The Hyperreal Substrate

Any computer program can eventually be compiled down to a sequence of binary logic operations, and any general-purpose computer can reproduce its equivalent behaviour. This didn’t happen on its own, and it’s the product of a very long historical process that had to bring together and reconcile lots of individual strands of scientific progress both on the hardware and maths side.

During the development of computer languages, the syntax and formal meaning of computer programs were precisely defined and could be checked mechanically, but the judgement of the correctness of the behaviour has always been left to humans. The deference to humans at the moment of interpretation is a common theme; symbolic computation, Cyc^[1], and the Semantic Web have failed to create a binary encodable ontology which could bridge the gap between the real world and it’s mathematical models.

Generative AI centred coding practices subvert the deterministic nature of this process. Instead of humans handing in instructions to a computer in a programming language, which doesn’t need interpretation (of intent), we now hand in natural language, which does need a lot of interpretation.

In the Death of the Author, Roland Barthes claims there is no stable meaning in text, the meaning is created when a reader interprets the text through their subjectivity. On the machine side there is no subjectivity, hence no meaning. The meaning of the text generated by AI is deferred – it is unmeant until reading.

Generated code sits in a special category, because when generated it’s still unmeant, yet operative. It would compile and work, deterministically, thanks to all the effort that has gone into ensuring its reliability through layers of formal definitions beneath. By contrast, prose is unmeant, and inert until read. It exists because the model was trained on natural language that described what the code does. The descriptions were meaningful, and provided a map for the model. The generated code would run and function, without the need for meaning.

Baudrillard talks about the “precession of simulacra”, where the map comes before the territory. The natural language description of the software is the map, and the generated code is the territory it engenders. It’s not a translation of a perception of the world but its own referent. When the code is generated, it would run before a human reads it, and the closed-loop operation sidesteps the subject entirely. LLMs aren’t trying to bridge the subjectivity gap, they’re trying to abolish it – just like Baudrillard described. The hyperreal is no longer burdened by a subject intervening, slowing it down with the translation from reality and interpreting it. This is what the hyperreal intends to do, operate on the pure sign level.

The subjectivity gap is structural, it’s there because we don’t know what knowledge, meaning or intelligence is. We keep circling around it, have another take, and fail. Come up with new methods during an AI summer, and then go back to licking our wounds during an AI winter – formalise our learnings and create another layer of determinism to build upon. We’re hopefully making gains, but like everything based on statistics, the returns are diminishing with each run.

This doesn’t stop us turning things around slightly each time. The gap changes shape, because the way we perceive the world changes (it now includes LLM coding), though never closes. New things get built on top of the previous efforts, our current hyperreal becomes the new normal for the next generation. They build their own simulation on top of it, and the cycle continues.