Earlier today, I attended a lecture by Antonio Somaini at the University of Oslo’s seminar series on aesthetics. There were many interesting things in the lecture, which apparently was based on the manuscript of a forthcoming article. Here, I will focus on one thing that caused my interest: the thinking of latent spaces as an “anarchive”.
Latent spaces
A latent space is a mathematical representation used in machine learning where complex data (like images, text, or sound) is mapped into a lower-dimensional coordinate system so that similar items are positioned closer together. Instead of storing explicit labels or complete files, models encode underlying features and relationships, enabling interpolation, clustering, and the generation of new outputs by moving through that space.
In practical terms, the latent space is where a model’s learned structure of the world is organised and explored. In musical thinking, it can be thought of as the plan we make before an improvisation. It is bound by a certain musical style and the possibilities and limitations of the people (and machines) and instruments involved.
An interesting consequence of this is that, in large models, you may never produce the same output. Because the way you “move” through the latent space will always differ, hence the output will also vary slightly, and will never be exactly like any of the inputs it was trained on.
Conceptualising archives
In his talk, Somaini introduced three interrelated dimensions when it comes to understanding archives:
- Technical apparatus – the media that store information (books, photographs, databases)
- Institutional apparatus – the organisations that collect and organise it (libraries, museums, states)
- Epistemological apparatus – the systems that determine what can be known or said (classification systems, historical discourse)
Drawing on thinkers like Jacques Derrida, Michel Foucault, and Bernard Stiegler, the lecture showed how these layers are always intertwined. The short story is that archives do not simply (and statically) preserve the past; they shape it.
Latent Space as archives
Here came the interesting twist in the lecture. Given that large learning-based models can “store” lots of information, should they also be considered archives? Yes, in the sense that they do preserve information. However, they do so very differently from traditional archives. Instead of preserving discrete records, they reorganise data into latent structures based on similarity and statistical correlation.
This, Somaini argued, has several consequences:
Provenance becomes unstable: We can no longer easily trace who created what, or when.
Context is stripped away: Data is detached from its original conditions.
Meaning becomes relational: What matters is not origin, but position within a network of similarities.
Following from the last point, this relationality means the information is not fixed but potentially changing. Facts become relative.
The Anarchive
It was at this point that Somaini brought up the concept of the anarchive. Pierre Cassou-Noguès and Gwenola Wagon have proposed this term to describe computational memory systems that do not conserve fixed documents, but continuously reorganise cultural traces into new, generative forms. In this sense, the anarchive is less a repository of stable records and more an operational field where new historical imaginaries can be produced from statistical relations.
Latent space is precisely such a structure. It does not preserve the past as a set of traces, but transforms it into a field of potential recombinations. If the traditional archive is about what was, the anarchive is about what could have been.
From Third Memory to Fourth Memory
Bernard Stiegler famously described modern technical media as forms of “tertiary retention”, externalised memory stored in material supports such as books or recordings.
Latent spaces (and anarchives) could be the sign of a “fourth memory” that is recursive, automated, and generative:
- It does not simply store traces of the past
- It generates new outputs from patterns within data
- It blurs the boundary between document and creation
This could be seen as a form of “ghost memory”, images and texts that feel familiar but are not tied to any specific origin. Thus, AI can be understood as both a preserver of cultural memory and an effacer that erases its context and provenance.
A new historicity
What is at stake, Somaini argued, is not just technology, but historicity itself. We are currently moving from a world in which the past is preserved and interpreted in archives to one in which the past is generated and reconfigured.
Within the new anarchives, history becomes probabilistic, memory becomes dynamic, and the archive becomes a space of possibility rather than stability. This offers many possibilities but fundamentally changes how we think about data as facts.
What remains is something both powerful and unsettling: a vast field of ghostly traces, waiting to be actualised.
Thanks to CoPilot for helping draft this blog post based on my scribbles and photos of slides from the presentation.