#1/n: Language Models and Writing
Hello again & intro to a new series
I am resuming posting here (under a new name) after a multi-year hiatus. I will be writing much shorter posts, about anything that interests me, and I won’t take much care over making them perfect. The one below is an intro to a series I have started about language models and literary / philosophical value. Probably I will also write about the other things that people who know me hear about rather too often: some novels/poetry, climbing, British political history, wine, learning a language, outlandish data-free political speculation. I hope you keep reading.
At work I use language models (LMs) both for productivity and as components in technical systems, and after work I sit at my kitchen table and try to write great works of literature. If you’d asked me a year ago, I would have said that by now these two versions of my life would be overlapping. But in fact they aren’t. I don’t use language models in my daily writing, except for small things like sourcing half-remembered quotes or concepts I don’t understand. (Most recently: share buybacks.)
Even so I get – almost daily – a confidence crisis as to whether there’s any point in trying to write. Writing, all things considered, looks like pretty crap use of my time: it’s a skill that gives you approximately zero-leverage, is psychically hard, is materially and emotionally unrewarding, and carries relatively low social status if you compare it something like acting or musicianship. Yeats claimed never to have done ‘more than five or six good lines in a day’. I’m very happy with 100 words. Benchmarked against an LM, that’s pre-industrial vibes.
My lack of confidence in myself as a human writer I think stems from a realistic assessment of the development of these models’ capabilities. If you’re someone that thinks LMs produce only a ‘fog of bafflegab’, then probably you’re prompting them wrong, or else you have not tried different kinds of fine-tuning, scaffolding, or RAG-style bolstering to give them material to work with. Plus the models we use as consumers models are not trained with a view to producing good writing, where ‘good’ means adhering to criteria derived from the literary canon, whatever that may be. The RHLF step used by most of the LM vendors is there to - among other things - prioritise lucidity and simplicity in a technical sense, but you could no doubt train the reward model such that you got it to elevate responses which look more like Austen or Emerson or whoever.
For example, the other day I vectorised my entire corpus and got Claude to write in my voice. I think it got within a few kliks of the uncanny valley.
I stood in front of the noodle shelf. Rows of packets, all the same shape, different colors. I picked three without thinking. Chicken, beef, whatever. The plastic felt cool and smooth in my hand. At the till, I put them down. The cashier scanned them. Beep. Beep. Beep.
It’s fine, not amazing. Whatever. Sure, we can list the defects of the paragraph: the double triples in ‘chicken, beef, whatever…beep, beep, beep’; the cliche ‘without thinking’; the bathetic pomposity of ‘At the till, I put them down’. Yet this list is missing the point; it feels like shitting on a football prodigy because he can’t tie his laces. I have no doubt that I could build a system which could write in my ‘style’ – that would be judged by a panel of good readers to be by me – and that would be better than anything I’d written – judged by a panel of good readers to be of superior quality.
The main issue is simply this: that I never wrote that paragraph, nor have ever wanted to write it. For me, the process of writing a novel for me is the process of writing itself. This is not a truism, or it is not just a truism. I am sure that there are writers out there for whom writing is the process of expressing ideas they already have. But for most of us, thoughts and commitments get determined by the very process of trying to write them. This is partly what we register when we note the sincerity of great writing: that no matter how much editing or repetition there might be, we feel like we are there with another mind earnestly struggling or under the influence of something important and perhaps difficult.
And it is not only that we perceive any person working with and through language, but there is also undoubtedly some sense in which it matters that I, in particular, was the one who wrote it. As Sloterdijk says in Rules for the Human Zoo, ‘books are thick letters to friends’. You read the words of an author because you are to a certain extent interested in what the author has to say and see such an example as relevant to your own thoughts and behaviour. It’s not hard to see how this gestures at a new humanism of sorts. The consciousness(es) behind the words might be what matters. Justin Smith suggests that we might ‘come to see writing not as a simple transfer of information, but as the witness and record of an inner life.’ It might be that LMs make us realise, contrary to ideas about material power, coercion, the intrinsic play of language itself, that you may want to read my writing simply because I am a person and I am limited and impractical. It is an index of actual experience, of ‘the shady media of the heart’.
So this is going to be a series of posts about how language models make us think about writing, writing-as-thinking, and literary quality/value. I will treat the following topics:
What could this ‘new humanism’ be, if anything? What could or should be the dominant model of literary interest/value in a world where LMs can write ‘better’ and ‘more’ than a human agent?
Is there instead a flavour of literary materialism encouraged by LM authors? If these models read the internet, are they in fact an index of some kind of power or influence in the sphere of communicative action?
What kinds of objects are novels written by LMs, and what they can be used for? Can we do the same kinds of historical or textual criticism that we do with human novels? Can we do more? Will Claude’s great masterwork actually be a better record of its time than was, for example, Bleak House, or a better retrospective than War and Peace?
What does it mean for writing to be an index of ‘quality of thought’? In a world where we can see ‘concepts’ as represented by weights in the hidden layers of a neural network, to what extent can we see thought as bound to the language in which it is constituted? What is thought without a thinker?
What can be distinctive about a human writer? Must the human writer define itself as the agent who investigates the things an LM can’t or won’t say? Will there always be aporia in language models in which human writers can find their role?
I hope to see you in later posts.