Tips for a successful internship during your PhD

by Muhammad Khalifa (PhD student, Michigan AI)

Introduction

During my master’s and PhD studies, I secured many research internships, some of which were up to 10 months long. My internships experience spanned large companies, startups, and industrial labs. Most of my internships led to publications in top-tier AI conferences, yet they involved a fair amount of failing. To learn from these failures, I developed a set of rules or guidelines that I rely on when applying or selecting internships. In this article, I will share these with you, hoping they will help you navigate your own internship experiences.

This article tackles three key aspects when it comes to working internships: (i) why do an internship in the first place, (ii) how to choose an internship offer, and (iii) how to maximize your chances of success during the internship, and internship risks to be aware of. The article is concerned with research internships—internships where you work on a research project with the aim of publishing a paper. I am not referring to other types of internships (e.g., software engineering internship), since these are generally not recommended if you are already doing research.

Disclaimer: The advice in this article is based mainly on my own experience. I expect you to take this advice with a tiny pinch of salt, adapt it to your own internship, and make it your own. 

1. Why do an internship?

Internships can substantially contribute to your growth as a researcher. I recommend aiming for at least two research internships during your PhD, even if you do not plan to pursue a career in industry afterward. But why, you may ask—invest time and effort in internships when you could focus entirely on your PhD research? I believe there are at least three key benefits to research internships during your PhD. Let’s explore them one by one.

1.1 Chance to explore different topics

Internships can be an opportunity to go beyond the boundaries of your primary PhD focus, providing a low-risk, high-reward avenue for research exploration. Have you ever thought of research ideas that seemed very intriguing, wrote them down in your research journal hoping to get back to them in the future, but never did? Luckily, internships can give you the “luxury” to pursue some of these ideas sitting lonely in your research journal—all without the pressure associated with it being directly related to your PhD success. 

Personally, I have always found it easier to feel peace with failing during my internships. Yes, of course I am occasionally stressed during an internship, but overall it feels that the stakes of failing are low. For instance, I can always resume my PhD research after the summer. 

1.2 Industry exposure and connections

If you intend to stay within academia after your PhD, connections with industry may prove invaluable in the future (e.g., once you are a faculty member with your own lab and are looking for collaborations). If you are on the fence about industry vs. academia, then internships will give you a taste of the industry environment, giving you a more complete picture of what to expect. A bonus to connections is the exposure to different industry tools/paradigms, some of which could transfer to your PhD research. I will elaborate on these below. 

1.3 Access to resources

It is more common for tech companies now to have significantly more compute power beyond what is possible in academia. An internship, therefore, will give you access to the horsepower you may need for your project. Especially for those working with Large Language Models, having access to more compute is almost always preferable. But resources from industry are not simply limited to compute. Many companies will cover your conference trip costs, gym membership, home desk improvements, or more. Personally, I have used much of the money I received from my internships to improve my home working environment–-something that was not easily possible with my income as a PhD student. 

Now that you are (hopefully) convinced of the value of internships, the next step is finding a matching position. Let’s delve into that.

2. Applying and choosing an internship

2.1 Before applying

If you are early in your research career and have not made much impact yet, relying solely on your resume when applying may not be the wisest thing to do. AI internship positions have become increasingly competitive. For example, one of the labs that I interned at received over 1000 applications, but they could only accept 60 interns. That is around a 6% acceptance rate!

But I will provide one piece of advice here: to maximize your chances of getting an internship at strong teams in big companies, you will really want to consider networking. By networking, I do not mean cold-emailing researchers at different companies and asking them for internships; I mean talking to people in-person at conferences and events, which is much more effective. This way, it is much easier for them to remember who you are when you later email them asking about internship possibilities.

Whether you have a few internship offers to choose from, or you are still thinking about internship applications, there are a few aspects that you should carefully consider before making a decision.

2.2 Selecting the mentor

Selecting the internship mentor plays a major role in the success and enjoyment of your internship (much like selecting a PhD advisor, albeit with shorter-term consequences). First, you want to make sure they have experience mentoring an intern. Mentorship is a skill that takes time and experience to build. Having a strong research portfolio alone is not a sufficient indicator that someone will be a good mentor to you. Second, you need to make sure that the type of work you will do is appealing to you. One way to judge is to take a look at their recent papers with other interns and ask yourself if that is the type of work you’d like to do during your internship. I want to stress that you want to look at papers where an intern is the first author; papers written solely by company employees may not give you an accurate picture of work based on internship projects. Last but not least, if possible, reach out to previous interns that worked with that mentor and ask them about their experience. 

2.3 Selecting the Topic

The second most important thing is the internship project topic. There are usually two scenarios: either the topic is pre-determined, or it’s up for discussion. In my experience, the mentor/manager has a vague idea about the type of problems they want to explore, but nothing set in stone. Still, this fuzzy idea can still help you formulate a project direction. 

There are two strategies that I’d like to highlight when it comes to selecting internship topics. The first is picking a topic that aligns with your PhD research. In this case, you likely have the advantage of a few years’ expertise under your belt, and could easily minimize internship time needed to get familiar with a new topic. The disadvantage is that you end up missing out on expanding your research “palette”. But why would that be an issue? In the current state of AI/LLM research where divergent ideas are moving the field, PhD students need to be more of generalists. Having a portfolio of diverse research projects is preferable, and you may need to reconsider before passing on an internship because it is outside the circle of what you currently know. 

I would recommend that you explore earlier in your PhD, and less towards the end, in a simulated annealing sense. For example, if you do a total of 3 internships during your PhD, the first internship could be exploratory, while the other two are more aligned with your research.

3. During the internship

3.1 Tackling the research problem

Keep in mind that 3 months are not insufficient to complete a strong research project from start to finish. This means that time management during your internship is crucial for success. I will not go into detail about time management, but I will provide some tips on managing the primary internship period. 

First, you should focus most of your energy in these 3 months on finding a clear problem setup. While having a reasonable problem setup for your research problem is always important, it is even more crucial for an internship due to the timeline. But what do I mean by “setup” here? Three things: the problem, experimental setup, and type of approach—in that order.

When you and your mentor are in the process of agreeing on a research problem, you need to strike a balance between the problem significance and the time needed to make satisfying progress. This is easier said than done: it is generally hard to predict the time required to work on a research problem, and there is no easy shortcut that I can provide you with. However, I recommend you iterate over problem selection at least 2–3 times, while talking to others—both inside and outside the company, and get their feel for the problem impact/feasibility tradeoff.

Having a good experimental setup involves building a project codebase and choosing the right train/eval tasks and metrics to quantify how well your approach is doing. You need a quantifiable measure of progress. Let’s make this more concrete. Suppose the problem you chose is how code-training of language models makes them worse at non-code-related tasks (e.g., machine translation). You have a codebase that evaluates different models on translation to and from five different languages. At this stage, you have the problem and the setup covered.

Now you can shift gears to considering the type of approaches/methodology you’ll explore during your internship. You might look into prompting methods, fine-tuning approaches, or inference techniques—to name a few. By selecting a family of approaches rather than a particular one, you keep your options open while maintaining a manageable scope. The rest of the internship time can be spent iterating on your method. With the research problem and setting now in place, it is easy to measure how much progress you are making. 

Note that the advice provided above primarily applies to typical ML/NLP “method” papers. Some of this guidance may not be entirely relevant for projects with different natures, e.g., empirical studies or dataset papers. However, the general template of finding a compelling problem and a clear experimental setup still generally holds.

3.2 Making the most out of the internship environment

Different companies use different tools, frameworks, and infrastructure. One of your goals during an internship is to learn the tools that 1) are essential for your internship project and 2) will still be relevant for your own research after the internship ends. To give an example, I have learned a lot about Docker during my internship at AI2, as it was necessary for accessing the infrastructure. As Docker is a popular tool that I will likely need in future projects, dedicating time to learning it was worth the effort. Needless to say, certain tools are specific to a company’s infrastructure, so I recommend dedicating minimal effort to mastering those.

In my article on lessons learned during a PhD, I stressed the importance of talking to people about your research. This advice is even more true for internships. Your goal should be to engage with as many people as possible—other interns, researchers on your team or other departments, or anyone working on related or unrelated projects. Personally, I have found it surprisingly easy to make friends with other interns during my internships.

You never know which of these people you will meet at a conference, reach out to during job searches, or inspire your next project. During my last internship at Cohere, I set a goal to connect with at least one new person every week. While I did not always achieve this goal, this mindset kept me actively seeking opportunities for conversations.

4. Pitfalls

I do not want to give the impression that internships are always a good move. Like any worthy life endeavor, internships come with their own risks. It’s crucial to weigh the pros and cons, and realize that there may be situations where forgoing internships may align better with your goals.

The most obvious risk is that your internship project might not turn out as expected. Even if you apply all the tips mentioned and more, you could still fail to achieve the project goal or solve the intended problem. Internship compatibility (topic, mentor, environment etc.) involves many factors outside your control. I have heard internships being described as a lottery, which means you should keep an open mind and understand that an unproductive internship does not reflect poorly on your abilities as a researcher.

Another pitfall I’ve seen others fall into is going on internships at the wrong time. Remember, an internship is a 3–5 month time investment—time that might be better spent in a different avenue. For instance, if you’re approaching your PhD defense and need to work on your thesis, it might not be wise to pursue an internship during that particular summer. A discussion with your advisor and a deep reflection over your priorities can help you make a good decision. 

Did I miss something? Do you find my advice to contradict with your experience? I want to hear your thoughts!Muhammad Khalifa is a fourth-year PhD student in Computer Science & Engineering at the University of Michigan, advised by Lu Wang and Honglak Lee. He works on Large language model reasoning, attribution, and controlled generation. Contact: khalifam at umich dot edu.

More about the author: https://mukhal.github.io