What was the language model?

Encoding/Decoding Cybernetics via Cultural Studies

Dr. Sarah Ciston

Fellow, Center for Advanced Internet Studies, Bochum

Visiting Researcher, Stuart Hall Archive Project, University of Birmingham, UK

Hello & Thanks I’ll talk about this project I did with the Stuart Hall Archives at the University of Birmingham, where I’m grateful to have worked with Rebecca Roach as a visiting researcher.

This project is broadly interested in the question “How did words become vectors?” Not just numbers but spatialised probabilities, reducing meaning to measurable comparisons and arrows in latent space?

It’s interested in how the technical choices that create numerical representations come to inform large generative text outputs — and how historical developments come to inform those technical choices. Both how those historical developments are represented and historicized and their potential alternative narratives that exist in parallel, for ways they might have or might still create alternative futures.

As an extension it’s asking about the current AI accelerations we see, where they require reductions in meaning making and in modes of representation of identity making, where they work to transmit and amplify power and ideology, how they do that work.

I have been interested in how cultural studies and other perspectives taken up from outside a white western canon might have differently informed a cybernetics in formation and by extension an AI discourse in formation.

With that in mind I came to the Stuart Hall Archives in Birmingham to look at drafts of Hall’s essay Encoding/Decoding.

We can therefore define the diagram in several different, interlocking ways: it is the presentation of the relations between forces unique to a particular formation; it is the distribution of the power to affect and the power to be affected; it is the mixing of non-formalized pure functions and unformed pure matter’. –Gilles Deleuze, Foucault (1986, 72-73)

Ideology and power affect the transmission of meaning

Cultural studies theorist Stuart Hall, “Encoding/Decoding,” combined versions from archive drafts and published editions of 1973, 1980. Illustration by Ciston.

Started by compiling the various subtle variations of the diagram found in versions of Hall’s Encoding/Decoding essay drafts.

As well as connecting the ideas found there to Hall’s intellectual history in other works around cultural studies and diasporic identity.

Plus I looked for any references to cybernetics or early computing.

Some things to note: - Hall focuses on meaning in this diagram, and how it moves. - For Hall, he says that ‘meaning is made in practices’ - He highlights this moves as a circuit (connects to his later collaborative work on cultural circuits) - And he sees the transmission of meaning as always imbued with and reflective of practices of ideology and expressive of power dynamics

Information theory

Mathematician Claude Shannon, 1948, in “The Mathematical Theory of Communication,” 7. Reproduced from original by Ciston.

My initial impetus for working with Hall and Cybernetics was the contrast of that diagram with Claude Shannon’s famous diagram that launched the field of Information theory.

It studies how information can be relayed as a signal from a source to a target, by a transmitter. And it accounts for noise that might interfere.

Shannon explicitly sets aside the content or meaning of the message as unimportant to his analysis, for the purposes of studying signal transmission.

And this move is perhaps fundamental to a kind of mindset seen in many approaches in what would become the field of computer science from the mid 20th century onward.

Machine learning’s visual tropes through archives

1. Layers of truth
2. Proximity & spatialisation of language
3. Nodes of thought
4. Encoding-Decoding Circuits of Transmission

Noticing what is not there

- Where is meaning?
- Where is power?
- As meaning and power combine: Where is ideology?
- How would machine learning today be different if these concepts were part of their formative structures?

So from these starting points I began looking for other diagrams, in particular asking what visual patterns I might follow and asking what is not there.

Some key learnings so far are that > the concepts of computing have been around much longer and come from non-traditional, multiple places at once > concepts could easily be traced from elsewhere, formed in much different ways w diff results

Hall & Shannon are the kickoffs for these four in particular…

The diagram itself is a useful trope for thinking about power, technology, and communication of meaning, as Deleuze argues, when he says,:

Layers of truth

Claude Lévi-Strauss, 1955, in “The Structural Study of Myth,” Journal of American Folklore, 435.

This visual becomes reminiscent of the representations of LLM layers, datastacks and n-of-however-many sets of invisible layers in black boxed systems.

Key to this idea is the sense of a super dimensionality that can be chartable and useful as a text beyond the version of the text on a page before us, a para and hypertext ripe for embedding and extracting.

Lévi-Strauss suggests charting myths in three dimensions, with the depth of multiple versions, and the diagram portends later deep learning model illustrations:

“Thus, our method eliminates a problem which has been so far one of the main obstacles to the progress of mythological studies, namely, the quest for the true version, or the earlier one. On the contrary, we define the myth as consisting of all its versions; […] We will then have several two-dimensional charts, each dealing with a variant, to be organized in a three-dimensional order so that three different readings become possible: left to right, top to bottom, front to back. […] experience shows that any difference to be observed may be correlated with other differences, so that a logical treatment of the whole will allow simplifications, the final outcome being the structural law of the myth.”

LeCun et al., 1998, “Gradient-based learning applied to document recognition.” Proceedings of the IEEE, 1998-11, Vol.86 (11), p.2278-2324 (convolutional neural nets)

Convolutional neural nets began to be developed in the late 1990s, and built on an argument from the bio sciences that believed brain neurons could each store pieces of information and be combined to collectively process more complex, higher-ordered information.

Similar to Levi-Strauss’s figure, this illustrates the layers of a CNN designed to process a handwriting dataset and identify its characters.

Spatialisation of language

Linguist Charles Osgood, 1957, in The Measurement of Meaning

Look at the Spacialization of language — we may take for granted this element now, but I am interested in how this came to be.

In natural language processing, it is often attributed to this oft repeated quote from 1957 from JR Firth, “the meaning of a word can be inferred by the company it keeps.” And this remains a foundational principle of language-based ML today.

One of the first places we see this applied is in work by Osgood from 1957, where he maps language to a spatial, cartesian plane.

“Even in technical use, vectorization is a tool for abstraction, for transforming ordinary tabular data into malleable orientations in multidimensional space (cf. Mackenzie 2017). Vector spaces are the symbolic terrain on which much of the labor of machine learning works, and they provide a widespread metaphorical language across the software industry. Startup founders describe their employees as vectors; venture capitalists describe the companies they fund as vectors; in ordinary conversation, engineers will describe unrelated things as “orthogonal” to each other.” ¹

1. Seaver, N. (2021). Care and Scale: Decorrelative Ethics in Algorithmic Recommendation. Cultural Anthropology, 36(3), Article 3. https://doi.org/10.14506/ca36.3.11 . Mackenzie, A. (2017). Machine Learners: Archaeology of a Data Practice. The MIT Press.

Schematic-illustration-of-semantic-projection-Word-vectors-in-the-category-animals

Become the contemporary conceptualizations of ‘latent space’ with which computer scientists imagine a secret world of language and meaning spatialized and navigable in points between words-as-vectors, where king equals queen minus woman plus man.

But as Nick Seaver argues… (back slide)

Google computer scientists Mikolov et al. 2013. “Efficient Estimation of Word Representations in Vector Space” (Word2Vec)

As a segue into the next section on nodes, can also think about how the processing of these vectors are represented, with weights (or mathematical probabilities) attached to vectors that modify their numerical values at each step. This is a diagram describing Word2Vec, a popular algorithm for making word representation using numbers.

And you can see how it resembles Shannon’s diagram, as well as the node-based diagrams we’ll look at now.

Nodes of thought

Frank Rosenblatt, 1958, Mark I Perceptron

The first proposal for modeling computing using a set of nodes, what would become neural networks much later. Was still an analogue process. Was fairly widely dismissed.

Computer scientists Kanal & Randal, 1964, “Recognition system design by statistical analysis”

An example of computer science’s military roots: This early object recognition ML was developed to detect tanks in images

Visually has in common with the convolutional neural nets from the late 1990s that ended up spawning a whole erra of contemporary computing.

Computer scientists Marvin Minsky & Seymore Papert, 1969, “Perceptrons: An Introduction to Computational Geometry”

These same visual tropes plugging into zones of an image and processing them in parallel (which we’ll see continue today) get taken up by Marvin Minsky and expanded

Computer scientists Rumelhart, Hinton, et al. 1986. “Learning representations by back-propagating errors”

This idea of nodes gets combined with the idea of layers we saw before in order to produce multi-layer perceptrons, operating in parallel, to make perceptrons useful at scale. Visually we’re getting closer to the simplicity of Shannon’s initial diagram from 1949 or Hall’s revision of encoding/decoding happening a few years before this (unlikely they would have seen).

MLP backpropagation was developed simultaneously by several groups of researchers.

Computer scientists Sepp Hochreiter and Jürgen Schmidhuber, 1997, “Long short-term memory,” Neural Computation

Then by the late 1990s lstm networks are producing reliably competent early language models that still echo Rosenblatt’s 1950s attempts and the handmade craft practices of weavers and patternmakers.

Encoding-decoding circuits of transmission

Linguist Roman Jakobson, 1960, in “Closing Statement: Linguistics and Poetics,” Style in Language, 353.

In 1960 Linguist Roman Jakobson who was close with the cyberneticists presents this model of communication theory that appears a riff on Shannon’s but with message included (still not meaning.)

Communications theorist Wilbur Schramm, 1960, The Process and Effects of Mass Communications

Also in 1960, Communication studies was developing as a field at the same time as computer science, both in some ways out of cybernetics, and also got in on the diagraming of the transmission of content or meaning.

Here we see encoding-decoding language of computation appear (not necessarily for the first time), and it acknowledges a mass culture element of communication — that it does not happen just between individuals but is influenced by social and cultural forces.

Computer scientist Joseph Weizenbaum, 1964-1966, archive sketch for ELIZA chatbot

This is an example of another archive I worked with previously, the Joseph Weizenbaum papers at MIT Archives. His papers include the original source code for the early chatbot ELIZA, as well as this diagram describing the flow that the software should take. It includes input and procedural processing of text, and a capacity for short term memory storage that gave a more robust illusion of verisimilitude.

With the divide into two parts, a reader system and response system, which is not quite how the system ended up being built in practice, the diagram echoes many of the other diagrams here that include encoder and decoder or similar transformation rules.

Robert Darton, 1982, Communications Circuit in “What is the history of the book?”

This is Darton’s communications circuit that appears in 1982, which attends to the production of writing — manufacturing and editorial processes. It includes the medium of text and the culture around it — but again not the content of the book itself or any meaning it might contain (especially in relation to anything else)

Cultural studies theorists Du Gay, Hall, & Janes. 1997, 2013. Doing cultural studies: The story of the Sony Walkman

By 1997 Cultural studies is well established and presents methods of ‘close reading’ mass media and its objects, and this circuit of culture thinks through how a transformative media object and physical technology like the Sony Walkman also shapes users’ sense of identity in relation to it as well as is shaped by its manufacturer’s processes and so forth.

Hall’s work in the archives lays really interesting ground work for this work, including unfinished projects directly around Meaning and Methods.

Game studies scholar Espen Aarseth, 1997, Cybertext

There are nods at this from e-lit which is also beginning to get a foothold with the more widespread introduction of hypertext in the 1990s, and Arseth’s “text machine” is a nice analog to the different kinds of text machines being built in natural language processing circles.

Google computer scientists Vaswani et al, 2017, “Attention is all you need” (Transformer models (GPTs)). Additional detail added, illustration by Ciston.

And then I touched on a few innovations of NLP in between, but another big transformation happens in 2017 with the introductino of the Transformer model, which presents the concept of ‘self-attention’ and builds a mechanism for word vectors to be altered in relation to every other word vector in their vacinity — at massive scale. Ushering in the current wave of Large Language Models.

The transformer model’s circulatory diagram seems a reflection of many of the other diagrams shown — and is then directly embedded as a step in many of the systems that are currently used now.

Llama Team, AI @ Meta, 2024, “The Llama 3 Heard of Models”

Here is Meta’s Llama model, with a diagram showing their Llama 3 architecture and training, and their reinforcement learning processes. Reinforcement learning is the training that happens after the initial steps of creating a model, that then prepare the model to be honed for its specific purpose (in this case as a general-use chatbot with various capacities and guardrails)

The transformer model is present in the sections labeled ‘self-attention’ and ‘feed-forward network’ and I’d also point out the ellipses are a small visual rhetoric suggesting layers upon layers, like the boxes we looked at before.

Even this diagram has a certain Shannon-like simplicity that simplifies the model to a black box. Information or data go in, and responses from an intelligent agent come out.

Chat model with reinforcement learning

Thank you to the Stuart Hall Archive Project for the intellectual, material, and financial support to explore Hall’s work through these lenses. And with particular thanks to Professor Rebecca Roach for the invitation, encouragement, and many conversations. – Sarah Ciston

So with that I’d just invite thinking together about how the images we encounter in current systems, in current AI tropes, might echo archival imaginaries across many fields, or might limit the possibilities for systems we could otherwise imagine.

And I’d like to think more about what it could mean to apply alternative frameworks, like Hall’s work, to systems that have such formal visual language and mythology already but could be otherwise.

Thank you!