The recent expansion of artificial intelligence is usually measured in terms of capability: how much code a model can write, how much context it can process, how long it can sustain a task, or which problems it can solve while human intervention grows ever more distant.
Gemini 4 Argon, unveiled by Google on 30 September, explicitly extends that logic to prolonged professional work. Meanwhile, in Cape Town, an international university summit held in Africa for the first time placed artificial intelligence within a broader discussion of knowledge, justice and the future. From Brazilian Amazonia, Aparecida Vilaça raises a different difficulty: even the categories designed to recognise other forms of existence can transform them in the act of making them legible to institutions built on stable classifications.
These are three scenes far removed from one another. One belongs to the industrial frontier of AI, another to university politics and the third to Amazonian anthropology. Together they bring into view something less visible than the advance of the models. Every knowledge infrastructure also sets the conditions under which a form of intelligence can be recognised, used or translated.
When the response becomes the work
On 30 September Google unveiled Gemini 4 Argon, a frontier model geared towards software engineering, research and specialised writing, legal and financial work, and cybersecurity.[1] Its most significant feature does not necessarily lie in any particular benchmark, but in Google’s insistence on long-horizon workflows: processes made up of extended sequences of operations that require keeping track of context, reviewing results and taking intermediate decisions.
The company says that thousands of people across its teams already use Argon, and it describes internal uses that include large-scale migrations from C and C++ to Rust and memory optimisation in data centres. In one case, Argon-based agents analysed telemetry from Google’s own infrastructure, identified possible improvements and autonomously applied changes which, according to the company, freed up more than 300 TiB of memory.[1]
The external rollout is starting on a restricted basis. A select group of cybersecurity professionals has access through Google’s Fairwind Program, while the company evaluates the system ahead of a wider release and, by its own account, takes part in the voluntary process through which the US government gains access to models before launch. The company also states that, for this group and for its own internal teams, the model is provided without cybersecurity safeguards, so that they can make full use of its cyber-defence capabilities.[1]
Some of the tests cited by Google come from external organisations such as Vals, Harvey, Zapier and Gray Swan. Caution is nonetheless warranted. These are results selected and publicised by the company itself and, while access remains restricted, the wider community cannot yet reproduce the full set of claims about the model on any broad scale.
What is most interesting is the kind of relationship with the machine that these evaluations attempt to measure. For years, the dominant interface of generative AI has organised interaction around a relatively small unit. A person makes a request and receives a response. Systems designed to sustain tasks for much longer introduce a different architecture. Between the initial assignment and the final result lie hundreds or thousands of operations that can hardly be submitted for individual human approval.
The unit of delegation thus begins to shift from the response to the work. A person can review a paragraph before using it. It is far harder to reconstruct every decision taken during a migration of hundreds of thousands of lines of code, an extended piece of research or the continuous analysis of an infrastructure.
Autonomy may emerge less as an exceptional property of the model than as the result of combining duration, tools and the distance between intermediate decisions and human oversight.
Cape Town negotiates the terms of adoption
From 29 September to 1 October, the University of Cape Town hosted the Times Higher Education World Academic Summit 2026, the first edition of the summit to be held in Africa. The organisers reported more than 500 attendees, more than 100 speakers and more than 250 participating organisations.[2]
Its theme, Knowledge, justice and futures: Universities in the age of intelligence, placed AI within a conversation that also took in human and collective intelligence, inequality, academic cooperation, public trust and the future of the university. UCT had already said in July that African and Global South perspectives would be integrated structurally into the programme.[3] One of the summit’s strands further proposed that Global South universities should be able to lead partnerships founded on reciprocity and co-creation over extraction.[4]
The remarks of UCT’s vice-chancellor, Mosa Moshabela, bring that framing down to institutional ground. On the first day, 29 September, he explained that the university has decided to train all its students in AI and recalled that it had switched off automated AI-use detectors, a decision he called difficult and presented as a way of removing barriers to AI. He acknowledged the risks but urged people not to dwell on them, describing the technology as a partner in learning.[5] The decision remains an institutional one, even if its direction is to open up rather than to restrict.
Moshabela also linked the adoption of AI to inequality. Many tools developed by the industry are priced beyond the reach of a large share of the continent’s population, which is why he argued that the university should act as an institution capable of narrowing that gap.[5]
In 2025, UCT adopted an AI in Education Framework that Sukaina Walji, director of the Centre for Innovation in Learning and Teaching, described as the outcome of more than a year of consultation. From 1 October that year, the university stopped using Turnitin’s AI detection score because of its unreliability and the risk of false positives and negatives.[6][7]
The same process reveals a more material tension. UCT offered its students free access to some AI platforms through agreements with Microsoft and Google.[7] The institution seeking to define for itself how AI enters its educational practice depends at the same time on tools produced by the large companies that dominate that infrastructure.
There is no need to read this as a contradiction. It is a negotiation of dependence.
Institutional capacity lies precisely there: not in building every model or platform from scratch, but in keeping room to decide what is adopted, how it is evaluated, which practices should change and under what conditions these technologies acquire authority within learning and research. An African university thus appears not merely as a recipient of technologies created elsewhere: it also takes part in defining their terms.
When a stable category tries to house a relational world
On 30 September, Aparecida Vilaça, Professor of Social Anthropology at the Museu Nacional of the Universidade Federal do Rio de Janeiro, delivered the Henry Myers Lecture of the Royal Anthropological Institute (RAI) at King’s College London, with support from the King’s Brazil Institute.[8]
According to the abstract published by the RAI, the lecture — Conflicting perspectives on other-than-human beings: Amazonian predatory animism encounters the ‘Rights of Nature’ — began from a critique: in recent decades, a kind of ‘purification’ has stripped Amazonian animism of its predatory aspects, recasting it as a benign, cooperative animism that sits comfortably with the holistic models of many environmental movements and, above all, with rights of nature initiatives. Vilaça set out to discuss the equivocations this eclipse of predation produces in dealings with environmental actors and lawmakers.[8]
Vilaça has worked with the Wari’ people of Rondônia since 1986. The subject builds on research she published this year in the Journal of the Royal Anthropological Institute on the Laje River — Komi Memem to the Wari’ — which was legally recognised in 2023 as a living entity with rights.[9][10] It was the first river in Brazilian Amazonia to receive such recognition.[9]
The case might look like a particularly successful meeting between environmental law and Indigenous ontologies. Vilaça’s analysis reveals a more complex situation.
Her argument draws on the notion of equivocation developed by Eduardo Viveiros de Castro, an anthropologist at the Museu Nacional: different parties can use the same word or category without speaking from the same relational world. In the Laje case, laws need classifications stable enough to determine which entity holds rights, who may represent it and what obligations follow from that recognition. Wari’ perspectivism, by contrast, works with positions that can alternate according to the relationship and the circumstances.[9]
Vilaça recounts a conversation with Francisco OroWaram, a Wari’ councillor elected in Guajará-Mirim in 2020 and the author of the bill that recognised the river. In 2019 he also became the first Wari’ to complete a master’s degree, in Geography, at the Federal University of Rondônia. When she asked him directly whether the Laje was wari’ — a person — he said it was. The conversation also turned to a traditional eschatology according to which those who have died live at the bottom of the river.[9]
The paradox matters. According to Vilaça, it was not traditional cosmology that inspired the law; the influence ran the other way. By treating the river as a person, the law made room for — and even called for — the return of Wari’ eschatology, which now coexists with Christian eschatology and is encompassed by it. Francisco, like many Wari’, is an evangelical Christian.[9] Institutional translation thus produced something more ambiguous than a direct equivalence.
The material outcome deserves attention too. During Vilaça’s most recent fieldwork, in February 2025, the law’s practical effects were still limited. Soya farming near the river continued, and the pollution was still causing white foam and skin irritation, especially among children. Francisco had received threats after pushing for action against crop-spraying and gave up his bid for re-election. The 2024 elections returned no other Indigenous councillor to the municipal council.[9]
Legal recognition, then, does not automatically amount to an effective capacity to protect the river.
The connection with artificial intelligence is not part of Vilaça’s argument and should not be attributed to her. Her work nonetheless offers a useful tool for thinking about contemporary infrastructures of classification.
A language becomes tokens. A museum collection becomes metadata. An image becomes vectors. A scientific practice becomes an executable procedure. Each operation makes it possible to relate and process knowledge on a previously impossible scale, but each also needs to stabilise what it represents.
The problem arises when an infrastructure built on computable categories tries to house entities whose position depends precisely on shifting relationships.
The question is then no longer only how much can be translated. It is also which forms of difference can be preserved once they have been converted into something the system knows how to represent.
Three ways of drawing a boundary
Gemini 4 Argon widens the space in which a machine can act without immediate supervision. Cape Town is a reminder that incorporating that capacity remains an institutional decision. Vilaça points to an earlier problem: to enter an infrastructure, something must first be made recognisable to its categories.
Technology, institution and ontology thus appear as three layers of a single infrastructure.
A technology’s capacity does not depend solely on what it can execute. It also depends on what its architecture manages to represent.
And some differences begin to be lost before the machine has even acted.
References
[1] Kavukcuoglu, Koray. ‘Gemini 4 Argon: our next era of frontier intelligence’. Google blog, 30 September 2026.
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
[2] University of Cape Town. ‘THE World Academic Summit 2026’.
https://uct.ac.za/research/the-world-academic-summit
[3] University of Cape Town. ‘World Academic Summit 2026: An update’. 30 July 2026.
https://www.news.uct.ac.za/article/-2026-07-30-world-academic-summit-2026-an-update
[4] Times Higher Education. ‘World Academic Summit 2026 — Agenda themes’.
https://www.timeshighered-events.com/world-academic-summit-2026/page/5240265/agenda-themes
[5] University of Cape Town. ‘VC shares insights on UCT’s adoption of AI’. 1 October 2026.
https://www.news.uct.ac.za/article/-2026-10-01-vc-shares-insights-on-ucts-adoption-of-ai
[6] University of Cape Town. ‘Introducing UCT’s AI in Education Framework’. 15 July 2025.
https://www.news.uct.ac.za/article/-2025-07-15-introducing-ucts-ai-in-education-framework
[7] Mashinini, Gugulethu. ‘UCT scraps flawed AI detectors’. Sunday Times; republished by UCT News on 24 July 2025.
https://www.news.uct.ac.za/article/-2025-07-24-uct-scraps-flawed-ai-detectors
[8] Royal Anthropological Institute. ‘Henry Myers Lecture 2026 — Aparecida Vilaça’. 30 September 2026.
https://therai.org.uk/events/myers-lecture-2026/
[9] Vilaça, Aparecida. ‘The birth of an earth being: “Rights of nature” in Brazilian Amazonia and elsewhere’. Journal of the Royal Anthropological Institute, 32(3), 833–851. Published online 18 March 2026.
https://doi.org/10.1111/1467-9655.70087
[10] Bacellar, Clarissa. ‘Rio Laje, em Rondônia, é o primeiro a ter direitos reconhecidos por lei’. Portal Amazônia, 24 June 2023. In Portuguese.
https://portalamazonia.com/meio-ambiente/rio-de-rondonia-e-o-primeiro-a-ter-direitos-reconhecidos-por-lei-entenda/