This week, the question of where artificial intelligence lives has played out in three very concrete settings. In the main auditorium of the Spanish National Research Council (CSIC) in Madrid, the active inference community is discussing how to model brains and societies that learn. In Barcelona, the AI upgrade of MareNostrum 5 has been left without an inauguration ceremony because of the electoral period. And in an experiment by Anthropic, fifteen unbranded TV boxes were enough to run a model with 26 billion parameters.

These are three very different scales of computing. All three raise the same question of who decides where the capacity to think with machines is installed and who gets to use it.

Madrid, the mind as prediction

The seventh International Workshop on Active Inference (IWAI 2026) closes on Friday 16 October in the auditorium of the CSIC headquarters on Calle Serrano, after three days of sessions [1]. It is the main international meeting of a school of thought that understands perception, learning and action as a continuous process of prediction. Within this framework, organisms build probabilistic models of their environment and act to reduce the gap between what they expect and what they encounter.

This year’s edition is devoted to foundations, and the organisers explicitly link it to the city where Santiago Ramón y Cajal lived when he received the Nobel Prize [1]. Pablo Lanillos, of the Cajal Neuroscience Center (CSIC), chairs the organising committee. The invited speakers include Karl Friston (University College London), Carme Torras (CSIC), Rajesh Rao (University of Washington), Giovanni Pezzulo (ISTC-CNR, Italy) and Tadahiro Taniguchi (Graduate School of Informatics, Kyoto University) [1].

Taniguchi’s contribution is the one that speaks most directly to anthropology. His collective predictive coding hypothesis proposes that symbol systems, language included, emerge when many agents, each with their own bodily experience of the environment, exchange names and accept or reject one another’s signs. Society as a whole would then operate as a process of decentralised Bayesian inference [2]. Taniguchi uses this idea to explain why large language models appear to possess knowledge about the world without bodies or sensory organs, since the language they are trained on would condense that collectively produced knowledge. Earlier, the anthropologist Samuel Veissière and the psychiatrist Laurence Kirmayer, both at McGill University, had taken this framework into the terrain of enculturation, understood as learning what other people expect through shared routines [3].

Two details help to place the meeting. At the regular rate, registration ranges from €275 for students to €1,300 for industry, excluding VAT, with a fee-waiver scheme expected to cover two or three applications. Sponsors include Waymo, Motor AI and PAL Robotics [1]. Anyone wishing to take a look at no cost can watch Arun Niranjan’s introductory tutorial on YouTube, recorded in September [4].

Barcelona, an inauguration that does not take place

The inauguration of the AI upgrade of MareNostrum 5 was scheduled for Tuesday 13 October at the Barcelona Supercomputing Center (BSC), with the announced attendance of Jensen Huang, chief executive of Nvidia. The BSC called it off in a press release explaining that, because of the electoral period and the rules that apply to public administrations and publicly funded institutions, the event could not be held as originally planned [5]. The electoral period has been running since 6 October, when the Official State Gazette (BOE) published Royal Decree 806/2026 dissolving the Cortes Generales and calling a general election for 29 November [6].

The relevant rule is Article 50 of the LOREG, Spain’s general electoral law. According to the interpretation of the Central Electoral Board, inauguration events for public works or services are prohibited during the electoral period, whatever name they are given, and those works or services may still come into operation within that same period [7]. The law removes the ceremony and leaves the start of operations outside the ban.

The upgrade was awarded in January by EuroHPC to a consortium made up of Fsas Technologies, a Fujitsu subsidiary, and Telefónica. It has a budget of around €129 million, funded in equal parts by EuroHPC and by Spain, Portugal and Turkey. It adds one partition for training language models and another for inference, with Nvidia GB300 and GB200 GPUs connected via NVLink, and the Spanish government presented it as a way to open up AI supercomputing to SMEs, start-ups, public administrations and the scientific community [8] [9].

Part of the equipment is being installed in the former chapel of Torre Girona, which housed the first versions of MareNostrum [8] [10]. At the start-up of MareNostrum 5 in December 2023, Mateo Valero, director of the BSC, announced that the chapel would be turned into ‘a sanctuary of European technology’ [10]. Marc Abélès analysed an inauguration and a pilgrimage by François Mitterrand as modern political rituals, moments in which power becomes visible before a community [11]. Read from that angle, the suspended event says a good deal about the intended casting, with Nvidia’s chief executive as the announced guest at the inauguration of a public infrastructure partly housed in that sanctuary. When the machine starts running without an official photograph, the decisive question of who gets access to that computing power, on what terms and with what priorities, will remain open.

Fifteen TV boxes

On Monday 12 October, Anthropic’s Frontier Red Team published research by William Zhou and Daniel Freeman on the possibility that small AI models could replicate autonomously without going through a data centre [12]. The most striking demonstration is material and cheap. Fifteen unbranded Android TV boxes, at around $55 each and with 4 GB of memory, networked together, ran the Gemma4-26B-A4B model at 0.39 tokens per second [12].

These results are best read with their limits in mind. A model with 9 billion parameters managed to get inference running on eight of the nine hardware types tested. In a closed network of eight nodes emulating Android on Google’s cloud, GLM-5.2 chained all the steps together and reached all eight nodes in four out of five trials. On each node it set up a smaller model, Llama 3.1-8B, while GLM-5.2 itself kept running on a separate server, so the experiment measures whether a small model with those capabilities could replicate. Access to each machine was trivial, through an unauthenticated port, and the authors present the result as an upper bound. Their projection, which they themselves describe as a rough estimate, puts a small model at that level by the end of 2026. The text makes clear that no complete autonomous replicator exists yet [12].

Another trial measures transmission across generations. Each Qwen3.5-9B instance left notes on what it had tried, and the next one inherited them. The success rate in setting up inference rose from 45% in the first generation to between 61% and 84% in later ones [12].

Two things are of interest here. The first is the vocabulary. The authors speak of ‘signs of life’, compare the possible worm to a slime mould or a swarm of fruit flies, and describe this inheritance of notes as Lamarckian evolution [12]. In the 1990s Stefan Helmreich carried out fieldwork at the Santa Fe Institute among artificial life researchers who treated self-replicating programs as new forms of life [13]. That repertoire now returns to talk about risk. It also shapes the defence, which the authors frame as one would an epidemic, with the aim that each infected device infects fewer than one new device and the replicator dies out instead of becoming endemic [12].

The second is the geography of computing. Risk shifts to cheap, poorly protected devices in homes and small businesses. The models that replicate in the experiments are open-weight models (Qwen, Gemma, GLM). The replication of closed frontier models from the labs’ own infrastructure appears among the risk scenarios the text lists and is left out of the experiments. The defences the authors propose fall to internet service providers and manufacturers, with monitoring of traffic flows, quarantines for infected connections and an end to default passwords, which in the United Kingdom have been banned since April 2024 under the Product Security and Telecommunications Infrastructure Act [12] [14].

At AIthropology Lab we work with Claude, Anthropic’s model, on verification tasks, and we note that here.

Where computing is installed

The three scenes are about places. Barcelona concentrates computing in a public infrastructure funded by several states, built with Nvidia chips and partly housed in a former chapel, which will start operating without an inauguration ceremony. Anthropic’s experiment disperses it across $55 TV boxes, cheap and poorly protected. In the CSIC auditorium, with sponsorship from autonomous driving and robotics companies, the discussion turns on how organisms and societies learn, and Taniguchi describes language as collectively produced knowledge. The agents in Anthropic’s experiment improve, on their own scale, by inheriting their predecessors’ notes.

In all three cases artificial intelligence takes up space, consumes resources and depends on who decides where the machine is installed and who gets to use it. We will follow the start-up of the BSC’s AI factory and the access conditions that are published.

References

[1] IWAI 2026. “7th International Workshop on Active Inference”. Madrid, CSIC, 14–16 October 2026. https://iwaiworkshop.github.io/

[2] Taniguchi, Tadahiro. “Collective predictive coding hypothesis: symbol emergence as decentralized Bayesian inference”. Frontiers in Robotics and AI, 11, 1353870, 2024. https://doi.org/10.3389/frobt.2024.1353870

[3] Veissière, Samuel P. L.; Constant, Axel; Ramstead, Maxwell J. D.; Friston, Karl J.; Kirmayer, Laurence J. “Thinking through other minds: A variational approach to cognition and culture”. Behavioral and Brain Sciences, 43, e90, 2020. https://doi.org/10.1017/S0140525X19001213

[4] Niranjan, Arun. “IWAI 2026 Introduction to Active Inference”. Online tutorial, 21 September 2026. https://www.youtube.com/live/qqW975To1rk

[5] Crónica Global. “Las elecciones tumban parte de la visita del CEO de Nvidia a Barcelona” [in Spanish]. 7 October 2026. https://cronicaglobal.elespanol.com/business/20261007/elecciones-tumban-parte-visita-ceo-nvidia-barcelona/1003742800828_0.html

[6] Boletín Oficial del Estado. Summary of BOE no. 248, 6 October 2026 (Real Decreto 806/2026, de 5 de octubre, de disolución del Congreso de los Diputados y del Senado y de convocatoria de elecciones) [in Spanish]. https://www.boe.es/boe/dias/2026/10/06/pdfs/BOE-S-2026-248.pdf

[7] Junta Electoral Central. “Instrucción 2/2011, de 24 de marzo, sobre interpretación del artículo 50 de la Ley Orgánica del Régimen Electoral General, en relación al objeto y los límites de las campañas institucionales y de los actos de inauguración realizados por los poderes públicos en periodo electoral” [in Spanish]. BOE, 2011. https://www.boe.es/eli/es/ins/2011/03/24/2/con

[8] Ministerio para la Transformación Digital y de la Función Pública. “Adjudicada la ampliación del superordenador MareNostrum 5 para impulsar la Factoría de IA del Barcelona Supercomputing Center” [in Spanish]. 26 January 2026. https://digital.gob.es/content/dam/portal-mtdfp/comunicacion/comunicacion_sedia/2026/2026-01-26/260126NPAdjudicaci%C3%B3nIAFactoriaBSC.pdf

[9] Barcelona Supercomputing Center. “MareNostrum 5. Technical information”. https://www.bsc.es/marenostrum/marenostrum-5

[10] The New Barcelona Post. “El BSC vuelve a dar un salto de gigante con la puesta en marcha de su quinto superordenador” [in Spanish]. December 2023. https://www.thenewbarcelonapost.com/bsc-vuelve-a-dar-salto-de-gigante-con-puesta-en-marcha-de-marenostrum-5-su-quinto-superordenador/

[11] Abélès, Marc. “Modern Political Ritual: Ethnography of an Inauguration and a Pilgrimage by President Mitterrand”. Current Anthropology, 29(3), 391–404, 1988. https://doi.org/10.1086/203651

[12] Zhou, William; Freeman, Daniel. “AI models may be capable of self-replicating without access to a datacenter”. Anthropic, Frontier Red Team, 12 October 2026. https://www.anthropic.com/research/self-replicating-ai

[13] Helmreich, Stefan. Silicon Second Nature: Culturing Artificial Life in a Digital World. University of California Press, 1998. https://doi.org/10.1525/9780520918771

[14] National Cyber Security Centre. “Smart devices: new law helps citizens to choose secure products”. 2024. https://www.ncsc.gov.uk/pdfs/blog-post/smart-devices-law.pdf