TLDR: As a junior technical AI safety researcher, do not try to write and publish your first paper alone; the chances of success on this are very low. Go to SPAR or another part-time research fellowship instead.
I facilitate the BlueDot Technical AI safety course, so I meet many people who want to transition to an AI safety technical researcher career. Usually, people are either from an engineering background or from research in other science areas. A reasonable first main goal for them is to publish their first AI safety research paper at a machine learning conference. Interestingly, often they consider the option to write and publish their first paper alone, before applying to any fellowships. Or they have applied to a high-tier fellowship like MATS, did not get accepted, realized that they need to improve their research skills and researcher’s CV, and decided to write a paper alone.
In this post, I argue that writing your first research paper alone is a bad idea, and instead, you should apply to part-time research fellowships like SPAR to write a research paper as part of the team there.
I will lay some background on part-time fellowships, then compare fellowships with writing a paper alone, and finally write a bit about useful solo research activities.
On a fellowship, you and several other fellows form a team mentored by a senior scientist, working together to produce research results, ideally a publication at a machine learning conference. Most fellowships are ~3 months long. Part-time unpaid online fellowships are less selective than paid full-time fellowships like MATS.
Part-time fellowships that I know about: SPAR, AI Safety Camp (AISC), AIgoverse AI safety, Apart fellowship, MARS. The default option is SPAR, it has a 20% acceptance rate when applying to 3-5 projects [1]. The other fellowships listed are more selective.
Outside view.
Here is a list of all papers accepted to NeurIPS 2025. Take a look at the “Authors” column. Note that single-author and two-author papers are very rare. There are some three-author papers, and the majority of papers have 4+ authors. It turns out that writing a machine learning research paper is a team sport, not a solo sport.
Workload estimation.
Here is my estimation of the workload for a typical NeurIPS paper with 4 authors. One of the authors is the senior scientist, who devotes 2-5 hours per week, and the other three authors work on the paper at least part-time, that is 15-20 hours per week, for ~4 months. Providing the same amount of work as a single author is equivalent to ~6 months of full-time work, which is a lot. On a fellowship, you contribute ~15 hours a week for 3-5 months (less if the project is less successful, and more if it is successful and worth pushing to publication).
What research problem to work on?
The problem should be simultaneously important/impactful (or at least interesting to other researchers) and solvable with at least medium probability. How to find such a problem? On the fellowship, you have a mentor who is a senior scientist with good research taste, and the mentor can point to such a problem. Working alone, you risk either working on a non-important problem or getting stuck with a problem that is too hard to solve. The ability to find a good research problem is a skill, and a junior researcher is not expected to have it.
Feedback speed.
If you work alone, you rarely get feedback on your work. Usually, you get feedback only when you try to publish or show the draft of the paper to a friend. On a fellowship you meet with a mentor at least once per week and communicate with teammates on a daily basis, so you get feedback on your work very fast. In particular, compared to working alone, the mentor and teammates can point out when you are reinventing the wheel or trying something that definitely won’t work, saving you weeks of effort. Wide skill set.
Producing a research paper can require many skills – engineering, theoretical foundations, domain knowledge, experimental design and evaluation, academic writing. Probably you lack some of these skills, so you need teammates with complementary expertise. Another advantage of teamwork is that you can learn from your teammate’s expertise.
Research skill.
There is a general research skill – what to focus on, what uncertainties to resolve first, when to run experiments and which ones, how to interpret the results, how to notice contradictions and missing parts. This is tacit knowledge that you can acquire during solo work, but it is much faster and more effectively acquired from a senior scientist who would mentor you on a fellowship.
Motivation, obligations, support.
When you work alone, you have only internal motivation to continue the project. If you get stuck somewhere, it is easy to get depressed and abandon the project.
At a fellowship, you have weekly team calls and clear commitments on what everyone will do until the next call, so you have some obligations. Essentially, everyone has external motivation from other team members. If you get stuck, you can immediately ask for advice from a mentor or teammates, and some of them can work with you on the thing you are stuck on. Also, fellows and the mentor have some formal or informal obligations to fellowship organizers to provide some results in a reasonable timeframe, which also adds motivation.
I hope the previous section makes the case that trying to write your first research paper alone is probably not the best idea. Still, other than fellowships, what can you do to grow as a researcher?
Maybe instead of a fellowship, I should rather write to a senior scientist directly, so I will do a project under their supervision?
I think on average it is still worse than a fellowship. The project will be mostly on you, with 1-3 hours per week of attention from the mentor. So you will still get the disadvantages of working without a team: more workload, potentially some needed skills missing, less motivation, fewer obligations, no clear deadline.
**Can I just not think about publications and career capital and focus on contributing to AI safety research directly? **
I think long-term it will result in less impact. To produce high-quality work, you need to work with smart, skilled, experienced people. It is hard to get a place in a team or organization where such people work. So it is better to focus not only on the value of your current research, but also on building career capital in the form of publications.
**What if I feel like I'm not ready for the fellowship I was accepted to? **
I have met people who got accepted to their first fellowship, but thought that maybe they were not skilled enough yet and seriously considered not going to the fellowship and upskilling alone instead. This is impostor syndrome. If you are accepted to a fellowship, this means that fellowship organizers think you are skilled enough, and if they are mistaken, it is their problem, not yours.
I have participated in [fellowship name] and have not liked it. Should I rather do research alone?
Probably a bad fit with the mentor or with the team, or the topic of research was not that interesting. Apply to another fellowship, or reapply to the same one with a different mentor.
https://sparai.org/advice/, see section "Apply to lots of projects": Our Fall 2025 acceptance rates by application count were: