality with long-term research ambition.
For PhD students, the key question should not be simply: “Should I do an internship?”
A better question is: what capability should this internship help me build?
Before choosing an internship, students may ask:
If the answer is yes, the internship can be extremely valuable. If the answer is no, the internship may still be useful, but it should not be confused with the core of doctoral training.
The future does not need PhD students who are merely trained earlier to fit into existing pipelines. It needs people who can create new pipelines, define new problems, build new systems, and connect AI with real human needs.
That is the kind of leadership doctoral education should cultivate in the AI era. Internships can be part of that journey, but they should not be the whole journey.