🔮 Safety in numbness

One who has learned the elements of a science and wishes to go further will find in books [...] new signs needed to rise to further truths.

Standardization was meant to propel fields forward. It was a ladder to scramble onto the shoulders of the giants, then push beyond.

Beyond meant different. Sometimes it’s a matter of magnitude, of going further in the same direction. Sometimes it’s less about direction than about adding new dimensions, ideally orthogonal, to explore. Or revisiting well-trodden ground with new sensibilities. But most importantly, it meant new.

Reading Erik Hoel’s excellent essay, “How the MFA Swallowed Literature,” reminded me of what happens when standardization does the opposite. Instead of raising everyone up so they can see further, standardization can comfort us to flatten the world. Erik describes literary fiction being deracinated by a protocol that became ubiquitous. I think it foreshadows what LLMs might catalyse in other fields, both near and far.

Erik argues that the rise of MFA in Creative Writing programs has steamrolled literary fiction. The key is that these programs have a standard formula: group workshop after group workshop, critiquing each other’s drafts. Rinse and repeat.

It’s a protocol: take aspiring writers and, through repeated peer review, sand away peculiarity, weird ambition and the unfamiliar edges most likely to attract an objection. Out the other end comes writing that’s homogenized and optimized to survive the workshop: “wan little husks of ‘auto fiction’”, as Joyce Carol Oates put it.

An entire economy has grown around a defensive style of writing that minimizes the novel’s “attack surface.” There were 15 MFA programs in the US in 1975 and 250 by 2012. Today, Erik argues the craft of literature and the program supposed to teach it have become identical. “One ate, digested, absorbed the other”.

The point of a literary culture is to find authors who help us rethink the canon. I think of Salman Rushdie’s Midnight’s Children, and the unruly memory that carries the myths, languages, and history of an independent India. He expanded the canon, changed what an English novel could be. He also gave new writers taller shoulders to stand on. Rushdie’s innovation was first a style, then a category, then a set of conventions. And finally it became mundane. Now we need to find the next writer who changes the form. Homogenization, which I think of as safety institutionalized, makes this harder.

It’s a problem of systems meeting incentives. In literature, the system is how creative writing gets taught and how that slowly makes each writer’s pages remind you of the others. And downstream, the publishing machine molds to the shape of that output.

But it’s not a process limited to fiction. A key purpose of any creative domain, be it scientific or artistic, is to find the next, not to repeat the last, with a tiny increment. Some systems foster modest speciation; others, a Cambrian explosion. The wrong system paired with short-term incentives will encourage the former.

And this is where the LLM comes in. An LLM is a standardised system, trained against particular incentives, to output against a particular statistical distribution with its very own shape. We’ve all experienced it.

At its most shallow, it is the words they spit out. But when you work with them more and more extensively, you can feel a common conceptual architecture of LLM thinking. Applied as a rubric, they will amplify their sameness in everything we use them for. In that sense, they parallel workshops that sand away sharp edges; they norm by finding the middle of a distribution. Outliers are unwelcome.

Research is now done at an industrial scale. A University of Chicago statistician put his name to more than 200 papers and 14 books in nine months. SSRN, the preprint server, has pulled 257 of his works, citing “unusual submission activity”. Just as MFA workshops teach writers to avoid critique, AI helps academics game citation indices and peer-review checklists. Politicians too, increasingly, put a machine’s words in their mouths.

Standardization is useful. It’s a ladder that gets you up. But it can become too comfortable, and something massive gets lost – intellectual bravery.

A decade ago, I was racing to get one of my children to an exam. Parking was going to be terrible, so we took a black cab. Dave, the driver, had two phones, one running Google Maps and the other Waze. I joked that with a London cabbie’s hippocampus, he didn’t need them. He replied, “I don’t, but I need to know where the idiots are going so I can avoid them.”

Researching my new book got hardest in the past year. The temptation is to fall back on AI. LLMs produce feedback on demand, the critique-resistant surface Hoel describes. But the feedback is a norm, the middle of the distribution, anchored to the average of what has been said before, not what has yet to be said.

I use LLMs to stress-test my arguments. But the further I ventured into new terrain, the less helpful they became, insisting novelty was nonsense; dragging me back to the status quo. But I realised what Dave had told me a decade ago. Waze told him where every other driver would be. The LLM told me every path that had been trodden.

They hold all of human knowledge except what I’ve been thinking right now. But they can tell me what everyone else had thought. So when I find myself fighting an LLM to understand my argument, I take it as a signal that I’m not compliant. I can be confident the idea is new, even if that novelty widens its attack surface.

Standardization gives a veneer of solidity at the expense of creativity, idiosyncrasy and inspiration. It will create shelves of novels indistinguishable from one another, unremarkable except in their lack of originality. It will produce arguments that everyone else has already made.

Condorcet, writing more than two hundred years ago, was alert to this problem. That is to mistake the act of discovery for props “that are nothing more than the dictionary and grammar of its real language.”

Standards come in many forms – social technologies like the MFA, technical ones like Condorcet’s signs, computational ones like LLMs. All are helpful, as long as they’re just one step you take, and never the last.