A few weeks ago I was walking with Erik Groeneveld in the woods. He is very friendly and quite all-round brilliant; he has been building his own house from hemp and is at the moment helping young people in a skills school near his house. Last year, he explained LLM’s to me in a way I could understand – how texts are turned into tokens (pieces of text a computer can handle) and how each token is then given a position in a semantic space as an vector: a long list of numbers that functions like a set of multidimensional coordinates, locating the token in relation to countless others.
It took me some time to fully understand what that means. Take the word bank in a sentence. The sentence is broken down into tokens. The token is embedded in a generic position in semantic space. As the model processes the surrounding words, its vector (and therefore its position in the semantic space) changes. In a sentence about money, it moves toward finance; in a sentence about rivers, toward water and geography. Billions of learned weights govern this movement, directing the model’s attention to whichever surrounding words matter most for interpreting ‘bank’ in context – weights that are themselves the residue of statistical correlations learned during training. Strictly, embedding is the act of placing a token in this space; in practice the word is also used for the resulting vector itself. People cannot read or interpret these vectors; they are a mathematical representation that computers can manipulate, but humans cannot. In a model the relations are implicit and not fixed; there is only space, no boundaries between categories. This is a core difference with knowledge graphs, which represent entities and relationships explicitly.
Back to the woods. Sitting on a log, Erik told me he had just downloaded a new model and was playing around with it on his computer. He was quite surprised at the amount of general knowledge this 21GB download appeared to contain. He mused on the nature of the thing he downloaded: you cannot retrieve the original from it, but there is a certain compressed mathematical representation of it. It led him to the conclusion that it is helpful to think about these models in terms of ‘lossy compression’.
This reminded me of a book I had just been reading by William Marx, Libraries of the Mind. In it, he describes what is left of a book in the mind of the reader after the original has been laid down. Using Madame Bovary as an example, he suggests the languidness and frustration, the provincial context, perhaps a few strong scenes and images and a general feel for the book is what a reader retains. I would be hard-pressed to quote from Madame Bovary literally, but do strongly remember my mixed feelings of irritation and compassion in reading it. Books are created in the mind of the reader nearly as much as in the mind of the author. People are endlessly different and reading experience is mediated by personal history and identity. What a reader retains after reading is another form of ‘lossy compression’.
So is there a parallel between these two forms of ‘lossy compression’? Yes and no. Yes: because both involve s a real reduction. No: because in humans, that reduction is embodied and personal and in machines it is mathematical and statistical. What survives in a human mind is shaped by experience, identity, time and emotion. What survives in a language model is mathematical – a position in a space learned from statistical regularities across billions of examples.
The short-cuts of the human mind mean we often use words thinking we mean the same thing, when that is not necessarily true. I wrote a master’s thesis on how the future of the book depends on what you mean by the word book. As philosopher Richard Rorty argues: ‘Words create worlds’; and Alfred Korzybski poses: ‘The map is not the territory’. In that sense, human minds and statistical models have a correlation (they might even be near each other in the semantic space) – which is probably why the word artificial intelligence was coined in the first place. But they are fundamentally and unreconcilably different in that humans live and feel, and machines don’t.
This ties into a rich conversation I had yesterday with Marian Hellema, a friend with whom I have been sharing a two person reading-group for several years, making our ‘reading reception environment’ more cohesive than usual (we can relate back to books read before and/or make comparisons). She pointed out that she hears the word ‘synthetic’ being used in two different ways in the discussions about AI – as ‘anorganic’ and as ‘summarizing or bringing together’ (like Jo Guldi does in her research). This resonated with me; a language model is is synthetic because it is not embodied; it has no lived experience, emotions or memory. But it is also synthetic because it compresses the statistical traces of millions of texts into a single representation. Every individual voice is reduced to part of a larger pattern. The result is often bland, averaged and clichéd. Yet that same process allows the model to generate, in response to a prompt, a single, personalized text that feels remarkably coherent and true, even when it is not.
AI synthethises without experience, it is what makes it artificial. Humans synthesize without the range of the models. Both have a role to play, but not the same one.
July 2026