Recommender Systems (RS), the case study of this essay, are one of the most used intersections with media on cloud platforms⁶, producing addictive selections of media (Chun 2016) you didn’t know existed and making the unfathomable challenge of exploring the abundance of cloud media possible. RS does this through automated valuing, sorting, ordering, and breaking down the mass of cloud data into hierarchies and comparative assemblages (Noble 2018) reinforcing how we imagine relationships as essentially comparative and hierarchical. The more classic and clear examples of these hierarchies are the likes of Noble’s critique of google rank, where results are ranked within a linear list, from most valued to least, but with the likes of talkative search engines (GPT utilized by Bing), and the complex animatic dynamics of feeds these hierarchies are being performed through new complex and less transparent forms which I will attempt to critique and re-value through configure-able methods.

The current configuration of media platforms and their RS algorithms are performed through monolithic process of control and ownership, giving them greater agency within the social/political relationality of ability in their network. Their socio-technical infrastructures do this by gatekeeping, augmenting and ordering our experience and representation of the world through them. These monolithic configurations of platforms are currently being (have always been) challenged by open and free software commoning of potential other configurations. One of the predominant examples, especially after the twitter takeover, is Mastodon on the activity pub protocol, providing an infrastructure to create decentralized⁷ and federated⁸ networks and communities. Through these new configurations you see how there is a shift within the political/ relational model of ability (Kafer 2013) within the infrastructure, enabling the people using it to configure and form their communities to their needs and desires. This shift moves away from the gatekeeping and extractivist methods of corpo-social platforms to ones that enable people to control their data, move their accounts to different servers⁹, use the same data between different platforms or access it from different software¹⁰. Most importantly this relationship creates a space where people can develop their own abilities to create communities and places safe from infrastructural inequality and harm¹¹. This reformation of infrastructure to enable people to define their own configuration of imaginaries and their matterings is a prime example of what I am calling configure-able methods.

Much like Chun and Barnett’s analysis of Latent vectors¹⁷ (Chun and Barnett 2021), RL systems are configured not only through imaginaries and the language of segregation, exploitation, and extraction, but also mathematic principles such as optimal control. In many AI systems, optimal control is embodied in their configuration, but is implemented through their loss function. Loss values the performance of a system to then tweak its potentially many thousands to billions of parameters towards an optimal outcome. Loss as a metaphor figures these complex dynamics into singular narratives of prospective loss and gains (economically) which can be navigated through Ferreira da Silva’s (2016; 2017) analysis of efficient causality¹⁸. She traces this thread through Newton, Descartes and Galileo, and sees it extracted from the rest of Aristoteles' four causalities (Falcon 2023). This configuration of efficient logics relies on the functionalities of mathematics to legitimize claims, instead of accounting for the whole, or final cause. Through this reductionist approach algorithms are only legitimized if they work, and not through a more complex relation of how they work, and what effect they may have on the environment around them. The legitimizing mathematics of the loss function that “work” are defined by rather rudimentary logics of differencing¹⁹, how far spatially and geometrically is the “truth” from the action, often squared to exaggerate differences and clearly define outliers. Through configure-able methods we must question these outdated and destructive dynamics to then imagine how we could configure these infrastructures otherwise (Ferreira da Silva 2016; Pritchard 2018; Soon and Cox 2021), implementing contemporary STS, social science, crip trans*feminist and intersectional approaches and models into their configuration.

Much like Chun and Barnett’s analysis of Latent vectors¹⁷ (Chun and Barnett 2021), RL systems are configured not only through imaginaries and the language of segregation, exploitation, and extraction, but also mathematic principles such as optimal control. In many AI systems, optimal control is embodied in their configuration, but is implemented through their loss function. Loss values the performance of a system to then tweak its potentially many thousands to billions of parameters towards an optimal outcome. Loss as a metaphor figures these complex dynamics into singular narratives of prospective loss and gains (economically) which can be navigated through Ferreira da Silva’s (2016; 2017) analysis of efficient causality¹⁸. She traces this thread through Newton, Descartes and Galileo, and sees it extracted from the rest of Aristoteles' four causalities (Falcon 2023). This configuration of efficient logics relies on the functionalities of mathematics to legitimize claims, instead of accounting for the whole, or final cause. Through this reductionist approach algorithms are only legitimized if they work, and not through a more complex relation of how they work, and what effect they may have on the environment around them. The legitimizing mathematics of the loss function that “work” are defined by rather rudimentary logics of differencing¹⁹, how far spatially and geometrically is the “truth” from the action, often squared to exaggerate differences and clearly define outliers. Through configure-able methods we must question these outdated and destructive dynamics to then imagine how we could configure these infrastructures otherwise (Ferreira da Silva 2016; Pritchard 2018; Soon and Cox 2021), implementing contemporary STS, social science, crip trans*feminist and intersectional approaches and models into their configuration.

So many of the tools to do these sorts of reconfiguring either lay within the complexities of code and its hierarchies of accessibility and understanding, where they may take years to learn and be able to implement, otherwise they are often very awkwardly arranged through predefined and limiting paid for and/or extractavist cloud platforms. Runway ML, one of the original for-profit creative AI tool platforms, is configured by taking open-source AI models and implementing them through their web interface, charging for the computation of the training of models and generating of content. The key here is that most accessible home machines and the people using them would not be able to train or run these algorithms, forcing most people to only enable them through some sort of cloud computing infrastructure. This can also be said about the creation of datasets for these models, which would be difficult and very laborious to produce if you did not use some web scrapping or other cloud extraction methods. This relational ability is again not primarily situated in one place, such as the hardware, but it is also inferred in the attitudes and politics of investors and developers focusing on the growth of these large-scale and inaccessible tools and infrastructures.

To practice configure-able RS I am coming together with my studio, Imaginary Practices (IP)²¹, to collaboratively develop different works to imagine through practice how RS and AI infrastructures might be configure-able. Informed by accessible and transformative AI projects, like Rebecca Fiebrink’s Wekinator (Fiebrink 2015) and Gene Kogan’s ML4A(Kogan 2016), this project aims to form new intersections into these mechanisms, enabling RS and AI to be configure-able to different imaginaries and mechanisms. The collaboration with IP has so far produced a few preliminary tests starting to form new tools and dynamics within AI infrastructures, but for this essay I will discuss two.

The first project is called the F*MNIST dataset, it is a dataset that we made collaboratively with a wide group of feminists, informed by Fermenting Data (Tyżlik-Carver, Rossenova, and Fuchsgruber 2022). Whilst forming the data we created a conversation, around what a dataset and data could be through feminist approaches. The work’s title itself is a play on the MNIST²² dataset, which is a standardized and normalized dataset that was/is a foundational benchmark for computer vision models. By F* the MNIST dataset we questioned what the standards of a dataset are and how we could renegotiate what normalizing this data could mean and how we contextualize these indexes of data so that they are not isolated and efficient images. To do this people inputting data were encouraged to be disruptive and unstandardized (whatever that meant to them), allowing space for people to add their own new indexes to the dataset (Fig. 5), as well as more information about themselves and the context of their input. This dataset is fluid and continually being added to and is formatted in linked open data (LOD) to create a complex web of data, relations and people in semantic triples that are open and linkable, which radically differ from the anonymous and isolated standards of the original MNIST.

The second work with IP is the re-reader, an outcome from an exercise imagining what roles AI can play within text analysis, communal reading, and research. We have been playing with a few different configurations, focusing on how highlighting and other preexisting affordances of digital reading can be made intelligent through small scale, local technologies that can help us renegotiate what we read. One of the most promising experiments from this work so far is an AI variational autoencoder²³ that is trained from a local and small dataset of texts, compressing this set of knowledge into a three-dimensional latent space. This segregational latent space, which I mentioned Chun’s critique of, takes on a new form of non-prescriptive and performative, enabling the person to retrain it and reform its ontology and splitting of space. Once trained, this latent encoding or ontology is then used to highlight a text sentence by sentence to the color (RGB) determined by the three-dimensional latent encoding of that sentence (Fig. 6). This compositing of fluid and nondeterministic ontologies enables us to navigate texts simultaneously through these different dimensions and dynamics, the ones we read, and the ones re-read by the trained model and interpreted by us. The model used here is also lightweight and quick to run and train locally on most consumable computers.

These are two early attempts to enable AI and RS to be configure-able to different methods and, through these actions, we have realised that making RS and AI analytics configure-able also means to have their concepts, imaginaries and matterings adaptable and accessible too. To do this we have started to entangle these often siloed approaches into a collaborative wiki/docs alongside the code. By doing this we are attempting to enable ourselves to form new dialogues, metaphors and matterings in the creation of these works which can move beyond the efficient ones of contemporary AI. Configure-able methods here means for these intersections of language and practice to be generative re-imaginings of infrastructures. These methods bring with them not so many answers, but create space for questions and conversation around how these systems are configured and what that means for us as people and communities living within their political/relational modeling, enabling us to rethink, rename, remetaphorize and revalue their processes.

From this essay, we have seen the need for this change within RS, but the same is true in many other intelligent and future imagined tech infrastructures, as they need a deep rethinking and reapproaching. This project and its understanding of configure-able methods is very much at the beginning of its journey, but I see hope in this methodology and the radical agendas it grows from, as well as the promising steps made so far. It is a promising starting point for us, Imaginary Practice, to be able to begin this essential questioning of how we can configure these (intelligent) sociotechnical infrastructures otherwise.