Before going into the details, a disclaimer: I am not going to provide details about the work itself, the interview process, or any specifics of day-to-day events in this blog post. The main focus of the post is solely on what I’ve learned, some possible mistakes I’ve made while doing a research job, and the challenges I face juggling multiple projects.
To give a little background: I was accepted to a visiting summer fellowship doing research in automated interpretability, which I initially had almost no knowledge about. This was part of my intention to try out different subdomains in technical AI safety, possibly areas in which I have less or no knowledge. I applied for many fellowship programs over the past year for this reason. The main aim was because I had been working on nearly the same thing for the past few years and had been thinking about stepping out of my comfort zone, but not straying too far from my domain expertise. Luckily, AI safety is a highly popular research area, so there are “many” opportunities out there for me to apply. A three-month short-term fellowship is ideal because it’s long enough for me to learn something, work on some research ideas, and hopefully achieve some results. Also, it’s short enough that if I fail at this or eventually don’t like the work I’m doing, I can simply walk away and return to my comfort zone. Additionally, with my recent commitment to a PhD program, this short-term research experience allows me the flexibility to challenge myself without risking my longer-term academic plans. Let’s dive into the main content of this blog post!
1. Prepare yourself to learn
When I received the offer, I had about two weeks before onboarding and hadn’t been given any information about the actual work I’d be doing. All I knew was that my focus would be on automated interpretability, specifically, LLM mechanistic interpretability. At that time, I was too relieved and excited about finally securing an offer, after a long period filled with interviews and rejections. Despite not knowing what I would be working on, I decided to rest instead of doing a literature survey. I felt I deserved some downtime before starting my new journey. Technically, I wasn’t completely resting since I had another ongoing project with weekly meetings with my supervisors. But in hindsight, this was my first mistake. I’m not saying that taking a break was a bad decision; I had certainly worked hard to get to that point, but looking back, those two weeks could have been enough to learn some basics about the field. For example, I could have familiarized myself with: the main areas of interpretability research, the fundamentals of mechanistic interpretability, basic techniques for intervening in model internals (e.g., activation patching, attribution patching), common model internals studied (neurons, features, circuits, SAEs, cross-layer transcoders). Neel Nanda has compiled a list of mechanistic interpretability resources for his MATS streams, which I had opened but never started reading. If I could go back, I would definitely spend some time engaging with those resources during those two weeks. Additionally, since the onboarding period coincided with the NeurIPS rebuttal timeframe, I particularly regret not preparing beforehand. Managing my own rebuttal, brainstorming ideas, and trying not to appear uninformed within the new group left me with nearly no sleep during that initial period.
2. Juggling between multiple projects?
This remains an open question for me. During the fellowship, I was involved in three projects at once: the fellowship project, a paper resubmission, and an ongoing research project I’d been working on before the fellowship, plus some minor responsibilities. By the end, I realized that only the resubmitted paper had reached a state I was happy with, while progress on the other two was unsatisfying. I experimented with two different strategies for managing my workload. Since the fellowship project was my main focus and the other two were side projects, I first used a “day-and-night” approach: working on the main project in the mornings and side projects at night. This worked for a while, especially since rewriting a familiar paper wasn’t too mentally taxing, and switching focus was manageable. However, with my other side project, one requiring new ideas and method design, I realized that switching contexts left me with too little progress each week. Through regualar weekly meetings, I discovered my output and contributions to discussions became less meaningful. To improve, I switched to blocking longer stretches: dedicating Monday, Wednesday, Friday, and Tuesday/Thursday/Sunday mornings to the main project, and Tuesday/Thursday after lunch and Saturdays to my side projects. I thought longer chunks would help, but it didn’t boost my progress as hoped. I think this is because before trying out this strategy, I have been doing constant context switching for a long time, my mental elastic band feels worn out already. Instead, I started to feel mentally exhausted; the cost of constant context switching became very real to me. If all projects need deep focus, then context switching becomes a major productivity drain. To use a computer analogy, every switch means unloading one project from my brain (GPU) to general memory (CPU), then loading the next into focus (GPU). This overhead really drags down productivity.
Toward the end of the fellowship, conversations with friends like Aaron and Matthew was particularly insightful. They suggested focusing on one project at a time. Matthew remarked, “I’m not surprised things aren’t going so well, but I’m surprised you’re not burned out.” I recognized the signs in myself -> my progress had slowed to a crawl, and my brain felt sluggish, like it was running at half speed. Reflecting on this, I realize that being able to work on just one project at a time is a privilege that many people, including myself and my friends, rarely have. The real challenge is learning to commit only to projects that truly excite you, and building sustainable routines that move you toward meaningful goals. Working seven days a week and constantly optimizing for productivity eventually backfire and lead nowhere is a sign of overcommitting. Yes, that’s me… For me, giving myself larger blocks of uninterrupted time for each project and minimizing context switching is the first step toward improving how I juggle multiple responsibilities. This is something I expect to keep working on and refining over the next few years.
3. Plan your days in advance
I used to plan my days only one day at a time, which probably contributed to my overcommitting: I was so focused on short-term tasks that I lost sight of the bigger picture. Early in my time at Oxford, I sat beside Matthew, who always brought two notebooks: one for notes and one as a planner. Every Monday morning, he’d review his calendar and planner, then time-block his responsibilities across the whole week. Seeing that made me notice how chaotic my own days were, and inspired me to make a change. When I returned to Taiwan, I bought a paper planner. I dislike digital planners because they’re too easy to adjust, you can just drag things around and delay work without really thinking about your time. With a paper planner, every week (usually on Sunday), I plan my “ideal week”, dividing each day into three blocks: morning, afternoon, and night. For each, I write a key focus area (sometimes adding detailed tasks). Then, before bed each night, I add 2–3 actionable tasks for each timeblock next day. Throughout the day, I check off or cross out tasks as I complete them, and if something feels too big or out of order, I break it down or reorder it in the planner. If I miss a task or need to postpone it, I have to write that down, making me much more aware of my real capacity and intentions. This approach has a few major benefits. First, it helps me be deliberate about what I say “yes” to. If a project can’t ever seem to fit into my weekly plan, it probably isn’t exciting or important enough, and I should consider letting it go. Similarly, if I’m offered something new but can’t find space for it in my planner, I now know it’s better to turn it down. Another advantage of a physical planner is that it’s easy to look back and see what I’ve accomplished and flipping through old pages gives me comfort when I need motivation. After a few weeks of using a paper planner, I’ve started to see more order in my days and have begun to celebrate small wins. I hope this system continues to help me bring more clarity and balance into my life.
4. Reserve time for learning and reading papers
For the past few months, I’ve been constantly busy, jumping from one task to another with barely any pause. Recently, I realized I couldn’t remember the last time I genuinely spent time learning or reading something new. It was clear when I saw the 20–30 open browser tabs full of articles and papers I meant to read but never did. Whenever my browser slowed down, I would just randomly delete tabs (yes, I know it’s terrible). I forgot about the privilege of being a student (soon to be one again lol), which is having the freedom to focus on learning and curiosity, rather than always having to meet deadlines or performance targets. After returning to Taiwan, I organized all those open tabs (papers, textbooks, and online courses) by topic into a proper reading list. Now, I dedicate half an hour each day to going through emails and slack messages with new paper recommendations and add anything interesting directly to my list, instead of letting tabs pile up. Kudos to Matthew’s reading list system! I’m also aiming to set aside at least three afternoons a week for focused reading and learning. It’s a challenge, especially since I’m a slow reader and tend to get distracted thinking about my experiments, but I’m working on making this a sustainable habit.
5. Rest is as important as work
I used to envy people with lots of projects, thinking that meant you’d never get bored, you could just switch tasks when stuck. Before this fellowship, I had just one exciting project. When I hit a wall, I had to push through. I always thought that with multiple projects, I’d never waste time and wouldn’t even need rest. Like a superman, but not saving the world though! But during this fellowship, juggling several projects, I realized: what if I get stuck on all of them? Instead of feeling productive, I just felt the pressure to keep up consistent progress everywhere. Switching focus isn’t as simple as it sounds, it’s not like swapping games as a kid. As a former serious runner, I loved sticking to my workouts. But with my new workload, running became just another box to check. I rushed warmups, skipped recovery, and my sleep suffered. It all caught up with me after one intense Saturday 3x3 (3x3000m with 5:15/ 5/ 4:45 min/km) morning workout: I crushed the run, got coffee at Paper Boat Café, and was back at my desk by 10am… but my sore legs got worse, and old injuries flared up. That pain haunted my last few weeks in Oxford; I barely slept and had to stop running entirely. After coming home, I spent weeks and a lot of money on PT just trying to recover, barely able to walk at first! Honestly, a month without training left me feeling like the unfittest person in the world, but at least I still get back on track/ road. I learned this lesson the hard way: as a long-distance runner, a month without training means getting back to square one, and I definitely have a long way to go before regaining my original fitness and competing again. Still, I’m back to steady training, listening to MLST, MAD and Dwarkesh podcast (please don’t criticize my taste) during warmups and recovery now makes even rest time feel rewarding. Through all this, my little Sunday tradition remained: walking to Blackwell’s, wondering around for a while and picking a book, reading with coffee at Caffè Nero upstairs. I even bought seven books in three months and hauled them all back to Taiwan in my carry-on (quite an effort lol). Looking back, I know rest is crucial not just for enjoyment but for real productivity. I regret turning what I loved into a chore, and now I’m determined to honor rest as much as work. Sunday afternoons are still reserved for reading books for pure enjoyment.
6. You always have a choice
I consider myself a very lucky person, especially given how much freedom I have each workday to decide what to work on or how to use my time (meetings aside). Some jobs require you to be in a certain place, doing specific tasks, but I don’t have those kinds of obligations. Ironically, though, I often waste this gift of freedom by living in the most unfree way possible: I dump endless to-dos onto my days and feel like every waking hour is just about checking boxes. Life becomes a dull, intense blur, and suddenly I realize I spent 17 hours flying to a foreign country for an exciting research fellowship, only to come home telling people I feel like I barely experienced anything. For a long time, I told myself I “had” to finish this and complete that, always saying yes to each demand. But in reality, I always have a choice. I don’t have to join every meeting (especially the pointless ones), and I don’t need to respond instantly to every Slack or email thread (no matter how many inbox zeros I chase, they never last). Most importantly, I don’t have to work through every day nonstop. By mindlessly saying yes, I piled up obligations until I squeezed the freedom out of my own days (even though, technically, no one was forcing me). Reading 4000 Weeks (bought from Blackwell!) and reflecting on all my juggling and struggling, I learned that saying yes to one thing is actually saying no to countless others that might bring me joy. Without making intentional choices, I end up trying to maximize excitement on the surface, but I’m really just diluting my time, energy, and enjoyment. Now I see that learning to say “no” and saving my energy for what truly matters is the real skill. It’s definitely tough, but after all this, I know it’s one I have to master.
Some personal wins
I’ve been talking about things I think I should improve or master, but actually, there are also things I’ve managed to fix and gain throughout this fellowship. First, I think becoming a “morning” person is a skill worth celebrating. My daily routine was pretty messed up before this fellowship. Since I didn’t have strict working hours, my sleeping schedule got… let’s just say, I could blame those conference deadlines. I’d usually go to bed at 2–4 am and wake up somewhere between 9–11 am (so in a way, I was a morning person, just sleeping and waking in the morning!). But after arriving in Oxford, transitioning from an unhealthy, unstructured sleep routine became surprisingly natural. I’d finish work around 8–10 pm, then go straight home and sleep, mainly because there wasn’t much to do in that small, uncomfortable room. I started waking up around 5-6 am and heading to work, and I’ve actually kept this routine going even now that I’m back home.
Second, I joined the OAISI roundtable, which I loved so much I’m considering organizing something similar at my new job. In short, the OAISI roundtable brings together a small group of students interested in technical AI safety for discussion once a week: chatting about new developments in the field, with one person sharing what they’re working on or diving deeper into their area of research. The rest chime in with questions or thoughts. As an introvert, just introducing myself is a hurdle, I often need to memorize my self-introduction. Since I have low social energy and tend to stutter (my supervisors definitely know this) while I speak, talking itself can feel exhausting. But the OAISI roundtable made conversation feel much easier: I could just sign up and show up, and people there were so welcoming. They didn’t mind if I spoke slowly or a bit unstructured. It’ll be a great challenge to try creating such an environment at my new job which I’ll be responsible for building a space where people like me feel welcome to engage in dicussions. But, life should be about occasionally stepping out of our comfort zones to keep things interesting! Before ending I would like to share and excerpt from the book 4000 Weeks, which I just finished reading as a summarization of this post:
The average human lifespan is absurdly, terrifying, insultingly short. But that isn’t a reason for unremitting despair, or for living in an anxiety-fuelled panic about making the most of your limited time. It’s cause for relief. You get to give up on something that was always impossible - the quest to become the optimised, infinitely capable, emotionally invincible, fully independent person you’re officially supposed to be. Then you get to roll up your sleeves and start work on what’s gloriously possible instead.
I’m really thankful for this fellowship and all the tough lessons I’ve learned. From a results perspective, the ending feels a bit disappointing or like a failure to me, but I guess we learn more from failures than from successes, right? Thanks to everyone (Matthew Farrugia-Roberts, Aaron Maiwald, Elias Kempf, James Oldfield, Louis Thomson, Jess Rapson, Adi Simhi, and Tung-Yu Wu) who had meaningful discussions with me and inspired me to write this post.