Where, then, shall I be brought?
—Kafka
I’m a big fan of George Saunders’s Story Club, which hosts discussions about stories and about writing in general. It’s got an incredibly cozy, wholesome vibe. In his “Office Hours” posts, Mr. Saunders teaches an approach to writing that is very reassuring—relaxing, even. The name “Office Hours” is fitting, because it really does feel like I’ve walked into my favorite professor’s office, distraught and discouraged, and he is taking the time to give me the exact guidance I need to feel capable again. You don’t have to take my word for it, check out the comments section on any of the Office Hours posts, and be uplifted yourself.
George Saunders’s approach emphasizes editing: playing around with your writing, jostling it around to get it to “work,” and developing your “ear”. He also emphasizes being honest: trying to figure out how the characters would actually behave, and pursuing the underlying “truth” about the story.1
This approach takes a lot of the pressure out of writing. He shows that it’s not necessary to just be this amazing artist that pulls magic out of thin air; instead, you can be more like someone patiently whittling a pretty little figure out of wood—as long as you are being yourself. As long as you are saying something True. I would attribute a lot of whatever confidence I have writing to him, and for that I am eternally grateful. So, thank you, Mr. Saunders, you changed my life.
Anyway, my continual ingestion and digestion of Story Club has led to me to a soft belief that writing is inherently stochastic.
What do I mean by this? Let’s see. What I mean is that no one really knows how to write something good.2 Instead, all you can do is start somewhere basically arbitrary, and then you just perturb it, and perturb it, and you listen to it, and then you have to either accept it or reject it, ruthlessly, and then you perturb it some more, and it’s moving around sort of randomly, but hopefully it’s going somewhere, because you have this dark sense of an ineffable, pre-conceptual Truth, and you are drawn to it.
This reminds me of Metropolis Hastings (MH),3 which is an advanced type of Monte Carlo Markov Chain (MCMC) algorithm, that is, an algorithm that uses randomness to generate a “chain” of samples—very important in domains like Bayesian inference, where you often can’t sample the target distribution directly or with simple transformation techniques. Another example of an MCMC algorithm is the little word-prediction that your phone does for you as you’re typing. That’s a really simple Markov Chain, where you have a chain of words, and the suggested words are chosen based on how frequently they appear in similar chains of words. MH is a more sophisticated example of MCMC, where the chain is generated by taking an arbitrary starting point, perturbing it, and then either accepting or rejecting the perturbed point based on some heuristic. In this way you can create a chain, where each point is an accepted perturbation of the previous point, and you just keep on going until you have as many samples as you need. The interesting thing about MH is that if you define your perturbation function and your accept/reject criterion well, the resulting chain reflects the true properties of the underlying distribution you’re trying to sample. It’s an extremely important algorithm in Bayesian inference, but is it important for writing, too?
If writing is more like an MH algorithm, than being good at writing is more about developing a good perturbation function (how wildly should you change things in your edits? should you cut things out? add things in? rearrange?) and improving your accept/reject heuristic (what will the reader think of this? is it provoking the effect I am interested in? is it honest; is it True?)4
Or think of it this way. In “Borges and the Essay,” we discussed how writing a good essay is a matter of being like a river. Rivers look like they wind randomly (they look like Markov Chains), but the path they find to the sea is actually the most efficient. They run a simple algorithm: at each step, flow down. The algorithm of the essay is: at each step, flow interesting.
But the territory you’re exploring in fiction is much more treacherous. Simple Markov Chains might work for essays, but you need something more sophisticated for fiction. The target distribution is infinitely more complicated. Points no longer obey the properties of arithmetic. There is no obvious up or down to flow to, because you’re in hyper-dimensional space. “Flow interesting;” but make sure that the flow runs backwards and forwards, and repeats (with extra interesting) on the next run, and you can flow from any point in the chain to any another, by crossing beyond linearity, through transformations defined by tensors with ranks you could hardly fathom.
The Office Hours of Story Club always begin with a reader question, so in that spirit I’ll end with a question: if you’re a writer (or creative of any kind), what’s something that has helped your attitude and confidence with respect to your craft? Leave a comment, or reply to this email, or just reflect on it to yourself.
Having a (virtual, distant) mentor in George Saunders has made a world of difference for me, but there are other, smaller, things too—for example, having a nice computer makes writing a lot more enjoyable. Another thing I noticed early on is how beneficial it is to write down as many thoughts as I can, as soon as they occur to me. Before I started “actually” writing, I only wrote down the thoughts that possessed me—the things recurred endlessly, and that felt like they were good enough to try not to forget. But once I started writing down any idea or writing-related thought I had, I was not only relieved of the burden of these thoughts,5 I suddenly felt like I had so much to write about, and I realized that the really powerful thoughts were creating all sorts of other thoughts that were also worth keeping. Indeed, Orbis Tertius was born out of the labyrinth which is the Notes app on my phone; it is a labyrinth which was constructed compulsively, the work of an insane Daedalus… but here we both are, navigating it.
This idea of “honesty” and “truth” existing in writing, especially fiction, is very compelling to me. If you haven’t already, you should read Henrik Karlsson’s essay “Dostoevsky as Lover,” which gets at this idea. It’s also a big emphasis in Stephen King’s memoir On Writing. King makes the argument that a story is something that a writer discovers, not creates; that a writer is closer to an archaeologist than a sculptor.
William Faulkner’s claim that he wrote As I Lay Dying start to finish without changing anything notwithstanding. It’s an inarguably great novel, and it’s one of my favorites, but I have a hard time believing him. Sorry, Bill.
I worked on software implementing advanced versions of MH while I was in college. I think the important detail about that experience that’s relevant here is that the distribution you’re trying to sample from could be anything—we’re not talking like, Normal or Uniform or Chi-Squared, here. We’re talking weird shit.
In the same vein as [3], the perturbation function and the accept/reject criterion could also be anything. There is a lot of art to it, because, for instance, two different implementations could both give the same final results, but one could take orders of magnitude longer to run.
Big thanks to Brian Leli for helping me deliver the particular thought behind this essay from evil and into the light.