Daniel Bolojan, the principal designer of our case study, is the founder of Nonstandardstudio, an agency specializing in the application of artificial intelligence to architecture, established in 2013 and based in the United States. The agency explores, in particular, deep learning techniques and their uses in architectural design. This approach to artificial intelligence, whose origins date back to the 1940s, achieved significant progress during the 1980s before experiencing a spectacular breakthrough from 2010 onwards (Ganascia, 2025). Trained as an architect, Bolojan graduated in 2008 from the Ion Mincu University of Architecture and Urbanism in Bucharest, Romania. He also obtained a design degree in 2014 from the University of Applied Arts in Vienna, where he is currently pursuing a PhD in architecture. He subsequently worked at internationally renowned firms such as Coop Himmelb(l)au and Zaha Hadid Architects, contributing to various avant-garde projects. In parallel, he has taught workshops and seminars at the Institute of Structure and Design of the University of Innsbruck (Austria), as well as at the School of Architecture at Florida Atlantic University (United States). His research focuses on the potential of artificial intelligence to enhance human creativity in the built environment, particularly in the fields of architecture and design.

Among his notable works are Machine Perceptions: Gaudí + Neural Networks and Deep Learning Sagrada Familia, carried out between 2018 and 2019. In these projects, Bolojan developed neural prototypes based on generative adversarial networks (GANs), capable of analyzing the aesthetic and semantic principles of Antonio Gaudí’s works and generating new reinterpretations. In 2019, he also contributed to Deep Himmelb(l)au, a project undertaken within the Coop Himmelb(l)au office, which sought to accelerate the design process through a generative program inspired by the firm’s previous projects.

Emmanouil Vermisso, the second contributor to our case study, is an Associate Professor of Architecture at Florida Atlantic University, while simultaneously pursuing a PhD at the University of Patras in Greece. Vermisso studied architecture at the University of Westminster in the United Kingdom and Syracuse University in the United States. He also worked at internationally renowned firms such as Foster+Partners. His current research focuses on artificial intelligence in design, with the aim of better understanding the creative and cognitive processes of the architect. Some of his works have been awarded and published in specialized journals such as the International Journal of Architectural Computing, receiving recognition from notable organizations, including the Association for Computer Aided Design in Architecture (ACADIA), Education and Research in Computer Aided Architectural Design in Europe (eCAADe) in 2007 and 2008, Computer-Aided Architectural Design Futures (CAAD Futures) in 2011, and the Sociedad Iberoamericana de Gráfica Digital (SiGRADi) in 2019.

Machine Perceptions: Gaudí + Neural Networks (Figure 1) is a study that explores the possibilities of generative design using artificial intelligence to learn, at different scales of reading, the aesthetic characteristics of Antonio Gaudí’s work. This project relies on neural networks to analyze and interpret photographs of the interior of La Sagrada Família, as well as video sequences captured during a walk in a nearby forest (Figure 2). La Sagrada Família, Gaudí’s most significant and complex work, embodies the genius of this visionary architect of Catalan Modernismo, who combined the skills of sculptor, painter, blacksmith, and mathematician. In this study, Daniel Bolojan and Emmanouil Vermisso sought to explore the collective perception of visitors in relation to this iconic building. Their objective was to generate a “hallucinated” space, founded on alternative realities derived from subjective and unconscious interpretations of these perceptions (Bolojan & Vermisso, 2020). La Sagrada Família was chosen as a case study because of its exceptional spatial complexity. This complexity is expressed in the ambiguity of tectonic distinctions between its various components, blurring the boundaries between architectural elements (ibid.).

Initially conceived as an “expiatory temple” dedicated to the Holy Family, La Sagrada Familia has since become an emblematic basilica and an indispensable monument of the Barcelona landscape. Since 2005, its crypt has been inscribed on the UNESCO World Heritage List. Although still unfinished, the monument inherits its Gothic floor plan from the project originally designed in 1882 by the architect Francisco de Paula del Villar. Antonio Gaudí took over the construction two years later and remained in charge until his tragic death in 1926. To date, only the Nativity facade, one tower, and part of the apse wall have been completed. Despite the destruction of the architect’s workshop by fire and acts of vandalism during the Spanish Civil War in the 1930s, construction has continued to this day, testifying to the remarkable commitment to completing this architectural masterpiece. Built with millimetric precision, the building is scheduled for completion in 2026, coinciding with the centenary of Gaudí’s death.

Although its construction began more than a century ago, the unfinished work of La Sagrada Família underscores the necessity of advanced digital technologies to interpret and understand Antonio Gaudí’s conceptual vision (Makert & Alves, 2016). Gaudí left behind neither a clear architectural theory nor defined axioms, but rather a set of spatial reflections. When asked whether La Sagrada Família followed in the tradition of the great cathedrals, he replied: “No, it is the first of an entirely new series” (Zerbst, 1999). Indeed, Gaudí approached space in a resolutely organic manner. He conceived the structure by drawing direct inspiration from vegetal forms and developed an abstract, fluid architectural language, blending parabolas, hyperbolas, helicoids, conoids, and ruled surfaces - complex forms requiring colossal mathematical calculations. To achieve this, he handcrafted numerous prototypes, both in miniature and full scale. In La Sagrada Família, particularly in the main nave, towering tree-like columns rising harmoniously toward the sky emerge as the dominant formal figure. Gaudí summarized this intention by stating: “It will be like a forest of trees” (Curti, 2020-07). In this stylized forest, “we are confronted with the joy of life, flora, fauna,” observes Judith Urbano (Urbano, 2023-07). Gaudí also enriched the design with abundant ornamentation populated by figurines of birds, wild animals, and sacred characters. The result is a masterpiece where nature and the sacred fuse into a visionary architecture - at once neo-Gothic and retrofuturistic.

Often described as a “Bible in stone” (ibid.), La Sagrada Família draws inspiration from Gothic, Moorish, and Catalan vernacular influences. Nicknamed by Gaudí as the “cathedral of the poor,” it was conceived as a place of worship for the workers of Barcelona. The building adopts a basilical plan, symbolizing the cross of Christ. Gaudí designed it as a basilica with five longitudinal naves and a transept composed of three naves, forming a Latin cross. It is encircled by a rectangular cloister and twelve peripheral towers representing the apostles. In addition, six central towers are planned, one dedicated to Christ and another to the Virgin Mary. The choir is completed with an apse and bordered by an ambulatory. The altar is surrounded from the outset by seven radiating chapels, each dedicated to the sorrows and joys of Saint Joseph. In his architectural conception, Gaudí proposed a symbolic reading of the mystical body of Christ (Zerbst, 1999). Each façade illustrates a major stage in the life of Jesus: the Nativity façade, facing east and constructed between 1883 and 1926, celebrates his birth; the Passion façade, to the west, erected between 1954 and 1977, evokes his suffering and passion; and finally, the Glory façade, still under construction, symbolizes the resurrection and eternal life.

Machine Perceptions: Gaudí + Neural Networks is an architectural creation beyond the human, insofar as much of the creativity emerges from an alternative approach to human authorship. The work is based on an artificial intelligence program applied to an architectural corpus, particularly that of La Sagrada Familia, and enriched by advanced digital technologies. The creativity of the AI is revealed through a process of “machinic hallucination,” whereby hybrid images are generated from a vast database of digitized visuals. The outcome takes the form of an immersive animation, offering a novel aesthetic that merges Antoni Gaudí’s organic naturalism with the algorithmic inventiveness of an AI capable of transcending the boundaries of human creation. Within this framework, Machine Perceptions: Gaudi + Neural Networks was developed during a workshop entitled Gaudi’s Hallucinations, organized as part of the eCAADe-SIGraDi 2019 conference in Porto. This two-day workshop explored the role of architecture in the era of the Fourth Industrial Revolution through more than two hundred presentations and workshops, the results of which were published in the proceedings. During the event, nearly three thousand images of Gaudí’s works were processed using a generative adversarial network of the CycleGAN type, an artificial intelligence technique that enabled the production of novel spatial and aesthetic visualizations of La Sagrada Familia. Unlike popular programs such as DALL-E, MidJourney, or Stable Diffusion, which generate images from textual prompts, this AI technique relies on training a visual corpus directly introduced into the machine, thus opening the way to radically new spatial and aesthetic reinterpretations.

Machine Perceptions: Gaudí + Neural Networks was published online as a video of approximately two and a half minutes (video 1). Online on June 11, 2021, on the Vimeo platform, it was presented at the 17th Venice Architecture Biennale as a demonstration of the Gaudi’s Hallucinations workshop. The video opens with a quotation from Jacques Derrida, taken from his book The Truth in Painting, which interrogates the notion of perception (Derrida, 2010). From the fifteenth second onward, hallucinatory high-resolution images appear, depicting dreamlike manifestations of the nave of the building. These representations are derived from a corpus of nearly 500,000 photographs, processed and trained within the latent space of the AI program - a kind of “computational mind” (Oktan & Kevser, 2024-09).

At the twenty-third second of the video, a second grid of images appears, representing the main nave of the monument - what Bolojan and Vermisso refer to as a “feature map” (Figure 3). This grid undergoes a topographic transformation generated pixel by pixel. At the twenty-ninth second, metamorphosis effects unfold within the latent space, exploring the infinite possibilities of the algorithm (Figure 4). By the forty-third second, the process of deep learning is illustrated. This process superimposes and blends, layer by layer, a low-angle photograph of the building’s interior with an image of trees photographed at a similar angle in a nearby forest (Figure 4).

By the end of the first minute, the outcome of this spectacular data training is revealed: AI-generated images presenting an alternative and fascinating vision of the interior space of La Sagrada Família (Figure 5). What is perceived oscillates between different sensitive impressions: at times a staged mummified skull evoking death and resurrection; at times a fragment of mountainous rock finely chiseled by a natural hand; at times again a marine sponge, revealing and concealing itself with the rhythm of a tide. These atypical visions animate twenty seconds later, establishing a dreamlike and hallucinatory atmosphere—between optical illusion, echoing Gaudí’s aesthetic, and immersion in the depths of an organic world of remarkable plastic power. Finally, at the beginning of the second minute, a concluding note drafted by the designers appears, outlining both the advantages and the challenges related to the confrontation between human perception and that of an emergent GAN algorithm.

Generative Adversarial Networks (GANs) constitute the deep learning technique widely used in the project by Bolojan and Vermisso. Developed in 2014 by Ian Goodfellow and his collaborators, GANs are based on the competitive interaction between two artificial neural networks: a generator and a discriminator (Ganascia, 2025). The two networks confront each other in a contest where the first generates a sample from a database trained within the latent space of the AI program, while the second attempts to identify it as real or artificial. In other words, the generator seeks to deceive the discriminator, while the latter evaluates the credibility of the generated data. This process results in “hyper-fabricated” data of often astonishing creative quality. The ultimate goal of this interaction is to reach a “Nash equilibrium,” that is, a situation where the strategies of both networks are optimized such that neither can improve its performance. This equilibrium, however, is difficult to achieve due to the complexity of interactions between the two networks (Goodfellow, 2016).

Generative Adversarial Networks have given rise to several variants, among which the CycleGAN, employed by Bolojan and Vermisso in their project. CycleGANs, short for Cycle-Consistent Generative Adversarial Networks, were developed to transform preexisting images or to generate new ones that are photorealistic, hybrid, or entirely abstract (Figure 7). In the work Machine Perceptions: Gaudí + Neural Networks, Bolojan and Vermisso train their model using real images of the Barcelona forest and photographs taken inside La Sagrada Família, thus enabling the discriminator to reliably assess whether or not an image corresponds to the interior of the monument’s central nave. To achieve this, they used metrics such as the Structural Similarity Index Measure (SSIM), as well as very deep convolutional networks (VGG network). The generator then created new images based on these inputs, which were subsequently submitted to the discriminator for categorization. Through this iterative corrective process, the generator progressively refined its image transformation until most of the generated images were validated by the discriminator.

From an architectural perspective, the results obtained through this artificial intelligence approach are impressive, particularly when it is used to enrich the formal language more organically, pushing beyond what is traditionally present in the work of La Sagrada Família. Although it relies on two-dimensional images, this approach to creation via artificial intelligence could greatly assist architects in pushing the boundaries of design, while offering insights into the importance of specific architectonic elements of the sampled building. However, exploration in the third dimension represents both a technological challenge and a promising advancement in creative practice (Bolojan & Vermisso, 2020). Furthermore, this experiment reinterprets the human perception of what Gaudi’s work represents in the unconscious of observers, while also mapping these spatial impressions to “generate mutated versions of spaces based on alternative [real/virtual] realities” (ibid.).

The work Machine Perceptions: Gaudí + Neural Networks illustrates a rich graphical ingenuity, revealing a previously unknown synergy between a historical masterpiece and an AI-generated creation produced by cutting-edge algorithms. Nevertheless, the CycleGANs employed raise several challenges. First, the relevance of image hallucination depends on the available computational power, particularly the computer’s graphics processing capacity. Moreover, these algorithms produce only two-dimensional images and approach three-dimensionality solely through perspective representations, thereby limiting their sublime transcription of La Sagrada Família. Finally, the identification and categorization of the generated images are influenced by the comparison technique used and the number of pixels, which condition their quality, currently considered “limited” and “failing to account for many nuances of human perception” (Zhang, 2018).

Born in the wake of Viennese Actionism, Coop Himmelb(l)au is an architecture firm founded in 1968 by Austrian architects Wolf Prix and Michaël Hölzer, together with Helmut Swiezinsky, a Polish-born architect. Wolf Prix, the firm’s principal director, is a prominent figure of Deconstructivism and an active member of several architectural institutions. Michaël Hölzer and Helmut Swiezinsky remained members of the firm until 1971 and 2001, respectively. The name Coop Himmelb(l)au, chosen by the founders, is a German wordplay meaning “to build the sky.” It also alludes to Himmelbau, referring to the “blue of the sky,” which explains the parenthesis around the “l” in Blau.

Coop Himmelb(l)au achieved paradoxical success through its intervention on a Viennese rooftop for the design of a law office, which earned it an invitation to the landmark Deconstructivist Architecture exhibition held at the Museum of Modern Art in New York in 1988. This exhibition brought together a handful of architects labeled as “deconstructivists,” who sought to produce buildings capable of establishing a critical distance between architecture and the meanings intrinsically tied to it. Another spectacular project was Open House, designed for an elderly Austrian psychoanalyst who wished to spend his final years in Malibu, California. Notably, the first sketch of this house was produced with eyes closed, hand-drawn “like a seismograph” (Rambert, 2015), recalling certain automatic drawings practiced by Dadaists and Surrealists in the 1920s and 1930s (Zukowsky, 2006). The unsettling imbalance of the house’s spaces generates sensations of slippage, instability, and dramatization that “implode the relation to reality in the name of heterotopia” (Hugues, 2010). Working in this way in a state of “semi-consciousness” (Mönninger, 2010), Coop Himmelb(l)au seemed to aim at “opposing the vital function of aesthetics to the leaden dogma of profitability” (Roder, 2002). Bold projects such as the East Pavilion of the Groningen Museum in the Netherlands (1994), the Seibersdorf Research Center in Austria (1995), the Media Pavilion of the Sixth Venice Architecture Biennale (1996), and more recently BMW Welt or the European Central Bank in Germany, as well as the Musée des Confluences in France, all bear witness to an unprecedented distortion of our habitual sensory perceptions.

The works of Coop Himmelb(l)au have consistently adopted a visionary approach that integrates computational advances and cutting-edge technologies, including artificial intelligence. According to Wolf Prix, architecture must embrace the possibilities offered by artificial intelligence, considering it not merely as a “tool” but as a “collaborator” that enriches creativity (Feng, 2024). This ambitious perspective has raised ethical concerns among critics such as Timnit Gebru, who views AIs as “stochastic parrots”: “parrots because they repeat the [information] from which they have learned; stochastic because they rely solely on probabilities” (Ganascia, 2025). Prix, by contrast, perceives AI as an unexplored potential, comparing it to the feathers of a dinosaur developed by nature without foreseeing that they would one day enable flight. For him, artificial intelligence is a tool that will allow “architects to fly” (Prix, 2024). It is within this spirit of openness and adventure that Deep Himmelb(l)au emerged - an unprecedented research project at the crossroads of architecture, professional practice, and artificial intelligence (Figure 8). This experimental work, as its name suggests, employs deep learning techniques to teach computers to interpret, design, and enrich the creativity of the practice through images of its past projects (Prix et al., 2022)

Generative Adversarial Networks (GANs), introduced in the previous chapter, constitute a subfield of artificial intelligence aimed at extracting and automatically generating knowledge from massive datasets projected into latent space. Inspired by the functioning of the human brain, they reproduce with high fidelity its mechanisms of learning, prediction, and decision-making. In the context of the Deep Himmelb(l)au project, the primary models employed were CycleGANs. Similar to the Machine Perceptions: Gaudi + Neural Networks project, these networks were used to generate stylized architectures, reinterpreting the distinctive aesthetics of Coop Himmelb(l)au. This process enabled the creation of novel visualizations that merge the agency’s identity with the creative potential of algorithms.

The generative networks used in Deep Himmelb(l)au operate in the same manner as those in Machine Perceptions: Gaudi + Neural Networks, but this time they were trained on a corpus of images from the practice’s earlier projects, spanning nearly fifty years of production and preserved in both analog and digital archives (ibid.). These archives include simple hand-drawn sketches, miniature models, photographs, 2D drawings, 3D models, and construction details. Immersed in an iterative network of generator and discriminator, the CycleGANs not only recognize and distinguish iconographic content within batches of images, but also generate new architectural spaces. Each neural network trains itself to understand the analyzed semantics and underlying compositional rules in order to produce novel images without requiring human supervision (Leach & Prix, 2021). The result is not a mere replica of the practice’s past projects but rather an amalgam - a genuine “augmentation of the design process” (Bolojan, 2022) - constructed from hundreds of thousands of data inputs introduced into the system during its training. This gives rise to a spectacular cinematic immersion in an imaginary landscape of hallucinatory building forms, reminiscent of Coop Himmelb(l)au’s characteristic aesthetic (Leach, 2021).

Deep Himmelb(l)au is a collaborative endeavor bringing together Wolf Prix and Karolin Schmidbauer on the conceptual side, Daniel Bolojan and Efilena Baseta on the computational side. This project explores the potential of deep learning to produce a semantics intrinsic to the formal language of Coop Himmelb(l)au. It is worth emphasizing that Deep Himmelb(l)au represents the most advanced research to date in artificial intelligence applied to architecture, particularly in its capacity to explore three-dimensionality through supervised learning models beyond CycleGANs. Furthermore, the work has been recognized by DigitalFUTURES, a platform dedicated to architecture that annually organizes a series of scholarly activities focused on the latest technologies in computational design and fabrication, including artificial intelligence.

The Deep Himmelb(l)au program is based on a network of nodes structured around several GAN models—primarily CycleGANs—themselves grounded in the dialectic between generator and discriminator. These interconnected GANs iteratively and incrementally interpret the agency’s earlier projects. They are organized around three thematic axes: Gestalt, organization, and technique (Bolojan, 2022). The first, Gestalt, addresses the formal dimension of architecture and relies on image generation. The second organization focuses on the functional dimension, training models with plans and sections. Finally, the third technique deals with technological aspects, particularly those related to environmental challenges.

The Deep Himmelb(l)au program includes several modes of architectural generation. First, it presents itself as a cloud of miniature images (Figure 9), gathering the agency’s various projects and thereby constituting a visual mapping of its latent space. Certain areas within this space correspond to specific aesthetics that are easily identifiable by the user, thus facilitating interaction with the program. Selecting a particular region of the cloud allows the generation of an animation, whether as a two-dimensional representation—fixed through plans and sections—or as a three-dimensional projection producing hallucinatory perspectives and volumetric simulations. Furthermore, the program offers the possibility of modeling physical prototypes, digitizing them, and subsequently assessing their environmental performance (Figure 10). Finally, a third module enables the user to generate new variants of architectural images from textual prompts (Figure 11). The user then enters a description, in the manner of Midjourney or DALL-E, to obtain building proposals corresponding to the given statement, ranked according to text–image likelihood ratios.

It should be recalled that Midjourney and DALL-E are generative AI programs capable of transforming textual inputs—sometimes absurd—into coherent images (Ramesh, 2021). Developed by the American company OpenAI in 2021, DALL-E uses a deep learning model to produce detailed, high-quality images that demonstrate a refined semantic understanding of the user’s textual input and remarkable compositional abilities. Midjourney, which appeared a year later, distinguishes itself through greater efficiency in generating painterly images. Trained on a vast corpus of artistic works, it has acquired an understanding of artistic canons and of painting in particular (Mahmoud, 2022). Thus, while DALL-E seems more oriented toward the formal composition of the images it generates from textual prompts, Midjourney privileges its visual and aesthetic qualities (D’Armenio, 2024).

Deep Himmelb(l)au is manifested through a series of videos published online. The first and most notable, lasting one minute and forty-seven seconds, was released on June 4, 2020, on Vimeo and on the agency’s website (video 2). It presents a subtle-hued 3D simulation, immersing the viewer in a promenade around a constantly transforming building, thereby generating new spatial interpretations (Figure 12). The sequence’s aesthetic draws on a metaphorical interweaving of the agency’s landmark projects, such as the UFA Cinema Center (Dresden, 1998), the JVC New Urban Entertainment Center (Mexico, 2001), and BMW Welt (Munich, 2007). In this simulation, the ceiling - sometimes constructed and fixed, sometimes absent and dissolved - disappears under a succession of overlapping visual layers that intersect and distort, thus altering the architecture of the building.

The second video, about one minute in duration, was posted on October 18, 2019, on the agency’s website and YouTube channel (video 3). It presents an accelerated tour around buildings with blurred and perpetually metamorphosing contours (Figure 13). This animation appears to correspond to a proposal for a port development competition in Crimea, Russia, mentioned by Wolf Prix during a videoconference in 2021 (Leach & Prix, 2021). Its deformed, at times even crushed, aesthetic evokes Francis Bacon’s paintings, particularly the Portrait of Michel Leiris (Figure 14). The animation also integrates formal reminiscences of projects realized by the agency, such as the Dalian Conference Center (China, 2011), the European Central Bank (Frankfurt am Main, 2014), and the Musée des Confluences (Lyon, 2014).

The third video, lasting about ten seconds, appears around the fifty-ninth minute of a lecture entitled The Project Himmelblau, delivered at the FAU School of Architecture on February 23, 2021 (video 4). It also circulates on social media, particularly Facebook (video 5). Dating from 2021, this sequence presents the conceptualization of a congress center associated with a hotel complex, dominated by a sloping roof with exaggeratedly knotty and protruding forms (Figure 15). This metaphorical roof, oscillating between bulging volumes and unfolded surfaces, hollows out and expands to evoke an armored shell. Enhanced by sophisticated material effects, it leaves the viewer with a singular impression tinged with strangeness, even surrealism.

The programming of Deep Himmelb(l)au mobilizes a variety of Generative Adversarial Network (GAN) models, enabling it to perform complex conceptual tasks. CycleGANs, which play a predominant role, are distinguished by their ability to establish correspondences between two different visual domains without requiring aligned image pairs (Bellale, 2022). They notably allow the extraction, within the program’s latent space, of the aesthetic features of previously trained images, which can then be synthesized to generate coherent stylistic transformations. StyleGANs, for their part, offer refined multi-scale control over the manipulation of image styles and achieve such a degree of realism that synthetic images often become indistinguishable from their real counterparts (Lago, 2022). Pix2Pix models condition image generation based on input data (Joyce et al., 2021). In this context, they ensure the transformation of images while preserving the iconographic characteristics of the source data. By combining these different models, the algorithmic approach of Deep Himmelb(l)au provides artificial intelligence with a significant advantage for complex tasks such as the conversion of two-dimensional spaces into three-dimensional ones, self-learning, and the synthesis of multiple digitized representations (Yousif & Vermisso, 2022).

Deep Himmelb(l)au is capable of addressing new spatial perspectives inherent in the morphogenesis of Coop Himmelb(l)au and making them communicable, rather than evolving in the traditional way from experiences, knowledge, and creativity passed down from generation to generation. Despite the heterogeneity of its formal language, Deep Himmelb(l)au has succeeded in faithfully creating unique works of art consistent with the agency’s predominant style, thereby pushing the boundaries of architecture and technology. These creations affect the invention of architectural aesthetics and challenge our very definition of the discipline, without, however, fully surrendering control of the design process to artificial intelligence (Feng, 2024). AI-driven creation, while unable to access human consciousness or emotion, nonetheless exceeds our cognitive capacity—an aspect that raises ethical concerns for some and skepticism for others.

Since its earliest works, Coop Himmelb(l)au’s creative approach has been engaged in an “open process” (Prix et al., 2022), initially initiated through seismic and intuitive sketches, rich in meaning and open to interpretation. With the advent of digital tools, this method of manual and analog drawing first gave way to parametric models that made complex geometry manageable and controllable, then to photogrammetry for measuring and reproducing models, next to robotic arms and additive manufacturing that broadened the field of possibilities, and eventually to computer-aided manufacturing, simulations, and digital animations that temporalized space and its representation. Remaining faithful to this openness to the “other,” the engagement with machine learning and artificial intelligence in Deep Himmelb(l)au combines all these techniques in a retroactive and hybrid process oscillating between the poietic and the technological. Indeed, computational tools, since their emergence in the 1990s, have been welcomed at the agency—not as substitutes for conceptual means but as amplifications and spatial optimizations. It is, in fact, a form of astonishing co-creation that surpasses human talent to yield an architecture beyond the human—oriented toward the future, yet deeply rooted in the past.

Not applicable

The author(s) declare that there is no competing interest.