I’ve been thinking a lot about how I learn recently, and more specifically, about how that has changed now that there’s a machine on my laptop that can do most of my work for me. I’ve had this idea floating around in my head for a while that the struggle of learning isn’t some tax we pay on the way to knowing something, and that the struggle is actually the part where the knowing happens. I’ve started calling it friction. I’m definitely not the first person to think about it this way (I’ll give credit where it’s due below), but I wanted to write down what it has looked like in my own life, and the heuristics I’ve been building around it, mostly so I stop forgetting them.

If there is no struggle, there is no progress. - Frederick Douglass, American abolitionist and writer

background#

My parents instilled one fundamental in me very early on: learning is an infinitely scalable activity with unlimited reward. The more you learn, the more you learn how to learn, and that turns into this self-fulfilling, cyclical process where every new thing you pick up makes the next thing a little easier to pick up. I’ve always loved learning because of that idea, but I don’t think I actually understood it until COVID-19.

When the pandemic hit, I was at a loss for ways to engage my time, and I let myself go a lot in video games. At one point my mom asked me who I wanted to be in this world, and I had to admit to myself that I really wanted to do great things - I felt like I owed it to the wonderful community I had, and to the world en masse, to make the best use of the resources I’d been given. So I started exploring. I tried directing stop motion movies first, and realized I loved the storytelling but hated physically building everything, so I moved to animation, which I enjoyed more but could never quite get to match the visions in my head. Eventually I found a TechWithTim video about machine learning, and around the same time, a Mark Rober video where he used machine learning to predict whether baseball runners would steal based on the coach’s hand signals, and I thought that was so unbelievably COOL. I was obsessed with the idea of building an oracle.

My first real project was a grades predictor: I wanted to see if an algorithm could predict my future grades based on the ones I’d already gotten, and then generalize that to a broader population. After that, I tried to predict the stock market, which was super fun precisely because people considered it hard at the time, and my naive self was like, I can try something here. None of it really worked out too well. But I was insatiable from that build, test, repeat cycle, and it meant so much to me because I could feel myself getting past the hard parts by doing the thing, rather than by finding a way around doing the thing.

That same energy carried into high school robotics, where I joined a Metal Crusaders team of two other students and a head mentor, with $0 in the bank and a closet to work out of (I wrote about this whole journey in proof of life, so I won’t rehash it here). Of course, there was no AI back then. If I wanted the robot to move, I had to understand why it wasn’t moving, and I learned everything I could from our head mentor and the internet to figure that out. I’m deeply grateful for that now, because every one of those hours is still with me.

the metaphor of friction as learning#

Think about what it means to push something. You don’t push hard on a box for the sake of pushing hard on it; you push hard on it because you want to move it somewhere, and you’re willing to deal with the friction as a means to an end of getting it there. My theory is that learning works the exact same way. Friction is the resistance between where your understanding currently is and where your goal needs it to be, and the work you put in against that resistance is what eventually turns into skill.

I think it’s important to note that the research world has been circling this idea for decades. Psychologists Robert and Elizabeth Bjork call it desirable difficulties: conditions that make practice feel harder and slower tend to produce learning that lasts much longer. My favorite example is a 2019 Harvard study where physics students in an active learning classroom learned more than students in a polished lecture, but felt like they had learned less (Deslauriers et al., PNAS). In other words, the friction registered to them as failure even while it was the thing doing the work. More recently, people have started using this exact word in the context of AI - Willem Van Lancker wrote about productive friction as the thing that builds judgment and taste, and Emily Zohar, Paul Bloom, and Michael Inzlicht published a piece literally titled Against Frictionless AI. They both say a lot of this better than I can, so please go read them.

why this is an epic metaphor#

What I want to add to all of this is the physics, because I think the metaphor keeps giving back the more you poke at it.

In physics, static friction is greater than kinetic friction, which means it takes a lot more force to get a box moving than it does to keep it moving. I think anyone who has tried to pick up a new skill knows exactly what that feels like: the first few weeks are always the hardest, and once you have some momentum, the same amount of effort carries you much further. And if that’s true, then I’d argue the hardest problem in learning is the very first push. For me, that first push came from two YouTube videos and a mom who asked the right question at the right time. I don’t think I overcame static friction because I was especially disciplined. I overcame it because someone showed me something cool enough that I wanted to move toward it, and I feel like that’s a huge part of how we can inspire people to learn more generally.

The other piece of physics I keep coming back to is what happens when there’s no friction at all. If you’ve ever tried to walk on ice, you know that a frictionless surface doesn’t make moving easier; it makes it nearly impossible, because there’s nothing to push off of. You can slide around for a while, but you can’t really steer, and you definitely can’t propel yourself anywhere you meant to go.

the frictionless path#

Now, in today’s day and age, AI has largely restructured the way we operate with respect to friction. For most tasks, it’s the path of least resistance - we want to get things done, and we now have an extremely smooth pathway to getting them done. I want to be clear that I don’t think that’s inherently bad, but I do think it puts a lot of us on the ice.

I learned this the hard way. In the spring of 2026, I worked on a project where I let AI take the lead, because I was really nervous about the thing being implemented and only had a high-level understanding of it myself. It felt fast at the time. But I definitely paid the price later on, because I didn’t understand where the project had evolved, and we didn’t end up submitting the paper to a conference I was really excited to submit to. I had slid a very long way without ever really knowing where I was.

The research is starting to show this pattern too. In a study of about 1,000 high school math students in Turkey, the students given plain GPT-4 access did much better on practice problems and then scored 17% worse on the exam than students who never had it at all, while a version of the tool that gave hints instead of answers mostly eliminated that harm (Bastani et al., PNAS 2025). Similarly, a study from Anthropic found that developers learning a new library with AI scored 50% on a follow-up quiz, compared to 67% for those who coded by hand, though the developers who used AI to ask conceptual questions (or to explain the code it had generated) held on to most of the learning. So in my opinion, the question isn’t really whether you use AI, it’s whether you’re still the one doing the pushing. And I feel like a lot of people right now are unfulfilled in what they do precisely because they aren’t - they feel like meat puppets, prompting AI over and over again without ever feeling the work move.

what can we do?#

I don’t have a complete answer here, and I’m actively figuring it out myself, but I’ll share the heuristics I’ve been building.

The big one, and the one I want to spend the most time on, is to choose your friction. In physics, there are ways to reduce friction besides changing the thing you’re pushing or how hard you push it; you can change what’s between the two surfaces (I don’t want to say lubricant, because that sounds kind of weird, but you get the idea). In real life, I think the thing that goes between the surfaces is meaning. Over COVID, I only worked on projects that meant a lot to me, and doing things on my own terms, when I wanted to do them, made the friction feel like something I was choosing rather than something I was being forced through. I feel like that’s lost a lot in modern education, where most of the friction we experience is assigned to us. So my advice is to pick the friction you actually want: the project you’re curious about, the problem you’d be thinking about in the shower anyway. Then let AI handle the friction you don’t care about (the boilerplate, the formatting, the setup), and protect the friction you do. Perspective plays a role here too. The reason people put insane hours into studying the night before a test is that the deadline suddenly makes it matter, and I think you can borrow some of that feeling, without all of the panic, just by deciding deliberately what matters to you.

The rest of my heuristics are smaller, but I’ll list them anyway:

  • I keep ideation human. When I’m in the ideation stage of a project, I deliberately stay away from AI. There’s a growing body of research showing that AI help makes individual ideas better while making everyone’s ideas more similar to each other (Doshi & Hauser, Science Advances 2024; Anderson et al., 2024), and I find that genuinely scary, because we still haven’t really unlocked what makes human idea production such a unique facet of our intelligence. My personal theory (which I’m not exactly sure holds up, since the evidence is pretty thin) is that it has something to do with language - linguistic relativity suggests that the way we use language shapes what we’re able to think, and if that’s even partly true, then outsourcing the words means outsourcing some of the thinking too. Either way, your ideas are precious, and you shouldn’t let AI steer you too heavily before you know what you actually want to say.
  • I stay conscious of what I make. I try to keep some sort of awareness of the code I write and the prompts I edit. If I can’t explain why a line exists, I haven’t really learned it, I’ve just received it.
  • I say it out loud. I’ve recently started using FreeFlow (shameless plug, link here) to just say what I’m thinking - it’s a nice free alternative to WisprFlow and Willow, which are also great options. Articulating my reasoning in full sentences forces me to actually have reasoning, and it gets around the fact that I’m lazy with prompt writing.

q.e.d.#

I’ll admit something: I’m currently working on seven different research projects, and I would not be able to work on all of them in parallel if not for AI. So I’m definitely not writing this from some place of purity, and AI has made a lot of things possible for me that I couldn’t have imagined even a few years ago.

The real question for me has become making sure that I take the time to be deep, be introspective, and slow down when I need to, simply so that I’m actually on top of everything I’m working on. And I think this is the part I most want to say to students, and to anyone who is worried about falling behind because of AI: it’s okay to go slow. Going slow is how you bring yourself up to speed, and the friction you feel when you’re learning something for real is a sign that you’re moving, even when it feels like the opposite. Once you actually understand the thing, that’s when AI lets you move as fast as possible, because you know where you’re going and you can tell when it’s taking you somewhere you didn’t mean to go.

It is this friction that I personally seek, and that I hope you’ll start choosing for yourself.

- Karthik