Context [optional]: 

I’m a Ph.D student at MIT doing research in technical AI safety. I mostly do work related to interpretability, adversaries, and robust reinforcement learning (see more here). Often, I am referred either by some Effective Altruist friend of mine or 80k to talk with someone who is interested in technical AI safety research – usually an undergrad (feel free to email me if you’d like to talk – scasper@mit.edu). One piece of advice that I give to almost everyone is to do a “deep dive.” 

What’s a “Deep Dive?”

Sometimes there is a catch-22 when it comes to pursuing research work – you need opportunities to gain experience, and you need experience to get opportunities. A deep dive is one way I recommend to gain experience and demonstrate initiative on one’s own. It’s also a good way to explore an area if you’re not sure that you’re interested in it or not. Having done several of them in the past, I believe that they are the best academic experiences I’ve had aside from research projects themselves. 

More concretely, a “deep dive” is a procedure that I recommend for reading papers that I think allows you to learn a lot about an area of research that you are interested in. There are 5 steps.

1. Get a broad level of familiarity with active research in the field that you’re interested in. This could include looking through high-level survey papers, survey posts, reading lists, etc. For AI safety, there is no shortage of resources like this. Textbooks usually won’t be so useful for this because they typically cover established concepts in a field rather than active areas of work. 
2. Pick a fairly niche subfield of active research that you want to learn a lot about. The exact scope should be up to you. One possibility could be to let the citations of a survey paper define your subfield. Another could be to focus on the research area that a particular professor or organization you’re interested in works in. Getting advice on subfields from someone experienced could be useful at this stage.
3. Read a ton of papers from the subfield. Ideally, you’d be able to read 50 or more. This will probably take over a month. I think that a pace of approximately one paper per day is good. When looking for papers, the “cited by” button on Google Scholar will be your friend.
4. Take notes on all of the papers as you read them. The goal is to take a set of notes that quickly summarizes the key insights related to the paper’s methods and contributions. 
5. Do a capstone project. This could be a blog post, a research proposal, a survey paper, a usable github repository, etc. One of the good things about a deep dive is that it can be done anytime anywhere by anyone without mentorship. But at this step, reaching out to someone in the field to get their help or feedback could be useful – whether it’s about initial advice on what the capstone should be on or getting feedback on the final result. 

Why is each step key?

Simply reading a lot of papers is good. It will make you better/faster at reading them and teach you a lot. But there are some very specific reasons I recommend all 5 steps. 

1. The first step is to get background. In addition to the general-usefulness of forming a mental map of the field, this will help in selecting a good subfield for the deep dive. 
2. In this case, depth > breadth. Selecting a niche subfield to familiarize yourself will give you an understanding comparable to people doing the research. It’s also the best way I know of to gain ideas for things to work on in the future. 
3. Not-so coincidentally, ~50 is comparable to the number of papers that many papers cite. 
4. This will help you with memory, the next step, and future reference. I recommend taking notes in a centralized, searchable system whenever you read a paper you care about – not just in a deep dive. I use Notion for my notes, and it’s good, but anything similar will do. 
5. This will help you (a) revisit, apply, and remember what you’ve learned, (b) get a feel for doing work in the areas, (c) give you a shareable item to keep in your portfolio and that you can sometimes use when looking for future research opportunities. If you want to go above and beyond, you could write a survey paper and put it on arXiv, but if you go this route, I’d recommend taking a lot of time and getting outside advice.

What do I do now?

I’ve done several deep-dives in the past, but lately, I just try to read papers in general now that I am more familiar with the work that I do. Whenever possible, I read at least a paper per day that I wouldn’t have read otherwise as an exploration heuristic. I still take notes on them. More recently, thanks to advice from Dan Hendrycks, I bookmarked, ​​https://arxiv.org/list/cs.AI/recent and started checking through the new titles every weekday. The deep learning Twitter space can also be useful for finding papers.