Sucking at something is the first step towards being sorta good at something.

– Jake the Dog, Adventure Time, Season 1, Episode 5, "His Hero"

As a child, I spent a lot of time playing The Sims. Being a grown-up intrigued me, and The Sims was a great way to experience adulthood by proxy. I couldn’t wait to become a rational agent and make my own decisions. And what’s more: the The Sims let me do it for other people. In the beginning, I was fully reliant on cheat codes – I’m sure move_objects on is ingrained deep in my muscle memory, for instance.1 The point of the game became to buy the most expensive objects for my Sims’ houses, and to have them spend all their time scoring relationship points with other Sims. After a while this became tedious, though, and I wonder if I would have enjoyed the game at all had I had a cheat code that instantly maxed out Sims’ relationship points.

At some point in early adolescence I discovered that it would be more fun to play the game without cheat codes; that it was more rewarding to have my Sims take on a poorly paid job as a test subject or mailroom clerk and to watch them progress through a career while simultaneously making friends, expanding their families, and honing their abilities. (And if you sneer at this as an example of a game teaching children to be docile workers, I invite you to consider the life lesson that’s to be learned from hammering out rosebud on a keyboard several times so that your Sim can afford the top-of-the-bill §6,500 Meet Marco desktop computer.)

Lately, I’ve been thinking about things that require some effort. These thoughts are in no small part inspired by the barrage of generative-AI marketing material which proffers “a helpful friend with PhD-level intelligence.” What does that mean, though? Back when I decided to pursue a PhD – working below my earning potential on an ill-defined project for several years – I noticed people rolling their eyes at this and making it known they in fact saw it as an unintelligent choice.

A good graduate student necessarily has a deep knowledge of the field in which they’re working, but so does any dentist, bricklayer, or software engineer. If anything, the best graduate students are those that have a sense of which problems are worth persevering through. Identifying such problems takes something that goes beyond ‘expert-level knowledge,’ it requires a sense of where a field is going, and above all, a sense of what problems suit your background and interests.

Submitting a paper or handing in a dissertation are important milestones of a PhD, but in the end these are perhaps not its most important goals. The actual point of doing a PhD is far less tangible: it’s the period leading up to the dissertation, the years spent working on untangling wicked problems with uncertain outcomes while devising creative workarounds for malfunctioning lab equipment. The process of writing articles or the dissertation is important, too; this is where you comb through the scientific literature again, organize your thoughts, and hone them while committing them to paper.

Most dissertations will vanish in university repositories, to be consulted only occasionally, and even then primarily by the handful of people that was involved in the research that brought them about in the first place. You could argue, then, that it is fine if students use generative AI tools to write their theses, since they typically do not have a wide readership. Why bother wrestling with written text if you can wrestle with lab equipment instead? I would argue the opposite: if you deprive (junior) researchers of the writing process, you deny them the chance to deeply reflect on the work that they’ve done, and it is very hard to develop autonomy as a researcher without doing so.

People advocating for generative AI seem to be primarily focused on the short-term rewards it brings them without considering too much whether there will be any costs to be incurred in the long term for them personally. When Terence Tao said that “AI tools are like taking a helicopter to drop you off at the destination of a distant hike,” this was received as a deep insight by the blabbersphere, but it should be fairly obvious to anyone in the business of solving problems that the process leading up to a problem’s solution can be just as valuable as the solution itself, especially for your personal development in whatever domain you do your problem solving.2

When I was a graduate student I came to enjoy indoor climbing. The university had a small bouldering gym right next to our lab with a decent rotation of routes, and while I had found out through empirical research that my dimensions put me at a disadvantage for “recreational” activities such as water polo, it turned out that they were favorable for climbing up a wall. Eventually I went three climbing times a week and could climb some of the more challenging grades. Being sorta good became a major source of pride during a period of time when I spent most of my days in a dark lab figuring out why stuff didn’t work.

The goal of climbing is to make it up the mountain, but that’s not the point of climbing. If it were, you might as well ascend by taking a cable car or, indeed, a helicopter; indoor climbing competitions would be won by those who were clever enough to take the stairs.3 The point of (recreational) climbing is to ascend in a way that’s challenging and enjoyable for you. I didn’t fully realize this at the time, though, and at some point my personal priorities shifted. As a result, I wasn’t able to climb as often anymore, I lost some strength, stopped making progress, and my enjoyment went out the window. I stopped climbing altogether after some time, which I regret in retrospect.

I doubt whether any of this would ever have dawned on me had I not picked up playing piano again two years ago. As a child, I played piano for several years, but because of the complete and utter lack of cheat codes available for the instrument the point of it was lost on me. After having looked at phase vocoders a while ago, I wanted to learn a bit more about music, and playing an instrument seemed like a good way to do so. I bought a rinky-dink MIDI keyboard which I hooked up to GarageBand to practice, and started taking lessons. Unexpectedly, this was a lot of fun. I was surprised to learn I didn’t have to start from scratch, in spite of my juvenile dedication to practicing as little as possible, and I quickly retired the MIDI keyboard in favor of a digital piano.

During those first couple of months, I tended to extrapolate my trajectory to visualize where this newfound hobby was going to take me. I imagined that, upon winning the lottery, I would dedicate all my time and winnings to practicing piano and try and enroll in music school. Needless to say, given the limited resources available to me, my progress slowed down and I had to ask myself whether this would be a fun activity if ‘sorta good’ was the best I’d ever be.

Why make music if there will always be people who are much better at it, especially if you can listen to their recordings anytime, anywhere? The answer isn’t very difficult. It’s a lot of fun to work your way through a piece, and to notice improvement over the course of a couple of days or weeks. Making music yourself also gives you a new appreciation when listening to other artists, and it can be a lot of fun to play with others. Every amateur musician knows the value of making music yourself, as do most concertgoers. After all, player pianos (self-playing devices that mechanically rattle away at a composition) are found in museums, not in concert halls. Who would buy a ticket for a mechanical organ concert?4 Which audience appreciates it when they find out the singer is lip-syncing their way through a concert?

The insight that working through complicated things is rewarding and worthwhile, is an important lesson in life, but it seems largely ignored in the recent breathless coverage of the progress of generative AI. A collateral effect is that this tone has led to a new permissiveness in our professional lives. I’ve experienced colleagues answering my questions with text that seemed suspiciously AI-generated (“Sure, here is…!”), which misses the point of asking each other questions: answering someone else’s question is also an opportunity to hone your own understanding of the matter at hand. One of the standard teaching techniques in medical school is “See one, do one, teach one,” implying that you’ve only truly mastered a procedure if you can help someone else do so, too.

In my experience, most physicists don’t enjoy writing, and don’t want to spend a lot of time wordsmithing away at sentences until they sound just right. On its own, this is not a problem. Richard Feynman, for instance, produced a great number of books, but didn’t ‘write’ many of them in the traditional sense. Most of his works were assembled by teams of editors from interviews and lectures, which was possible because Feynman was a fantastic storyteller. He must have devoted a significant number of mental clock cycles to how he was going to package his material for a particular audience, which can be a challenge on its own (his introduction to the principle of least action is a masterclass in this). Working through these challenges leaves you with a better understanding of yourself, your audience, and the stuff that you’re presenting. If you ask a chatbot to do this work for you, you’re ultimately cheating yourself out of a better understanding of what it is you do.

A few months ago, I went to see a physiotherapist to get help with some persistent aches. He recommended some exercises, but was pretty adamant that I start bouldering again. I took him up on his recommendation, but I try to leave the expectation of ever getting sorta good at it at home. It’s been a lot of fun so far, and I look forward to it every week.

It beats me why someone would rosebud their way through a physics problem or the writing of a blog post, thereby depriving themselves of the opportunity to mull things over, get to know themselves a bit better, and perhaps arrive at an insight that hasn’t occurred to anyone yet. By writing some of this stuff down, I get to work through my own thoughts and perhaps build new ideas on top of them. I have a backlog of questions that I think are interesting to look into and write about. It would be trivial to instruct an LLM to write a blog post on either of these questions and to dump the result here; this may result in something that’s intelligible, but it would bring me none of the satisfaction of working through these problems while searching for connections to my own background and experience.

Footnotes

  1. After I graduated from the Sims franchise, I went on to play a non-negligible amount of Grand Theft Auto, a game which coincidentally also has a lot of cheat codes to circumvent its internal rules. Since the entire premise of GTA is that it’s a simulator that allows you to play fast and loose with the rules of society this is somewhat ironic. ↩

  2. Granted: the echo chamber has been catching up with this via the “If we stop hiring junior devs, where will senior devs come from?”-meme. The natural solution to this, though, is that curricula tend to adapt themselves to the times. Look through universities’ brochures, and you’ll find all sorts of programs that didn’t exist fifty years ago. Visit a hospital, and you’ll have no hard time finding specializations that were developed in the past couple of decades. ↩

  3. My thoughts frequently drift to this Alan Watts bit on life and music (animated by Matt Stone and Trey Parker here): “In music, one doesn’t make the end of the composition the point of the composition. If that were so, the best conductors would be those who played fastest, and there would be composers who wrote only finales. People would go to concerts just to hear one crashing chord, because that’s the end!” ↩

  4. Full disclosure: I have done so, but only for the Muziekgebouw’s Fokker organ, which was designed by Adriaan Fokker (of Fokker–Planck equation fame) and which has 31 notes per octave as opposed to the usual twelve. It can be played through a MIDI interface, because owing to its many keys it can be difficult for a human to play on it. ↩