An Exploration and Exploitation Game

Over the last couple of months, some high schoolers and undergrads have asked me what to focus on and what to learn as they head into their careers. I'm not really in a position to recommend anything, because everyone is different and has different goals. And I'm figuring things out myself. Here's some of my thinking, as of July 2026.

On what to learn

The first is learning how to learn. In high school I mostly just memorized things, and that was enough for good grades. A pretty limited objective. During the undergrad I fortunately had to actually understand things, really understand them. Fundamentals. In my case this was for computer science and math. More generally, being able to deconstruct a situation, understand it, and learn from it. This feels pretty universal to me. As long as humans need to improve on what they're doing, I'd guess that skill holds no matter how the technology changes. And it's fun.

The second is going deep. If you study computer science (or any other subject) and really study it, you go into the mathematics, how systems work at the operating system level, how bits work at the computer architecture level, and so on. Then when you move to something adjacent like electrical engineering, you don't start from zero. A lot of the fundamentals transfer. And if you already know how to learn, picking up the new knowledge tends to go quickly.

And in conjunction with learning and going deep, I think it's useful to find your own limit. Because without action, being able to learn and going deep isn't enough. They're necessary but not sufficient. So how much meaningful work can you actually do? Not just for a good week, but constantly, over months and years. It's easy to keep going when things go well. But what if they don't? Where do you top out, and not only in hours but intellectually? I don't think you can answer that up front. You mostly find it by pushing and seeing what happens. And once you have that data on yourself, you're equipped to take on and execute the interesting, meaningful problems.

For example, I think someone who can do all three tends to adapt wherever technology goes. I remember a talk by Jamie Dimon saying that the era where you learn one thing and do it for life is probably over. I've only been alive 25 years, so take this lightly, but even before GPT-2, the people who took the one-thing-for-life approach seemed to me to have a fairly limited mindset. The people who kept learning new things and kept trying to improve seemed to always find something meaningful to do. And I think the learning gets much easier, almost natural, when you actually care about the thing. So if I were an undergrad student now, I wouldn't just blaze through classes for good grades. I'd pick something I'm genuinely interested in and challenged by, something that really kicks my ass, and go deep. When you like what you're doing, staying curious and getting better stops feeling like effort.

Learn how to learn.
Go deep.
Find your own limit.
Continue.

On collecting datapoints

Exploration. Throughout undergrad I worked part-time or full-time, alongside my studies. Part of that was to pay rent, but mainly to look into different industries and subjects and figure out what I actually wanted. I went from cybersecurity research, to data science at an agriculture company, to software engineering at some bigger tech companies, and only at the end did I realize I didn't want to focus mainly on pure software engineering at all. I wanted to do research instead. The whole point, at least in hindsight, was to collect as many datapoints as I could and then make a more informed decision. How should you know what you like and what you're good at if you've done one thing your entire life? I think it makes sense, especially for a young person, to find their limit. Go hard.

Exploitation. A bit orthogonal to the exploration side, I think it's just as important to stick with something even when it doesn't pay off right away, and to go really deep into it. When I was researching at the Sky Lab, going deep into semantic caching was a genuinely satisfying experience, because I eventually understood the entire system (and many related subjects). Where it breaks. How to improve it. Additionally, on a meta level, that kind of depth reveals interesting learnings. What is your structured approach for understanding a completely new system (and this doesn't have to be a computer science system)? Where do you start? Which tools do you reach for first? Me handing you answers wouldn't cut it, partly because I'm still figuring them out myself. And I don't think reading them would be enough either. You have to actually experience it.

Exploration.
Exploitation.

On intrinsics

Over the last few years I've seen people with very different motivations. Some seem to be chasing goals they can't really explain. Ask them why they're pursuing it and they don't quite know.

Next, invert that. When I talk to someone who intrinsically enjoys what they do, driven by a constant desire to get better at it, I get the sense those people live fuller lives. Some external factors matter too, no question. But I've watched someone do a PhD because their parents asked them to, pursue it for six years, and still not know what to do afterward. Including all the prep for it, that's maybe 26 to 30 years spent on something they don't fully believe in. So I think it's worth taking seriously the idea of doing something you actually care about, even if it's hard to know what that is early on. And no one can tell you the answer. Eventually, figuring out the answer is a fulfilling journey. This goes back to exploration and exploitation.

Do what you intrinsically care about.

On geographics

I think that your geographic location is one of the higher-leverage decisions you make, because where you are largely determines who's around you and what you have access to.

The first reason is the people. There's this saying that you're the average of the five people you spend the most time with. I think it's true. But you can only draw those five from whoever is around you, and that pool is set by where you live. So place is upstream of the people, and the people are upstream of a lot. Ideally you end up somewhere dense with people who challenge you and make you better, where you do the same for them. Give and take, not only take. For example, if the people around you optimize for the easiest path through school and that doesn't match what you're after, it's worth asking an uncomfortable question. Should you change your pool? Do you have to move to a new city to do so?

The second reason is proximity to capital. If you have good ideas, in most cases you need capital to execute them. Sometimes a lot, sometimes a little. Some places concentrate the money and the skilled, ambitious people who deploy it, and being inside that environment helps. It isn't required. But it helps.

For tech specifically, a non-exhaustive list would be San Francisco, New York, London, Munich, Zurich, Beijing, Singapore. I don't have enough visibility into other domains to make a useful suggestion there.

Location decides access.
Optimize for your five people.
Proximity to capital helps.

On technology and problems

Technology seems to move in cycles. Call the internet n-1, the smartphone n, and the wide adoption of AI n+1. Each cycle builds on inventions of the previous ones, and I think that's what lets us solve more and more difficult, interesting, and impactful problems over time.

And each cycle also shifts which problems are worth solving. The old problems get absorbed by the new technology, and attention moves to harder ones. So the useful question probably isn't whether your role survives a shift. It's what problems the shift opens up.

I learned this backwards. While the LLM hype kept growing I was at Microsoft, still deep in the exploitation of becoming a software engineer, and I asked one of our directors how much this hype would affect the job market for engineers. I asked because I was uncertain whether my "role" would even exist. Looking back, I think that was the wrong question and a fairly limited way of thinking. It assumed a role is a fixed thing rather than a response to a problem. But what do you need an engineer, or any role, for? To solve a problem. And when that problem gets solved, often because of some related technological advance, you just move on to a new problem people care about. Two years ago I couldn't have foreseen going into machine learning models for physics simulations at all. The cycle opened a problem I didn't know existed. And it's a pretty cool problem.

Knowing how to learn, being able to dive into a space, and operating at your limit are what make it exciting to switch domains and take on new problems. When the next cycle opens something you couldn't have predicted, those are the things that let you move toward it instead of away from it.

Roles exist to solve problems, and each cycle just seems to let us work on harder ones.

I tend to think of life as an exploration and exploitation game, ideally an adventurous one.