If we were to transpose the famous ancient Asian parable of “The Blind Men and an Elephant” – in which several blind men touch different parts of an elephant and identify it as different animals individually – unto generative AI, then the student will find it to be a homework machine, the technologist might find a hype hope for a future Artificial General Intelligence (AGI)-inflected, Open-AI branded world, your uncle might find a customized fake news generator fit to tell him what he electorally wants to hear, and the economist would find a magical solution to productivity ills he has never been able to (and likely never will) cure. And here, in this corner, we may reach the school of humanistic and social scientific scholars huddled together, yet still touching different parts of the elephant: the optimist will find posthumanist promises, now finally coming to bear, after having been promised ages ago; the educator will look at his/her university memos and sees an academy destroying machine; and the critical scholar will find intellectual property theft, a massive economic extraction from the underpaid global underclass, and a massive infrastructural extraction of water and electricity that contributes to a rapidly warming world (Dhaliwal, Reference Dhaliwal2023; Offert and Dhaliwal, Reference Offert and Dhaliwal2024). Now, all of these folks are right in one way or another (except the economist, but that is to be expected, since his whole historical existence is based on being wrong [MacKenzie, Reference MacKenzie2008; Morgan, Reference Morgan2012; Sweezy, Reference Sweezy1970]), but my starting point for this position piece is something that allows this parable to work as one: the fact that the entity at the center is an elephant, an animal of such immense size that it necessarily defies the human scale of perception and consequently generates, for the purposes of the parable’s narrative structure, these conundrums of subjective knowledge that Hindu, Buddhist and Jain mythologies have found useful over two millennia ago. What happens if we take seriously the scale question at the heart of our differential diagnoses of contemporary artificial intelligence?Footnote ¹

Amidst the admirable work done by the nascent field of critical AI studies (Goodlad, Reference Goodlad2023; Raley and Rhee, Reference Raley and Rhee2023), many scholars have pointed out the extensive environmental toll of this recent scalar expansion – in research, deployment and usage – in generative AI (Anson et al. Reference April, Ballestero and Chahim2022; Bashir et al. Reference Bashir, Donti, Cuff, Sroka, Marija, Vivienne, Christina and Elsa2024; Bender et al. Reference Bender, Gebru, McMillan-Major and Shmitchell2021; Dobbe and Whittaker, Reference Dobbe and Whittaker2019; Hao, Reference Hao2024; Hogan, Reference Hogan2024; Pasek et al. Reference Pasek, Lin, Cooper and Kinder2023). It is now clear that staggering amounts of electricity and water are needed to power the big data centers that process the petabytes of information scraped from the internet for the calculation of weights that drive these foundation models.Footnote ² These models are designed to be broad spectrum; as several experts studying the economic hype around these models have shown, it would not be unfair to say that, while impressive in parts, these foundation models may be solutions searching for problems (LaGrandeur, Reference LaGrandeur2024; Merchant, Reference Merchant2024; Vinsel, Reference Vinsel2021). Casting a wide net here with the sheer scale of foundational models is not only a way for the tech industry to invite use cases and possible directions from all of us (who are now turned into, whether we want or not, the beta testers for such models) but also can be considered as a way of deferring into the future the question of what these models will do (instead of can do today). This is also evident from the fact that the business models behind these foundation models have still not proven themselves to be robust and sustainable at scaleFootnote ³; OpenAI, the industry leader, for example, despite its massive coffers and billions in revenues, loses more money with each passing week (Gray Widder et al. Reference Widder, Whittaker and West2023). It must be noted that this problem of scale – which is publicly only identified by the numerical metrics that denote the number of parameters of a model be it 3 billion or 671, not unlike the traditional question of an archive’s scale (Daston, Reference Daston2017) – is not merely an environmental or business sustainability concern but also an issue regarding ascendant tech monopoly powers (owing to the increasingly absurd costs of training bigger and bigger models [Hosanagar and Krishnan, Reference Hosanagar and Krishnan2024]), and because these tech monopolies are geographically concentrated, now further a broader global geopolitical (and also data sovereignty and privacy) issue as bigger models start becoming out of reach for most researchers outside the United States of America (Bratton, Reference Bratton2015; Gillespie, Reference Gillespie2020; Miller, Reference Miller2022); consider, for example, the recent trade policy decisions by America which prevent academic/industrial researchers (and even consumers) in most parts of the world from accessing the most advanced integrated circuits which are usually needed to train the state-of-the-art foundation models (Farine, Reference Farine2025).

If these ubiquitous large foundation models are synonymous, in public discourse, with AI as a whole (Dhaliwal et al. Reference Dhaliwal, Lepage-Richer and Suchman2024; Schmidt, Reference Schmidt2023; Suchman, Reference Suchman2023), then we must consider the very foundations of this equivalence. If we are but beta testers in an unfolding semi-solid global rollout of potentialities, what does that say about our limited agency in the realm of machinic and social collectivities? Unpacking foundation models’ scalar dimensions may help answer that question.

To recap, not only are foundational models unimaginably large, but they are also built on a non-consensual extraction of the labor of everyone in the information economy and are also somehow environmentally catastrophic. If you have been following along any part of critical AI studies, all of this would mark the baseline knowledge of the world today.Footnote ⁴ But the part that I want to focus on here is the epistemological correlate of that one, one that has to do with the fact that foundation AI models do not have an identity of their own. These are big archival projects – and I am using the term archive here totally in the Derridean, archive studies sense, where an archive anticipates yet unknown applications – that are not merely storing but actively transforming, processing, their contents so that one day some unknown application could make use of the transformation (Jo and Gebru, Reference Jo and Gebru2020; Plantin, Reference Plantin2021; Thylstrup, Reference Thylstrup2022; Thylstrup et al. Reference Thylstrup, Agostinho, Ring, D’Ignazio and Veel2021). And the industry has its fingers crossed really, really hard that the transformation is financially viable. However, we should recall that large models are in fact a small subset of what AI is, has been, and can be, and this piece calls into question our relationships with the scale and size of AI today and tomorrow.

In what follows, I shall here outline a couple of epistemological propositions (or ways of considering) regarding how we may contextualize these scalar conundrums; consider these to be yet more (hopefully interesting) additions to the “what animal is elephant” problem, additions that I hope lead to specific provocations on what is to be intellectually and experimentally done:

1. 1. My first proposition here follows simply the assumption that foundation models can be understood better as objects that invite wonder and thus tie more neatly into our history of fascination than our history of tools and implements. That the foundation models are solutions still searching for a problem to solve is not unrelated with the conceptual framework of techno-solutionism – the idea that social problems can be solved technologically, where the instrumentality of technology is front and center (Kneese, Reference Kneese2023; Richterich, Reference Richterich2024) – here, then, is a fundamental issue plaguing what our models can do. But if the solutionist principle is forcibly excised from the matter, these models come to be seen as material wonders, made of the matter that we are surrounded by, but nevertheless alien in how they appear to us (Amoore et al. Reference Amoore, Campolo, Jacobsen and Rella2024; Hayles, Reference Hayles2022). This sense of familiar alienation – and I am here inspired by Katherine Park, Lorraine Daston, Mary Baine Campbell and Caroline Walker Bynum’s landmark studies of wonders, miracles and premodern epistemologies (Bynum, Reference Bynum2020; Reference Bynum2015; Campbell, Reference Campbell2011; Daston and Park, Reference Daston and Park2012) – that the models can invoke in mundane interactionsFootnote ⁵ are not only moments of inquiry and investigation but also indicate a moment of pause, a moment of reconsideration of the inter-activity at hand. Treating these objects as material wonders does not mean becoming overwhelmed by their power; even miracles and Christian relics in the late Byzantine era, as Bynum reminds us, were often understood within logical systems of wonder and not always an overpowering theological force. In fact, understanding them as systems that produce surprise in one form or another may help us figure out what surprises us and why.Footnote ⁶ Nothing can be more closely analytical than reading for surprise in mundanity, adopting a methodological stance of what, why and how is our fascination piqued. That is the first step towards cultivating interest that does not take the big, grand sublime as the only wonder possible. If not closely looked at, surprise generation can often be taken as granted to be a feature of “bigness” (Rheinberger, Reference Rheinberger2011).

2. 2. My second proposition here is that any instrumentality we need/want from our models (however experimental or standard it might be) does not require any foundationality per se. In other words, if foundation models are synonymous with large models, be it large language models or large diffusion models, or now multimodal systems of scale, then, I am arguing that anything that we may need the models to “do” (if we need in the first place) does not require systems of this magnitude; this is not only true for humanities and social scientific applicationsFootnote ⁷ but also borne out by emerging work in computer science at large (Chen et al. Reference Chen, Zhang, Zhuang, Tang, Liu, Wang and Xu2024; Menghani, Reference Menghani2023; Rather et al. Reference Rather, Kumar and Gandomi2024; Wu et al., Reference Wu, Acun, Raghavendra and Hazelwood2024).

A quick look at the history of artificial intelligence shows us that connectionist paradigms – which can be understood as earlier versions of neural network like approaches today – have been conventionally plagued by improper scaling for over half a century now. When artificial neural network (ANN) paradigms were first introduced in the 1960s, their utility and efficiency was severely doubted by the field, and as a result, such paradigms “lost” to the proponents of symbolic AI (think step-by-step predetermined expert systems). And this loss was later recast as a problem of scale; the claim here being that technoscientists did not have enough data or processing power when neural networks were initially proposed. In other words, within the history of technology, it is widely understood that neural networks have always scaled unpredictably (Dhaliwal, Reference Dhaliwal2023; Dhaliwal et al. Reference Dhaliwal, Lepage-Richer and Suchman2024; Wiggins and Jones, Reference Wiggins and Jones2023; Jones, Reference Jones2023; Mendon-Plasek, Reference Mendon-Plasek, Roberge and Castelle2021); most in the 1970s did not predict that powerful machines with a lot of data one day would have ANNs “defeating” the symbolist approaches, just like most experts in 2008 did not realize that something as powerful as ChatGPT could come along only if we scraped the whole internet and tokenized/transformerized it successfully (Daston, Reference Daston2022; Jones, Reference Jones2023; Pasquinelli, Reference Pasquinelli2023; Plasek, Reference Plasek2016). These two particular momentsFootnote ⁸ are the extrapolatory bedrock of the expansive expansionist approach taken by the tech industry today, the presumption here being that every moment of previous disappointment needed to be followed by scaling more, and as a result, we need to keep on scaling more intensely today, despite the burgeoning evidence to the contrary. It is hardly a universally smart business decision, but since the start of the growth of global productive overcapacity in the 1970s, the economic surplus laying around searching for productive explosions has only taken wi(l)der and wilder speculative bets. Many bets never paid off, but ChatGPT, albeit with major legal IP loopholes underneath it (at least temporarily) did (Benanav, Reference Benanav2019a; Reference Benanav2019b; Koutras and Selvadurai, Reference Koutras and Selvadurai2024; Mims, Reference Mims2023). So, the production of bigger models – precisely what is targeted by the critical scholarship on environment, economy and technology mentioned above – comes from several such parallel bets playing out in real time. Returning to my proposition, I see no reason why we must participate in those bets, for we do have much to lose. This is also precisely what extant conceptual (and empirical) critiques of scale and scalability teach us (Hanna and Park, Reference Hanna and Park2020; Lee and Ribes, Reference Lee and Ribes2025; Ribes, Reference Ribes2014; tante, 2025; Tsing, Reference Tsing2012; Lempert Reference Lempert2024; Widder and Kim, Reference Widder and Kim2025).

Recent research in computer science has shown signs that larger models may not only be prohibitively costlier but also, in some cases, less likely to sustain their veracity and performance levels in general (Caballero et al. Reference Caballero, Gupta, Rish and Krueger2022; Kaplan et al. Reference Kaplan, McCandlish, Henighan, Brown, Chess, Child, Gray, Radford, Wu and Amodei2020; McKenzie et al. Reference McKenzie, Lyzhov, Pieler, Parrish, Mueller, Prabhu, McLean, Kirtland, Ross, Liu, Gritsevskiy, Wurgaft, Kauffman, Recchia, Liu, Cavanagh, Weiss, Huang, Floating Droid, Tseng, Korbak, Shen, Zhang, Zhou, Kim, Bowman and Perez2023; Zhou et al. Reference Zhou, Schellaert, Martínez-Plumed, Moros-Daval, Ferri and Hernández-Orallo2024). It is imperative then, if we have to (or must) live with AI – and it is a big if, for I am ambivalent about our possibilities – that one thing we must do is pick the scale ourselves, carefully, taking full cognizance of our agency as researchers. Small-scale models do very well on many regular tasks; most people do not need the unbounded and imprecise horizon of something as massive as ChatGPT. Smaller, more custom models, especially ones that come pretrained with weights, do not actually present anywhere near the same kind of ecological challenge that the constant expansion done by Microsoft, Oracle, Google, Meta, Anthropic, and OpenAI does; most of the ecological cost of AI comes from training a model and not running it. (And that is not to mention the data privacy cost of having local vs cloud-cartelized models.) So, for the vast variety of tasks, scales much smaller – and consequently, though not always, more sustainable – than the ones wanted by the likes of ChatGPT should/could/would suffice. Plus, given our funding situations and threats globally and the existing archival specificities in our portfolios, rigorous research and custom development of smaller models can actually be done at the cutting edge and unlikely to follow the hype cycles of an extractivist tech industry (which is the only one standing to gain if the models get bigger).

This brings me finally to an interrelated set of design provocations. As much as I have just critiqued the state of solutions searching for a problem, I do so only because of the social relations engendered by ChatGPT as a product, and not because I consider experimentation and failure to be inherently bad, or that the archival deferral is irredeemably problematic (Daston, Reference Daston2017; Hong, Reference Hong2020; Strasser and Edwards, Reference Strasser and Edwards2017). In fact, drawing upon theories of generative failure (Heller, Reference Heller2008; Marres and McGoey, Reference Marres and McGoey2012), in conclusion, I wish to suggest a few possible undercooked directions for humanistic research and exploration with smaller models.

Pardon my polemics and self-cringe but here is how I see the possible scalar queerings of foundational models.Footnote ⁹ On one hand, we may note a scalar engagement actively affecting the epistemic techniques of service and assistance. As several scholars have already noted, service as a cultural technique has been widespread in our ideas about technology; this is why the servers serve us (Canales and Krajewski, Reference Canales and Krajewski2012; Dhaliwal, Reference Dhaliwal2022; Krajewski, Reference Krajewski and Iurascu2018). But not only are there social and agential problems with the idea that our computers must, and exist to, serve us, there is also frankly a lack of interestingness in that notion. The human replacement problem – the idea that AI will or does behave and replace a human, a genealogy older than Turing test (Offert and Dhaliwal, Reference Offert and Dhaliwal2024) – is only a problem if “service” is at stake. But if other human modalities, be they creativity or aesthetics or phenomenological experience (Denson, Reference Denson2023), are to be explored intra-actively (Barad, Reference Barad2007) then a wider range of what can be done could perhaps open up. On the other hand, scalar shrinkage of these models also perhaps offers us capacities for weirder, queerer, contestations.Footnote ¹⁰ This includes the ability of – rather, the agency of – different communities and groups and peoples to make their own explorations, with paradoxically both more and less efficient models, or better yet, with efficiencies construed as per the assemblage at hand. So, in other words, I am suggesting models less good at generalized tasks but better at specific or idiosyncratic ones. I don’t think our computers should be more like us, but even if that was the case, what could be more like us than bad at most things and good at only one (if that)?

To sum, scale, technically speaking, is often taken by different discourses as a growth framework.Footnote ¹¹ Just as conversations around degrowth proliferate out of global material necessity (Saitō, Reference Saitō2022; Reference Saitō2024; Schmelzer et al. Reference Schmelzer, Vetter and Vansintjan2022; Vettese and Pendergrass, Reference Vettese and Pendergrass2022), I want to speculatively suggest that degrowing foundational models might be far more interesting, generative (or even the bare minimum, viable and possible) in our attempting to co/in-habit our human–machinic futures, or to return to the figure I began with, live with an elephant in our midst (Haraway, Reference Haraway2012).