What We Talk About When We Talk About AI

Last month, GPU manufacturer NVIDIA’s share prices tumbled by nine-point-five per cent, representing the greatest loss in market value in one day for any company ever at a loss of 279 billion dollars ¹. The GPU is the primary technology that powers AI, and harnessed en masse allows for inference to be conducted at industrial scale. While the share price has since corrected, this kind of blow to the world’s most successful GPU manufacturer could be understood as a moment of investor hesitation around an AI bubble, as well as a diagnosis of the technology’s cultural status. The fact is that – technically, aesthetically and existentially – AI is in a very strange place right now. The popular transformer and diffusion architectures have become wildly commercially successful, allowing for large language models (LLMs) and generative AI to be received as bona-fide technological paradigm shifts. However, like crypto, NFTs and other e/acc (effective-accelerationism) backed hype-trains that have been and gone, AI has also widely become cringe-ified, with a growing consensus between normies and creatives that the technology is morally bankrupt, derivative, aesthetically repugnant, copyright infringing, and will be used to automate the work of and ultimately displace workers.

Falling share prices and a growing sense of resentment towards AI does not give us the entire context: materially, the technology is in a little bit of trouble. Despite the transformer architecture’s coming of age through large language models, Open AI is in an arms race with Anthropic and Google, spending hundreds of millions on training their LLMs for marginal performance increases ². Meanwhile, Gen AI has supposedly begun to ouroborically consume itself as AI-generated content is fed back into the model, deteriorating its performance ³. In the music world, entire Bladee songs show up verbatim as training data in generative music models, as record label litigation says ‘sic’ em boys!’ ⁴, whilst Elevenlabs AI slop videos infest every reels, shorts, and viral video platform (I am quite fond of these videos, as I’m sure you are too). 

At this point, 'Artificial Intelligence' has become an umbrella term that captures not only a certain set of technologies but also a set of ideologies and their opposites. On Elon Musk’s X, the celebrities of AI, e/acc, ea, and other awkward acronyms publicly spat with one another around a dream of Artificial General Intelligence (AGI). At one pole, safety polemicists like Eliezer Yudokowsky urge against the danger of inducing a self replicating AI apocalypse ⁵, while at the other, Meta’s chief AI scientist Yann Lecun cautions that regulatory capture will stunt the development of newer, more sophisticated learning architectures ⁶. These narratives around AI and its ethics are firmly situated in the present toward future tense, where the current generation of technologies are considered kernels with the potential to metamorphose into radically life changing technologies – for better or for worse. The incredible progress of generative AI and LLMs in creative media in particular, is seen by both proponents and detractors to have become the technological battleground that supposedly portends our technological, as well as our global future.

In this discourse, the creative industries are uniquely positioned threatened by the burgeoning success of AI —  a particularly vicious canary in the coalmine, and at the coal-face. Simply put, as creators we intimately understand how the material landscape of our craft functions, and the ways in which our work will be changed. Further, as artists inheriting the new technologies that simultaneously threaten to make us obsolete, we have a tendency to (sometimes confusingly) engage self-reflexively and critically with our techne. Curiously, musicians seem to be at the forefront of this conversation, allowing select musicians such as Grimes and Holly Herndon to take on a kind of artist-scholar role amongst the discourse, appreciably impacting public perception of the tech.

Beginning in 2017 on Eora/Darug land, Sydney, SOFT CENTRE’s debut festival in Naarm/Melbourne was a three-day event under the theme of  SUPERMODEL, including a Discourse day on 30 August 2024. Set at Trades Hall’s own Solidarity Hall, the day was set for a series of lectures, panels, and short films by artists, musicians, writers and other thinkers such as Rowan Savage, James Parker, Emile Frankel, and Jennifer Walshe. Heraldic trade union tapestries framed the SUPERMODEL stage, backdropped by 2 dual (Chloe Newberry) biomechanical sculptures – a contrast that aptly highlighted the underlying theme of technology vis a vis labour. While it was (thankfully) not mandatory to speak on AI, there was at the very least a persistent narrative of technology’s ‘speculative future’ threaded through most of the works and presentations. And of course, many of them did directly speak to, or incorporate AI – thematically, literally or both. The festival’s subtitle obviously alludes to this: in a Mixmag feature on the festival, SOFT CENTRE’s co-director Thorsten Hertog stated: 'SUPERMODEL is a bit of a tongue in cheek word because it's very evocative. Like initially it brings to mind, you know, these impossible angel bodies walking down runways. But also, for me, it brings to mind this hyper object, or like an AI singularity moment. Something severe and mega and beyond comprehension'⁷.

The plurality of interpretations around the speculative essence of AI – its  'hyper-object'-ness – was on display at Discourse. In particular, talks by Rowan Savage, James Parker, Emile Frankel, Jennifer Walshe, and Deforrest Brown Jr. (partially) considered the cultural, creative, political, and ethical dimensions of AI in sound and music. Each talk came with its own ontology of AI – of what AI is: postcolonial creative tool, listening apparatus, destroyer of language, anything and everything, technological delusion. In an article for Unsound 2023, Jennifer Walshe suggested that AI should be thought of '...from multiple positions, simultaneously. Messily. Not one way of looking at AI, but many' ⁸. Here, this discourse of plurality also has its downsides – not because there is one prevailing meta-narrative around AI, but instead because there is a certain type of discursive abstraction occurring here, one that allows for a rhetoric that begins to mirror the speech of silicon valley C-suite executives. To the executives and marketing teams at Open AI, Google, Microsoft, Meta and Anthropic, AI too is multifaceted, pluralistic, and polyvalent – we just have our moral compasses flipped.

At present, the essence of AI is overhyped, overvalued, and deeply technically rickety – so why are we attributing this kind of power to the technology? Again, as per Walshe 'AI [is] a meme… a scam'. But it is not just an ideological fabrication of late-stage capitalism: AI simply doesn’t work as advertised. Despite the advertised power of large language models, neural networks and the transformer architecture are not able to capture syntactic meaning. And although the raging debate between a certain type of Xitter schizoid in the replies of celebrity machine learning engineers (see below) about whether ChatGPT is actually conscious, the reality is that no LLM at present is able to tell you how many letters are in the word strawberry (It’s three. Unless…?)⁹. This is one of many examples, and is not an error or ‘glitch’ in the technology, but more a statistical reality of modelling data: if a model presents a certain type of world-view, then this world view is ultimately probabilistically fallible. On the other hand, AI (and its speculative future) is repeatedly described as a probabilistically infallible and infinite model. If AI can eventually be anything, it is everything – the Landian genre of capital that promises an infinite scaling ad nauseam; you can’t overfit a model if it has been trained on all possible data. 

If the AI we talk about does not accurately capture the shaky materiality of the tech, then what are we talking about? It is clear that the discourse is centred about a sort-of novel technological configuration of epistemic knowledge, of (Jeremy Fragrance voice) Power, of a Golden Calf, of a Big Other (who itself is an infinite set of all elements) to rage against. The discursive reality of AI is quite clearly comprehensible: we are not talking about boring limp-dicked linear algebra of the present, but the virile potency of the self-learning, infinite fantasy of Artificial General Intelligence! This is not to say that AI in its current form does not bring upon calamity and woe upon the world: the terrible impact on the planet, awful treatment of data labellers and normalisation of surveillance technology are a series of blights. But really – how is this different from the rearticulation of the pernicious nature of tech throughout the twentieth and twenty-first centuries? The Algorithm, Platform Capitalism, The Stack, Surveillance Capitalism, The Cloud, and many such other Verso Books titles could all be substituted in place of Artificial Intelligence.

Musicians in particular, are well aware of the fetish of technology, that which seems to impart something beyond the music: The crackle of the record, the sound of ‘that’ pedal, of ‘that’ reverb unit, the sound of ‘that’ tone. Unlike the fantasy found in the AI of writing, the fantasy of AI in music has more often revolved around the unheard, the speculative, the ability to create beyond oneself. The sound of ‘that’ future: create any sound, model any instrument. Although AI can represent '...new, un-dreamt-of forms of art, genres which no one has ever heard, joyous and life-affirming ways of making art and music which are beyond our comprehension, just over the horizon of the vibe shift'¹⁰, there has not yet been a successful 'one-shot' architecture that doesn’t rely on at least some training data. With our current technology, the results will always exist in the sample data in some way. Maybe after many iterations, convolved, diffused, cross-bred and mutated, we will see a semblance of the new, but there will likely never be a technological rupture that is able to bury the archeological trace of that which came before.

One of the great beauties of music is really that it persists on a fantasy of an infinitely growing technology in the first place. Despite the oft quoted aphorism that goes something like 'there is no such thing as original music', we engage with instruments, writing, and the body in a way that suggests otherwise. Our use of AI is no different. The novel part of AI (in music and elsewhere), is that trying to decipher the mediation of the technology is extraordinarily difficult. Compared to other tools, the most popular learning models are infamous black-boxes that obscure the historicity of their technical composition. These huge, industrial scale models are trained on billions of features, and have many more resultant coefficients, and embeddings, but we cannot easily assess their contribution to the model’s output. That is, we don’t often know what the machine is ‘learning’. To the model, these are anonymous and probabilistic, and so we turn to the social messaging surrounding AI, and to mythology and fantasy.

The discourse around AI and generativity is currently stuck in a gear of talking about these fantasies, fetishes and projections. While this allows us to propose artificial intelligence in the abstract, as well as giving a diagnosis for the social construction of AI, we don’t yet have critical language around the technical-material and musical-material processes that AI models tend to mediate. We may yet require a more thorough effort in discovering how these models and architectures materially construct our art and music. Historically, music scholars have been quite good at this: we have media archeologies of different instruments, of various formats, of sound and noise itself. But what of the so-called AI? Do we know what the history of AI bears on the music we generate today, and do we understand its constituent parts? I think that there is good work being done: this is not intended as a Sokal affair ¹¹ type accusation, or to say 'don’t wear thrasher if you don’t skate' – but in my fantasy, amongst the messy pluralism, there is a middle path.

Asher Elazary is a musician and composer from Naarm. He writes and records music under ‘r hunter'.

Edited by Ella Howells as part of the un Extended Editor-in-Residence program. Supported by City of Yarra, Creative Victoria, Creative Australia and City of Melbourne.

1. [1] Noel Randewich and Suzanne McGee, Nvidia suffers record $279 billion loss in market value as Wall St drops, Reuters, September 4 2024.
2. [2] Katharina Buchholz, The Extreme Cost Of Training AI Models, Forbes, August 26 2024.
3. [3] Aatish Bhatia, When A.I.’s Output Is a Threat to A.I. Itself, The New York Times, August 25 2024.
4. [4] UNITED STATES DISTRICT COURT FOR THE SOUTHERN DISTRICT OF NEW YORK, Complaint against UNCHARTED LABS, INC., d/b/a Udio.com, and JOHN DOES 1-10, Defendant, filed June 24 2024.
5. [5] Eliezer Yudkowsky, profile on X.com. Accessed September 10 2024.
6. [6] Yann Lecun, Yann LeCun On How An Open Source Approach Could Shape AI, TIME Magazine, February 7 2024.
7. [7] Jack Colquhohn, Soft Centre's 'SUPERMODEL': Humanity in the time of AI, Mixmag, August 21 2024.
8. [8] Jennifer Walshe, Unsound Dispatch: 13 Ways of Looking at AI, Art & Music, Unsound Festival Substack, December 16 2023.
9. [9] [source/pic]
10. [10] Ed Dead Redemption, I am excited to reveal the incredible power of OpenAI's new "Strawberry" model (known as "o1"). This technology is the future, status on Bluesky, September 13 2024.
11. [11] Walshe, 2023.
12. [12] Ian Reilly, Public Deception as Ideological and Institutional Critique: On the Limits and Possibilities of Academic Hoaxing, Canadian Journal of Communication, Volume 45 Issue 2, July 2020, pp. 265-285.