The Différance Engine

A short note on a lecture by George D Montanez

In an old article I said that continental philosophers may have ‘deeper and more interesting things to say about the philosophical implications of AI’ than their analytical counterparts. Back then I used the example of Walter Benjamin’s essay ‘Art in the Age of Mechanical Reproduction’ (1935), showing how the twin concepts of ‘aura’ and ‘authenticity’, when combined with Frederic Jameson’s ideas about ‘pastiche’, seem almost ready-made for analysing AI art. On reflection I could have taken the more obvious example of the theory of hyperreality developed by Jean Baudrillard - a widely acknowledged guilty pleasure for Anglo philosophers these days - or just about anything in Deleuze. I used to think Adrian Moore had gone mad when he said Deleuze was one of the greatest philosophers ever to have lived. I now believe he was much closer to the truth than the popular despisers of post-modernism who see in Deleuze a heretic particularly deserving of fire and damnation.

A recent lecture by George Montanez, associate professor of computer science at Harvey Mudd College, has added another strand to my thesis. Montanez expertly threads together different examples of how chatbots can be gamed to produce manifest nonsense. The techniques used include forced reward hacking, interpolating irrelevant information during task-specification and training the AI on its own outputs (the so-called ‘curse of recursion’). His overall argument is that such malfunctions imply that LLMs are not really ‘thinking’. They are better understood as purely syntactic symbol manipulation machines ‘pushing around pieces on a game board’. The pieces do not have any intrinsic meaning. What a piece ‘means’ is determined by the position of every other piece that it can be distinguished from in the library of pieces the machine has been trained to process. In other words, its meaning is purely a matter of ‘difference’.

Yes, the theory-bros, primed by the pun in the title, know exactly where I am going with this. In his presentation Montanez likens what LLMs are doing to the sort of thing Hilbert wanted to achieve when he attempted to demonstrate that the entire ‘thought content of mathematics’ could be represented as a closed ‘formula game’ (ein Formelspiel) without any extraneous considerations. But to my mind the relevant philosophical comparison is with the big daddy of continental philosophy, Jacques Derrida. Specifically, all of this sounds eerily like the theory of meaning as ‘différance’ that emerges in his early critical writings on Husserl.

Having said this, I should come clean that I am, in my habits of thinking, much more formed by the analytical style of doing philosophy than I am by writers like Derrida and Deleuze. These are treacherous waters that I struggle to navigate. So let me instead quote from one of the few Theory-whisperers on the analytical side of the philosophical divide, the severely under-rated Christopher Norris:

‘... ‘Différance’ [is] a neologism that Derrida coined in order to suggest how meaning is at once ‘differential’ and ‘deferred’, the product of a restless play within language that cannot be fixed or pinned down for the purposes of conceptual definition. It is a cardinal precept of modern (structural) linguistics that signs do not have meaning in and of themselves, but by virtue of their occupying a distinctive place within the systematic network of contrasts and differences which make up any given language. And this picture is complicated, in Derrida’s view, by the fact that meaning is nowhere punctually present in language, that it is always subject to a kind of semantic slippage (or deferral) which prevents the sign from ever (so to speak) coinciding with itself in a moment of perfect, remainderless grasp. In French, the anomalous ‘a’ of ‘différance’ registers only in the written form of the word, since when spoken it cannot be distinguished from the commonplace, received spelling. And this is precisely what Derrida intends; that différance should function not as a concept, not as a word whose meaning could be finally ‘booked into the present’, but as one set of marks in a signifying chain which exceeds and disturbs the classical economy of language and representation’

When Norris says that signs do not have a meaning in themselves but only by virtue of their place in relation to other signs, he does not mean by that a logical relation. The theory of ‘différance’ comes out of an attempt to deconstruct Husserl’s ‘pure logical grammar’, analogues of which can be found in early Wittgenstein and Frege. As Derrida says in Speech and Phenomena (1967), the book where he first coined the term, ‘signification functions not by virtue of the compact force of [words] cores but by the network of oppositions that distinguish them and relate them to one another’. The topology of these networks or graphs does not ‘fall from the sky’, nor is it ‘inscribed in the wax of the brain’, but is the product of a completely arbitrary play of differences.

Given the forcefulness of Montanez’s examples, I am more persuaded than ever that (1) LLMs do not think and also that (2) differánce is not the whole story when it comes to language, though I do agree with Norris that it constitutes a genuinely powerful critique of what was a popular theory of meaning in the first half of the twentieth century. Reading Speech and Phenomena we are less inclined to go along with Husserl when he says that ‘everything that is can be known in itself’ and that the content of what is known can be expressed ‘through wholly determinate word-meanings’. And yet, I suspect that to explain the differences between LLM-style simulations of intelligent thought, and intelligent thought itself, we will find ourselves going back to Husserl, in particular to the critical chapters of the Logical investigations where he introduces the notion of ‘intentionality’ and the about-ness of consciousness.

Setting that aside, it is nonetheless amusing that when Samuel Butler wrote his fantasy Erewhon, in which mankind decides that the threat to civilisation posed by the possibility of thinking-machines is such as to necessitate making ‘a clean sweep of all machinery’, the most advanced machine around was Charles Babbage’s Difference Engine. Today, when some of us are persuaded that we are on the cusp of the sort of disaster that forced the people of Erewhon to take all of their machines to bits, the menace comes in the form of a Différance Engine.