You can just not do things.

For the past two weeks, I’ve been working five hours a day, and I’ve never been more productive.

In January, I moved from Montreal to the Bay Area for the Astra Fellowship, an AI safety fellowship. The people I met were incredibly hardworking, and it was very easy to get sucked in. Midnight or 1am meetups were normal, and I’m grateful for those late-night conversations. I made a lot of friends, and it was fun to be surrounded by people who cared so much about the same problems. But people here are a little crazy. Sometimes it feels like a competition to see how many hours everyone can work.

There’s something almost intoxicating about it. Before moving, I think I’d assumed that people in the Bay Area would somehow be smarter than people everywhere else. What actually stood out was agency: the attitude summed up by “You can just do things.” This showed up in people’s work, but what surprised me most was how they made plans. In Montreal, organizing something could involve a lot of deliberation and indecision. In the Bay Area, someone would say, “We should do X,” and five minutes later, it would be booked. I loved this. Being around people who just did things made me want to do things too.

In AI safety, that energy comes with an extra layer of urgency. At times, it feels like a race to save the world. There’s always more to do, and it’s easy to feel like you should be doing it. What is consistently prized in these fellowships is experimental velocity. How quickly can you run experiments, get results, and move on to the next thing? I think what we should prize more is thinking. You can run a lot of experiments without stopping to ask whether they’re the right experiments, or what the results actually mean. And when you’re exhausted, it’s easy to keep doing things without thinking very carefully about them.

I’ve seen this many, many times, and I’ve done it myself. I was staying in the office until 1am, trying to “solve alignment.” I felt productive. How could I not be if I was working 14 hours a day? Then came the burnout. I couldn’t work for a week.

During that week, I read Rest by Alex Soojung-Kim Pang, and it really changed my perspective. I realized that you can be agentic without overworking yourself. Pang makes the case for around four or five hours of focused, creative work a day, with deliberate rest built around it. My takeaway was: do your deep work, get through whatever shallow work needs doing, and then actually let yourself stop. At first, I thought this sounded crazy. But I tried it. For the past two weeks, I’ve been stopping after five hours instead of stretching work across the entire day. I feel better, and I’m getting more done.

I’d read Deep Work a few years earlier, and it talks about similar things: ending the workday at a set time, having a shutdown ritual, and giving yourself time to recover. But Rest got through to me in a way Deep Work hadn’t. I think that’s because Deep Work is focused on work. It says it in the title. It does talk about rest, but for someone already obsessed with productivity, that part is easy to overlook. The message that stuck with me was: “This is how you do deep work. Now do this all the time.”

What I took from Rest, and what I’ve started to experience myself, is that rest helps in three ways:

• It helps prevent burnout, so you can keep showing up for the work.
• It allows you to think clearly, which is probably the most important thing you bring to research.
• It makes you happier. That matters on its own, and it also makes it easier to do good work.

Clear thinking is the part I want to emphasize most. A few hours when you’re rested and paying attention can be worth much more than a whole day spent pushing through mental fog. I think AI makes this even more relevant, and also harder to put into practice. In a recent LessWrong post about psychological support for AI safety researchers, Ihor Kendiukhov describes how increasingly capable agents can make time off feel more costly. Every hour you spend resting is an hour you could have spent getting an agent to do something.

I understand that feeling. AI can automate a lot of the boring stuff and make things that used to require serious cognitive effort feel almost easy. There’s always another experiment to run, another task to give an agent. When getting started is that easy, it’s tempting to just keep going. But your cognitive energy is still limited. You still need to decide what’s worth doing, evaluate the results, and notice when something doesn’t make sense. Those parts take attention, even when generating the output barely feels like work.

This brings me back to experimental velocity. As running experiments gets easier, choosing good experiments and understanding their results become even more important. We should use these tools during focused periods, when we have the energy to think clearly about what we’re doing. Then actually let ourselves rest. If all you’re doing at 1am is blindly saying “yes” to your agent, maybe figure out how to automate that part and go to bed. Save your attention for the decisions that actually need it.

I still feel the urgency of working on AI safety. But I’ve changed my mind about what that urgency asks of me. These are incredibly difficult problems. Staying in the office until my brain was mush, then burning out for a week, wasn’t helping me solve them. If we care this much about the work, we should care about giving ourselves the rest we need to think clearly.

The Bay Area taught me that you can just do things. But you can just not do things, too. You can close your laptop after five good hours, let your agents keep running, and go do something else. Choosing when to stop is also agency.

So, to everyone working on AI safety: please rest. It’ll make you happier AND more productive :)