1. Invest in projects, not just papers.

Many graduate students place great value on publishing papers in the early stages of their research career. This is an important step for learning research, exploring initial topics, and demonstrating progress. But in the long run, what matters more than the number of papers is the impact of your research and your ability to communicate the big picture.

"Don't think of your research as a collection of individual papers. Think of it as a larger vision, a sub-field, or a paradigm you want to lead. Ask yourself: what change does your research want to make?"

One way to give that vision concrete form is to structure your research around open-source artifacts — models, systems, frameworks, or benchmarks. It takes more effort than simply running experiments and quickly uploading a repo, but it goes a long way toward ensuring the coherence and usefulness of your research.

2. Choose problems that are timely, have high headroom, and high fanout.

Not every paper needs to be a long-term investment. But if you're looking for directions to grow into a large project, consider these three criteria:

1. Timeliness: Find problems that are likely to become "hot" in 2–3 years but haven't yet entered the mainstream.
2. Fanout: Choose problems where your findings can influence multiple downstream sub-problems.
3. Headroom: Find problems where solving them could yield results that are 20× faster or 30% more effective.

For example, the ColBERT research in late 2019 started from the observation that applying BERT to retrieval was highly inefficient. The problem was timely, had headroom for a 1000× efficiency improvement, and had high fanout as a core component of retrieval systems.

"Efficient retrieval is a 'foundational' problem. Everyone needs to build on top of a retriever, but almost no one wants to build the retriever themselves."

3. Think two steps ahead and iterate quickly.

Once you've found a problem, don't just focus on the immediate solution. Instead, practice thinking two steps ahead. Anticipate the path people will commonly take, identify the limits of that path in advance, and research ways to overcome them.

In the case of ColBERT, encoding a document as a single vector was the "obvious" approach at the time. But thinking two steps ahead means asking where the single-vector approach will hit its limits. This kind of thinking led to innovations like the Late Interaction paradigm.

"Many people will follow the obvious approach. But think ahead to where that approach will hit its limits."

Iterating quickly is also crucial. Rapidly testing early problems and incorporating feedback increases your chances of solving the harder problems that follow.

4. Publicize your research and take the initiative in popularizing your ideas.

Once you've found a good problem and produced exciting results, before moving on to the next paper, share your research and actively engage with people. Don't stop at simply publishing a paper — use the big picture and open-source artifacts to continuously advance your ideas.

"A good idea doesn't end with a single presentation. You need to communicate it repeatedly, in many different contexts, so that people can truly absorb it."

The first step in publicizing research is posting a paper on arXiv and writing a tweet or post to announce it. The key here is not merely to announce the paper's existence, but to clearly and concretely communicate the core argument.

5. Tips for growing open-source research: convert interest into community.

Open-source projects don't end with uploading code. You need to build a project that has usability, utility, and accessibility. Consider the following milestones:

• Milestone 0: Make it usable Enable other researchers to run your code and reproduce your experiments.
• Milestone 1: Make it useful Design the project so a broader audience can make use of it.
• Milestone 2: Make it accessible Technically precise documentation alone is not enough. Provide documentation and examples that allow users to approach the project easily.
• Milestone 3: Show the limits of alternatives and be patient It takes time for people to recognize a problem and accept a solution.
• Milestone 4: Understand and engage diverse user groups Expand the project to serve both experts and general users.
• Milestone 5: Grow into a community Welcome contributions and discussion, and support the community in growing organically.
• Milestone 6: Expand into collaborative, modular sub-projects Modularize the project to create opportunities for new researchers to contribute independently.

"Open-source research requires both good research and good open-source artifacts. Striking this balance is hard, but when done right, it is deeply rewarding."

6. Keep investing in the project through new papers.

Open-source projects and research are not separate endeavors. On the contrary, an open-source project provides opportunities to intuitively recognize new problems, find collaborators, and effectively distribute research results.

For example, ColBERT did not end as a single paper — it expanded into more than 10 papers, covering topics like better training methods, memory optimization, and faster retrieval infrastructure. DSPy similarly spawned multiple papers on programming abstractions, prompt optimization, and more.

"Good open-source artifacts provide modular opportunities for new researchers and contributors to grow independently."

Key Keywords

• Project-centered research
• Timeliness, headroom, fanout
• Thinking two steps ahead
• Open-source artifacts
• Usability, utility, accessibility
• Community growth
• Sustained research investment

This piece offers a concrete guide on how an open-source-centered approach to AI research can achieve greater impact. Don't let your research end with publishing a paper — focus on continuously advancing it through a larger vision and a growing community. 🚀