Sourek Architectural Intelligence (2024) 3:4 Architectural Intelligence

https://doi.org/10.1007/s44223-023-00046-9

REVIEW ARTICLE Open Access

AI in architecture and engineering

from misconceptions to game-changing
prospects
Michal Sourek1*

Abstract
Artificial intelligence invades our lives and professions at an ever-increasing pace and intensity. Architecture, engi-
neering, construction, and operation of the real estate have been joining the trend only timidly and belatedly. The
paper overviews the basic concepts, methods, general background, and results of artificial intelligence in architecture
to date, discusses the achievements and prospects, and concludes the perspectives on the deployment of machine
learning in the field. The record of some of the most recent “famous achievements” in the field is set straight and chal-
lenged, the flawed idea of a (truly) creative potential of the technology is debunked. Its roots equidistributed both in a
farsighted vision of the next workflow of both productive and creative architectural and engineering designing,
and construction and real estate management on the one hand and state-of-the-art machine learning on the other,
an ambitious though realistic blueprint for R&D of AI-fostered architectural creativity, building design, planning,
and operation is tabled for discussion. The attention turns to open-source patterns platforms, generative patterns
processing, generative pre-design, parametric evaluation and optimization, latest achievements in machine learning
building on reinforcement learning, imitation-based learning, learning a behavior policy from demonstration, and self-
learning paradigms zooming in on the design-development processes instead of only on their results. Leveraging
the objectivity of assessments and streamlining workflows, artificial intelligence promises to unleash true architectural
creativity and leverage the productivity and efficiency of the design, planning, and operation processes.
Keywords Artificial intelligence, Machine learning, Open-source platform of patterns, Generative pre-design,
Parametric review and optimization, Imitation-based learning

1 Introduction “the next level” in the field of visual design - and archi-
Artificial intelligence, as algorithmic computer programs tecture, engineering, construction, and operation of real
are called that can generate data from heterogeneous, estate (AECO) hopefully, too. Only belated, however,
most often textual and image templates or patterns, eval- AECO has been joining the trend of artificial intelligence
uate these large data sets according to complex criterion that has been entering our lives and professions since the
structures, develop downstream new algorithms, and use 1980s.
them to generate outputs of a nature identical or subse-
quent to the initial templates and patterns, is considered 1.1 Reimagining AECO
Not by chance, the wavering approach to the new tech-
nology tunes with these disciplines´ failures to cope with
*Correspondence:
Michal Sourek societal and economic development; in AECO – in archi-
[email protected] tectural design and real estate (RE) development in par-
1
Department of Architecture, Faculty of Civil Engineering, Czech ticular, it has been a reality for at least 70 years. At the
Technical University in Prague, Prague, Czech Republic
same time, the influence and impact of the disciplines´

© The Author(s) 2024. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
licence, visit http://creativecommons.org/licenses/by/4.0/.
Sourek Architectural Intelligence (2024) 3:4 Page 2 of 21

general performance on the economy, social affairs and within the framework. Section (3) provides a general
issues, and sustainability are immense. Dissatisfaction summary of the achieved results, which the subsequent
with the state and development of the built environ- section (4) extensively discusses. The overview is a foun-
ment has been increasing through recent decades, while dation for identifying the misunderstandings and limita-
comprehension of the causes is lacking. Nevertheless, tions of recent attempts and expectations regarding the
the need for the field’s paradigm change renders: new, use of AI in AECO - for debunking some of the “fantastic
promissing technologies appear, such as virtual reality achievements of AI promising to make man redundant”
and artificial intelligence (AI), that could improve archi- contemporarily discussed in AECO professions and
tectural, planning, and operation practices. Virtual twins beyond. In contrast, other, so far not considered perspec-
offer unprecedented abilities to create, understand, and tives of AI in AECO are introduced: in pattern-based
communicate AECO and the built environment, while AI approach to design and RE management and operation,
provides machine-learning capabilities for design, plan- in reinforcement learning, imitation-based learning,
ning, and parametric review and assessment. Lack of learning a behavior policy from demonstration, and self-
comprehension of the nature and potential of the tech- learning paradigms zooming in on the AECO-design-
nologies is the issue so far. development processes instead of only on their results.
Discussion (4) sketches how AI could contribute to an
1.2 State‑of‑the‑art upheaval of the AECO professions by overcoming their
Since around 2010, global star-architectural studios recent and contemporary technological lacking behind;
alongside young enthusiasts combining information true, authentic architectural, engineering, and RE-man-
technology and architecture try to embrace AI’s poten- agement creativity will not be stifled by the technology
tial contribution to architectural design or, better to say, – the opposite, it will be unleashed by delegating para-
disclose wherefrom it might stem and what it might con- metric problems to AI that is unbeatable when it comes
sist of. In 2020, DeepHimmelb(l)au - a video of a journey to computational and iterative issues. A flowchart ren-
through an imaginary landscape of Coop Himmelb(l) dering visually the proposed unprecedented AI-aided
au-like building forms - has come into existence. The AECO-design development boosts the effectiveness of
elaboration of datasets of reference images of geomorphic the discussion. Finally, Conclusions (5) outline the prin-
formations on the one hand and actual Coop Himmelb(l) ciples of directions and particular goals of further devel-
au projects on the other by CycleGAN and other forms opment of machine learning for the good of AECO, the
of GAN (generative adversarial network) technologies built environment, and sustainable development.
provided “machine hallucinations” (Bolojan, 2022; Coop
Himmelb(l)au, 2023; Leach, 2022a) - represented prevail- 2 Methods and expectations
ingly in two dimensions, substantially lacking both spa- Though generally (and in this paper, too) labeled as AI,
tial comprehensivity and the for architecture inherent the term machine learning (ML) adheres better to the
interconnectedness of the experiential (poetic, in other use of the technology in AECOe: learning or training is
words) and material attributes, discussed further in sec- the keyword, and machine learning is a label for various
tion (4) of the paper. methods of how, in a shortcut, AI works. ML algorithms
create models stemming sample data that the algorithms
1.3 How is the paper organized have been trainin the models on to make decisions or
Reviewed by Sourek, (2022) and presented in Wear- proposals without being programmed to do so. The base
recho, (2023), the virtual twins’ technology, its benefits for the learning is a set of data possessing the same char-
and prospects attracts no more paper that concentrates acteristics as the data to be generated: a truly large file
on AI. The AECO domain on the one hand and data and as will be shown in (4), and a comprehensive one; what
computer sicence and development on the other are both is not in it, the AI cannot learn. Variations of machine
deeply complex and greatly different from each other: learning deserve reminder: supervised learning, unsuper-
achieving mutual understanding of the two is a challenge vised learning, reinforcement learning, and various alter-
that has so far defied attempts to overcome. Revealing natives and fusions.
such a common ground is the intention of this paper in
the broadest sense. 2.1 How a machine can learn
First, the paper challenges the existing approach to AI’s In supervised learning, the system is given a series of
deployment within AECO, its theoretical starting points, categorized or labeled examples and told to make predi-
and the perspectives put forward so far. To do so, the cions about new examples it hasn’t seen yet, or for which
framework of a state-of-the-art of field is introduced; the the ground truth is not yet known (Christian, 2020a).
methods used and expectations (2) declared are assessed Supervised learning uses labeled datasets, whereas
Sourek Architectural Intelligence (2024) 3:4 Page 3 of 21

unsupervised learning uses unlabeled datasets. “Labeled”, cognitive approaches, and human motivation that, as will
means the data already tagged with the requested answer. be resembled, were an etalon and inspiration for McCull-
In supervised learning, the learning algorithm measures och and Pitts (1943) at the very origins of “learning
its accuracy through the loss function, adjusting until the machines” does not cease to inspire current R&D as well.
error is sufficiently minimized. Two types of supervised Comprising the dopamine-releasing mechanisms, possi-
learning distinct - classification and regression. Classifi- ble-value- and actual-expectation-motivation, curiosity, a
cation uses an algorithm to assign test data into specific self-motivated desire for knowledge, imitation and inter-
categories. It recognizes specific entities within the data- active imitation, self-imitation and transcendence, count-
set and attempts to draw some conclusions on how to less facets of motivation and reward underwent R&D in
label or define the entities. Common classification algo- this regard. Random network distillation (RND) algo-
rithms are support vector machines, linear classifiers, rithms using prediction errors as a reward signal (Chris-
decision trees, k-nearest neighbor, random forest, and tian, 2020a, b) or algorithms approximating a state-value
others. Regression applies to understand the relationship function in a quality-learning framework (Mnih et al.,
between dependent and independent variables - com- 2022), or knowledge-seeking agents (Christian, 2020c)
monly to make projections. Linear regression, logis- are the results. In such algorithms, AI agents render able
tic regression, and polynomial regression are popular to come up with their own objectives, measuring intel-
regression algorithms (IBM, 2022a). ligence in the end effect in terms of how things behave
Unsupervised learning analyzes and clusters unlabeled – not in terms of the reward function (Christian, 2020d).
datasets. These algorithms discover hidden patterns or Imitation-based learning provides three distinct advan-
data groupings without the need for human intervention tages over trial-and-error learning: efficiency, safety,
- without the need for labeling the datasets. In unsuper- and, which also renders promising for AI’s deployment
vised learning, a machine is simply given a heap of data in AECO, the ability to learn things that are hard to
and told to make sense of it - to find patterns, regularities, describe (Christian, 2020e). Moreover (again promisingly
and useful ways of condensing, representing, or visualizing for “machine-learning-driven architecture”), learning by
it. (Christian, 2020a; IBM, 2022b). imitation zooms to the (design) process - “how things
Reinforcement learning concerns how intelligent come to existence” - instead of output - “how things
agents ought to take action in an environment to maxi- shall be”. This seemingly petty distinction will introduce
mize the notion of cumulative reward. Based typically in (3) a branching on the path to efficient deployment of
on the Makarov decision process (a discrete- and con- AI in AECO design - architectural design in particular:
tinuous-time stochastic control process in mathemat- a branching that the efforts so far have not noticed. Sec-
ics) (Jagtap, 2022), reinforcement learning differs from tion (4) of the paper discusses the so far overlooked path
supervised learning in not needing labeled input/output as the game-changer that can unleash the potential of
pairs to be presented, and in not needing sub-optimal deployment of AI in AECO.
actions to be explicitly corrected. Instead, it focuses on Another “next level” of the extrinsic-reward-free
finding a balance between exploration (of uncharted ter- schemes comes with interaction that allows the algo-
ritory) and exploitation (of current knowledge). In other rithm to work properly requiring incredibly little feed-
words, placed into an environment with rewards and back as Ross et al. (2011) dataset aggregation (DAgger)
punishments, [the system is] told to figure out the best way has shown. Brought closer in the Touchstones and Trai-
to minimize the punishments and maximize the rewards blasers sub-section, self-imitation and transcendence
(Christian, 2020a; Wikipedia, a, 2022). As the compre- render to be a “top” of today’s learning schemes (Chris-
hension upgrades of learning processes and - especially tian, 2020f ).
- of consequences of their details for the outputs, the Imitation learning is a framework for learning a behav-
reward signals to fine-tune the models tend to be human ior policy from demonstrations. Usually, demonstrations
preferences based (referred to as reinforcement learning are presented in the form of state-action trajectories,
from human feedback (Ayush, 2023)) instead of simple with each pair indicating the action to take at the state
automatic metrics. Indicated further in this section, the being visited. To learn the behavior policy, the dem-
safety and alignment problems are the starting point for onstrated actions are usually utilized in two ways. The
the deployment of these approaches that are much more first, known as behavior cloning, treats the action as the
time- and cost-consuming. Such is, for example, the case target label for each state and then learns a generalized
of InstructGPT - one of the most advanced language mapping from states to actions in a supervised manner.
models today. Another way, known as inverse reinforcement learning,
New implementations of learning paradigms and views the demonstrated actions as a sequence of deci-
new learning models evolve. The human mind, human sions and aims at finding a reward/cost function under
Sourek Architectural Intelligence (2024) 3:4 Page 4 of 21

which the demonstrated decisions are optimal. Finally, a principle of a GAN is an “indirect” training through the
newer methodology, inverse Q-learning (Q for “quality”) discriminator – competitive network agent that can tell
aims at directly learning Q-functions from expert data, how “realistic” the input seems, which itself also updates
implicitly representing rewards, under which the optimal dynamically. This means that the generator gets no train-
policy can be given as similar to soft Q-learning (Papers ing to minimize the distance to a specific image, but rather
with Code, 2023). All these schemes rely on Markov deci- to fool the discriminator. This enables the model to learn
sion processes, where the goal of the apprentice agent is in an unsupervised manner; however, GANs have also
to find a reward function: find it from the expert demon- proved useful for semi-supervised learning, fully super-
strations that could explain the expert behavior in rein- vised learning, and reinforcement learning (Leach, 2022b).
forced learning or find the agent’s objectives, values, or An artificial neural network works by computing. In
rewards by observing its own behavior in inverse rein- essence, two principles of computing apply in artificial
forcement learning (Gonfalonieri, 2023). neural networks: feedforward computing and backpropa-
Last but not least, dubbed the dark matter of intel- gation. The goal is always to train the models generated
ligence, self-supervised learning renders a promising to cope with the criteria inserted typically by vast collec-
path to advance ML. As opposed to supervised learning, tions of sample datasets. Feedforward computing refers
which is limited by the availability of labeled data, self- to a type of workflow without feedback connections
supervised approaches can learn from vast unlabeled that would form closed loops; the latter term marks a
data (Balestriero, 2023). way of computing the partial derivatives during train-
ing. When training a model in the feedforward manner,
2.2 Artificial neural networks the input “flows” forward through the network layers
Approaching the mentioned GAN technologies, artificial from the input to the output. By contrast, while using
neural networks are a type of ML model that can deliver backpropagation, the model parameters update in the
various tasks, including deep learning, Bayesian learn- opposite direction: from the (one closer to) output layer
ing, and more. An artificial neural network is a collec- to the (one closer to) input one. However, a backpropa-
tion of connected units or nodes called artificial neurons, gation algorithm does not equal training algorithms that
which loosely model the neurons in a biological brain. provide the model dataset updates; backpropagation is
Each connection, like the synapses in a biological brain, a strategy to compute the gradient in a neural network.
can transmit a signal to other neurons. An artificial neu- Backpropagation is a general technique; in terms of neu-
ron that receives a signal then processes it and can sig- ral networks, it is not restricted to feedforward networks,
nal neurons connected to it. A deep neural network is it works for recurrent neural networks as well (Barreto,
an artificial neural network with multiple layers between 2022).
the input and output layers; in a shortcut, a deep neural “Fed” by inputs, the networks deliver outputs “at the
network makes machine learning deep learning. (Bishop, end” that try to mimic the deliverables of human work.
2020; Chaillou, 2022a; Datamind, 2022). The principle is that artificial networks deliver relent-
Designed by Ian Goodfellow and his colleagues in 2014, lessly, very quickly, and in huge quantities - as opposed
GAN is a class of ML frameworks. Representing recent and by contrast to humans. The vision is that amidst
state-of-the-art, GAN is a milestone of R&D launched these quantities in no time at all emerge outputs that not
by neurophysiologist and cybernetician of the University only mimic but also attain, if not surpass the quality of
of Illinois at Chicago Warren McCulloch and self-taught human performance. The vision still leaves something
logician and cognitive psychologist Walter Pitts. Building to a human - the choice of the most suitable output pro-
on Allan Turing’s work On Computable Numbers (1937), vided and its fine-tuning, but who knows - 1 day... Objec-
McCulloch’s and Pitt’s foundations-laying paper A Logi- tively, an evaluation of solutions due to a given set of
cal Calculus of the Ideas Immitent in Nervous Activ- criteria is a task suitable for computer, too - a task easier
ity (1943) set a path to describe cognitive functions in in principle than a creation.
abstract terms showing that simple elements connected
in a network can have a huge computational capacity. 2.4 Networks´ and techniques´ evolution
GANs, a recent revolution in machine learning provides
2.3 Infering by computing results today that achieve appreciation, Leach (2022c)
Building upon “the founding fathers´” achievements, the puts. First introduced in 1987, the pioneers were con-
idea of GAN copes with evolutionary biology principle of volutional neural networks (CNNs), also known as shift
an arms race between two species. Two neural networks invariant or space invariant neural networks, most com-
contest with each other in the form of a zero-sum game, monly applied to analyze visual imagery (Wikipedia, b,
where one agent’s gain is another agent’s loss. The core 2022). The foundations of CNNs were laid in 1979 when
Sourek Architectural Intelligence (2024) 3:4 Page 5 of 21

Kunihiko Fukushima (1980) introduced neocognitron, of VAE’s generative potential to provide large quantities
hierarchical, a multilayered artificial neural network pro- of “outputs” (as AI enthusiasts heralding the twilight of
posed for Japanese handwritten character recognition design call it) - typically in furniture design, fashion, pho-
and other pattern recognition tasks. ImageNet (2022), tography, architecture, and urban design (Wikipedia, e,
a groundbreaking project from the 2010s builds on this 2022).
technology. Graph neural networks (GNNs) are another Oposed to the above outlined discriminative and/or
field of recent research aiming at the processing of graph decoding techniques that identify objects and infer what
data. And various applications of GANs are still emerg- is “real” and what is “fake”, generative AI systems create
ing: FrankenGAN for urban context massing, detailing, objects such as pictures, audio, writing samples, and any-
and texturing, Pix2PixHD by Nvidia for high-resolution thing that computer-controlled systems like 3D printers
photorealistic image-to-image translation, GAN Loci, or can build (Burke, 2022). Generative AI allows machines
GauGAN (Chaillou, 2022b). to create new works based on what they have learned
The “recent” indication adheres better to recursive from others. With such a straightforward deployment, a
and (more popular) recurrent neural networks (RNNs) question arises of how much the (so far existing) genera-
in fact. Recurrent neural networks are recursive artifi- tive AI systems are truly AI-driven in terms of computa-
cial neural networks with a structure of a linear chain. tional networks and processes; however, in the practical
Whereas recursive neural networks operate on any hierar- framework of this paper, the resolution is not of substan-
chical structure, combining child representations into par- tial meaning. As a principle, generative and discrimina-
ent representations, recurrent neural networks operate on tive or decoding systems most often operate paired in
the linear progression of time, combining the previous time GAN models setting the business-as-usual rather than
step and a hidden representation into the representation state-of-the-art of today’s AI industry. Typically, a sys-
for the current time step (Wikipedia, c, 2022). In 1925, the tem labeled as generative AI is self-learning, it uses unsu-
Ising model by Wilhelm Lenz and Ernst Ising was the first pervised learning (but can use other types of ML, too),
RNN architecture that, however, did not learn. Shun’ichi and deploys anomaly detection and problem solving - it
Amari made it adaptive in 1972, to be also called the can come up with innovative solutions or approaches
Hopfield network later. In 1993, a neural history compres- based on its experience with similar problems in the past
sor system solved a “very deep learning” task that required (Schmidt, 2022).
more than 1000 subsequent layers in an RNN unfolded
in time. Long short-term memory (LSTM) networks were 2.5 General context to compare
invented by Hochreiter and Schmidhuber in 1997 and There are ecosystems of natural language processing,
set accuracy records in multiple application domains. image processing, voice processing, gaming, and code
Around 2007, LSTM started to revolutionize speech rec- or software processing and development, further robot-
ognition, outperforming traditional models in certain ics, and expert systems or business intelligence. (Chris-
speech applications. In 2009, a Connectionist Temporal tian, 2020g; Firth-Butterfield, 2022; Gozallo-Birzuela &
Classification (CTC)-trained LSTM network was the first Garrido-Merchan, 2023; Rijmenam Van, 2023; Stylema-
RNN to win pattern recognition contests when it won sev- nia.it, 2023) These ecosystems exist, evolve, and (some of
eral competitions in connected handwriting recognition. them) work (though sometimes obscured, even covered
LSTM also improved large-vocabulary speech recogni- up) already over decades and render mature. It is a real
tion and text-to-speech synthesis and broke records for influx of ever-new AI tools what the present experiences
improved machine translation, language modeling, and (Kilian, 2023; Martin, 2023; Pichai, 2023; Storm, 2023;
multilingual language processing. LSTM combined with Urban, 2023).
CNNs improved automatic image captioning (Wikipedia, Transversally to the closed corporate releases on large
d, 2022). language models, there have also been clever ideas and
Variational auto-encoders (VAEs) develop another breakthroughs in the field of natural language processing.
technique (Stewart, 2022). Unlike GAN, instead of the A new training strategy of meet-in-the-middle (MIM)
generator – discriminator pair, Variational Autoencoder (Nguyen et al., 2023) has been shown to improve not
combines two distinct approaches - encoding and decod- only the performance but also the interpretability and
ing. Encoder abstracts data by compressing while decoder thus security of large language models that the over next
brings the data back to its initial format. Through the subsection will introduce. Significant progress arrived
decompression, or “reparametrization”, the decoder gen- in computer vision, in both diffusion models and neu-
erates variations of the modeled phenomenon (Stack- ral radiance fields (NeRF) (Mittal, 2023) - a type of ML
Exchange, 2022). The ability to emulate a phenomenon algorithm used for 3D modeling and rendering based on
by generating multiple versions of it is a starting point deep neural networks capable of generating high-quality,
Sourek Architectural Intelligence (2024) 3:4 Page 6 of 21

photorealistic images of complex scenes from multiple balance neither by trial and error (as would be typical for
viewpoints. The new MeshDiffusion (Liu et al., 2023) reinforcement learning) nor by human labeling of diverse
allows direct generating 3D meshes without any post- alternatives (as in supervised learning) but by imitation
processing, and also new FateZero (Qi et al., 2023) can of human moves employing one of the novel machine-
edit the style of the videos using text while keeping the learning technologies, also promissing in terms of AECO
pre-trained model weights intact. Last but not least, deployment as reminded earlier in this section.
a promising marriage of NeRF with CLIP (contrastive Fifteen years later, DeepMind’s AlphaGo system finally
language-image pre-training) (Weng et al., 2023; Zheng implemented Arthur Samuel’s vision of a system that
et al., 2023) arrived: LERF (language embedded radiance could concoct its own positional considerations from
ields) (Kerr, 2023). With it, natural language queries in scratch. Instead of being given a big pile of thousands of
a 3D fashion can apply within NeRF, targeting different handcrafted features to consider, it used a deep neural
objects in the scene. This brief overview highlights sev- network to automatically identify patterns and relation-
eral new types of algorithms that may represent the first ships that make particular moves attractive, the same
steps toward productive prospects for deploying AI in way AlexNet had identified the visual textures and shapes
AECO. These algorithms, developed (most likely) with that make a dod a dog and a cat a cat (Christian, 2020i).
no regard to AECO dedication, may break out of the Again, hence an inspiration for AI-led architectural pat-
misconceptions that have so far dominated efforts in the tern-based design and analysis that later the section (4)
field, as section (3) of this paper will show. will discuss.
As implemented already in Deep Blue, focus on the
2.6 Touchstones and trailblazers process instead of the output/result provides another,
Designed by Alex Krizhnevsky et al. (2012) with Ilya Sut- even more important lesson for AI’s deployment in
skever and Geoffrey Hinton, AlexNet set a benchmark AECO design. In October 2017, Google DeepMind
in image recognition. A composition of eight layers, the brought this paradigm to a (so far) ultimate level by
first five convolutional layers, some of them followed by going through with the playing-against-itself strategy in
max-pooling layers, and the last three fully connected AlpahGo Zero (AlphaGo, 2023b).
layers, AlexNet competed successfully in the ImageNet AlphaGo combines advanced search trees with deep
Large Scale Visual Recognition Challenge on September neural networks. The “policy [neural] network” selects
30, 2012, achieving a top-5 error of 15.3%, more than 10.8 the next move to play and the “value network” predicts
percentage points lower than that of the runner up (Ger- the winner of the game: a reinforcement-learning para-
shgorn, 2017). During training, the then-novel graphics digm. Initially, the developers introduced AlphaGo to
processing units (GPUs) delivered the high computing numerous amateur games to help it develop an under-
performance demanded by the (then) exceptional model’s standing of the play. Then it played against different
depth that was essential for its high performance. Hav- versions of itself thousands of times, learning from its mis-
ing spurred many more papers employing CNNs and takes. Over time, AlphaGo improved, becaming increas-
GPUs to accelerate deep learning, the paper introducing ingly stronger and better at learning and decision-making.
AlexNet is one of the most influential in computer vision. AlphaGo went on to defeat Go world champions in dif-
According to Google Scholar (2023), the AlexNet paper ferent global arenas and arguably became the greatest
had over 120,000 citations as of early 2023. Go player ever (AlphaGo, 2023a). Introduced in Nature
When IBM’s Arthur Samuel developed a ML system for Journal on October 19, 2017, AlphaGo Zero is a next-
playing checkers in 1959 (Christian, 2020h), he used 38 level version of the Go software created without using
considerations determining the strength of a position – data from human games, stronger than any previous ver-
the number of pieces on each side, the spatial distribution sion (Silver et al., 2017). By playing games against itself,
of stones, mobility and space, safety and risks, and on. By AlphaGo Zero exceeded all the old versions of AlphaGo
1990, the IBM team working on the chess supercomputer in 40 days (AlphaGo, 2023b).
Deep Blue used 8 thousand such considerations (Camp-
bell et al., 2002). This chess evaluation function … prob- 2.7 The Black box problem, security issue and a threat
ably is more complicated than anything ever described in to humanity
the computer chess literature, put the team lead Feng-hsi- With all these nice results, it’s not clear what these models
ung Hsu – and it deserves noting in this paper’s frame- are learning, Mathew Zeiler puts it (Christian, 2020j).
work that perhaps similarly complicated is the structure Leaving (in this framework) aside the vulnerability of
of considerations on – let’s say – a residential building AI applications to various types of attacks and data “poi-
spatial layout development … Moreover, in Deep Blue, soning” allowing for unauthorized access or control over
those thousands of considerations were brought into the system, the issue is the complexity of AI algorithms
Sourek Architectural Intelligence (2024) 3:4 Page 7 of 21

and models that are more or less impossible to interpret in AI systems by making their decision-making processes
or understand. Black box inflicts a lack of transparency more transparent and understandable. Interpretability is
and accountability in AI decision-making, as it can show the starting point of identifying and addressing a system’s
unclear how the system arrived at a particular output or biases or errors; it comes through various methods, such
decision. Causing unintended consequences or biases, as visualizations, explanations, or feature importance
also in ethical and legal concerns, obviously incorrect measures. Diverse techniques start to diagnose prob-
or misleading deliverables appear rather often. Inher- lems with the network’s training, identify biases in their
ently embedded in the nature of the learning process, decision-making, or optimize its performance for a spe-
an absence of categories “true” or “false” determines cific task (Christian, 2020k; Pandey, 2022). Nonetheless,
AI-driven decision making. The algorithm typically but achieving interpretability can sometimes come at the cost
deduces the degree of conformity or deviation according of the accuracy or performance of the system (Christian,
to patterns arrived at by own judgment, either without 2020l).
human supervision or under human direction, but always And fears remain. At the end of March 2023, Italy
covertly in detail. In cases - not exceptional - when sub- blocked ChatGPT (Zorloni, 2023) to secure the privacy
sequent analysis reveals systematic or occasional inaccu- of people and tycoons of global business claim pausing
racy of the outputs, confusion of cause and effect in the “giant AI experiments” - read the development of AI for
training data that diverts attention to the background of six months (Harari et al., 2023) to prevent an unman-
graphic inputs (bokeh) instead of their core or similar aged reaching of the singularity phase of AI development,
usually shows as the cause. when a spontaneous technological growth breaks out,
The bokeh salience feature of AI provides a com- and not only the society begins to be irreversibly changed
prehensive clarification of „famous “Coop Himmelb(l) by the effects of technology but humanity loses all con-
au “machine hallucinations” (Bolojan, 2022; Coop trol over further development of AI.
Himmelb(l)au and Meet DeepHimmelb(l)au, 2023;
Leach, 2022a). Figure 1 depicts not the creativity of AI 3 Results of applying AI in AECO
but a misleading perception of visual information hid- For a paper headlined Architectural, the preceding sec-
den in the black box of the algorithm; not creativity tion may seem too extensive in terms of both general ori-
but an error and accident. Computer hallucinations by entation and scope. However, the state-of-the-art R&D
unintended bokeh salience are examples of how tech- concerning applying AI in architecture shows the oppo-
nology can be misused to manipulate and misinterpret site: only the awareness of the vast achievements of the
visual information, either to fake art or to distort scien- other fields allows comprehension of how sidelined (not
tific research. The AI development community deserves only in terms of AI) AECO is and, especially, how great
credit for looking for and already delivering the first the opportunities for the development of the branch are.
applications that solve this problem that, nevertheless, A reason for AECO’s, in particular architecture’s side-
remains far from being solved. The goal is to build trust lining in this regard is the multiple dimensionality and

Fig. 1 Cloud augmentation of Deep Himmelblau universe© Coop Himmelb(l)au. https://coop-himmelblau.at/method/deep-himmelblau/

(accessed Aug. 20, 2023)
Sourek Architectural Intelligence (2024) 3:4 Page 8 of 21

diachrony of architecture that contrast with the nature “mood” - suburban, public park, etc. (Chaillou, 2022e;
of the fields of the biggest successes of AI deployment Steinfeld, 2023c). Blending the outcomes of Isola’s team
– one-dimensional language or two-dimensional image. and Steinfeld’s R&D, Sketch2Pix provides an interac-
The dimensions issue shows (close to) obvious - opposed tive application for architectural sketching augmented
to the complexity issue: architecture and development by automated image-to-image translation (Steinfeld,
of the built environment are complex to a degree that, 2023b). Unfortunately, these applications - and others
in practical terms, makes efforts to describe unfeasible. to come further only render the limits of output-focused
That should turn attention to the (relatively novel, in (2) algorithms so far aiming to contribute to architectural
presented) imitation-based learning or, more generally, designs.
to machine learning algorithms elaborating (design) pro- Tom Mayne of Morphosis employed AI to develop
cesses instead of the results. As will be shown later in this operational strategies to generate output that could
section, (processes-)imitation-based machine learning never be predicted. The studio developed Combinato-
should, but so far does not attract attention in the AECO rial Design Studies: a Grasshopper definition of one for-
realm, which remains the domain of GANs. mal study elaborated by GAN technology provided a
Compared to the double challenge – an additional range of further combinatorial options (Leach, 2022e).
dimension and a diachrony nature, and an extreme com- Foster+Partners (2022), another global-star architectural
prehensiveness of the task, the recent and contemporary studio cannot stay aside; in its Applied R + D team archi-
applications´ development efforts show even more dar- tects and engineers together with expert programmers
ing – no matter how unsuccessful a class of them is. Or combine the best of human intuition and computational
is it naivety? rigor working with new technologies such as augmented
reality, machine learning, and real-time simulation.
3.1 AI models and creates architecture – does it? In terms of practical use, predictive simulations render
Among multiple others, also Zaha Hadid (studio) met AI the etalon. ComfortGAN, for example, investigates the
using the technology to render forms not so free to cease challenges of predicting a building’s indoor thermal com-
resembling antic temples patterns that served as imagery fort (Quintana et al., 2020). Also structural design is on
datasets to feed the GAN (Zaha Hadid Studio, 2023). In the lookout for AI. Using Variational Autoencoders, for
doctoral research under the supervision of Patrik Schu- instance, research development at MIT investigates how
macher of ZHA in 2017, Daniel Bolojan created Paramet- AI can generate diverse structures while ensuring perfor-
ric Semiology Study using ML algorithms and other tools mance standards (Chaillou, 2022f ). However, due to the
of gaming AI implemented in Unity 3D to model the essential material liability of the structural design, the not
behavior of human agents in order to test the layout of a yet-solved problems of the algorithm’s black box that do
proposed space (Leach, 2022d). not allow to rely on the machine curb so far the deploy-
Stanislas Chaillou (2022c), Nvidia Company, and oth- ment of AI in structural design to the theory and concep-
ers provide AI applications to generate floorplans and tual drafting.
apartment layouts. ArchiGAN uses generative networks On an urban scale, attempts are ongoing to contribute
to create 2D and 3D building designs based on input by generating “typical style” road- and circulation pat-
parameters such as dimensions and space requirements. terns and networks using - among others - the Neural
Another model is CityGAN, which generates drafts of Turtle Graphics (Chaillou, 2022g; del Campo & Man-
city blocks and buildings. From a practical point of view ninger, 2019). Over the past decade, the deployment of
and concerning the efficiency of deployment, the results online platforms has provided an adequate infrastructure
of both applications are questionable - as in all other sim- to the end users (Tian, 2022), also to deploy Generative
ilar cases. On the principle of image-to-image translation AI: Spacemaker (Kyle, 2022; Spacemaker, 2022), Cove.
with conditional adversarial networks (CANs), Phillip tool (2022), Giraffe (2022), or Creo: Design (2022) are a
Isola (2023) Research Group provides series of machine- few examples of this growing ecosystem, offering sim-
generated facades following the “style” and character of plified access to AI-based predictive models, generative
the pattern deployed as the “input” (Chai et al., 2023; design, augmenting reality, real-time simulation, additive
Chaillou, 2022c). Introduced by the same team, Pix2Pix manufacturing, and IoT to iterate faster, reduce costs, and
(2023) is shorthand for an implementation of a generic improve product (Chaillou, 2022h).
image-to-image translation using CANs. Developed Not only start-ups, academia, and spin-offs of global
in 2019 by Kyle Steinfeld (2023a), GAN Loci generates architectural star-studios go in for AI: the global CAD-
perspective images of urban-like scenes assembled with tycoon Autodesk runs Machine Intelligence AI Lab – and
given facades-like textures, pathways, street furniture, much of Autodesk’s software, including Fusion 360, is
pedestrians, cars, etc., by training to achieve the required AI-enabled and applying generative design today (Leach,
Sourek Architectural Intelligence (2024) 3:4 Page 9 of 21

2018a), not to mention the acquisition of Spacemaker working environment of a design engineer, Creo con-
(TechCrunch, 2022). Nonetheless, as broad as all this list- tributes to the productivity and efficiency of his work by
ing may seem, the development of AI for AECO is still in model-based defining, simulations, additive and subtrac-
its infancy, failing to catch up with LLMs (large language tive modeling and manufacturing; Creo fosters the crea-
models), text-to-image processing, deployment of AI in tive potential of a designer by means of generative design
internet search, content placement, and advertising, but (PTC, 2022). Similarly, Cove.tool delivers performance
also healthcare, pharmaceuticals, insurance, or justice data of the building solutions in real-time employing the
referring to custody and bail (Christian, 2020l). power of AI (FinancesOnline, 2022). Cove.tool is a cloud-
based network of tools that provides interconnectivity
3.2 Assessment of results achieved in AECO within the teams working in the design and pre-construc-
The overview of the results of applying AI in architec- tion cycle on issues of daylight, carbon footprint, climate,
ture achieved so far renders shattered: it is neither by geometry, HVAC, cost, or performance.
accident nor by a lack of caring by the author. When AI Also famous as the 2 hundred and 40 million acquisi-
performs well concerning parametric aspects of diverse tion of the AEC-software tycoon-software-producer
materializations of architecture (such as construction, Autodesk, Spacemaker not only gives the architects and
energy efficiency, daylighting, or noise in buildings and developers the automation superpower to test design
neighborhoods) and fails to be effective and productive concepts in minutes and explore the best urban design
in conceptual architectural design, it is not only a tem- options. It enables users to quickly generate, optimize,
porary swing in the performance of particular efforts. It and iterate on design alternatives, all while considering
is a consequence of AI applications´ developers failing design criteria and data like terrain, maps, wind (as Fig. 2
to grasp and follow the starting points and the workflow depicts), lighting, traffic, and zoning, with the help of AI.
of creating architecture, whether on the scale of build- Utilizing the full potential of the site from the start, it
ings or the built environment. Only by overcoming this allows designers to focus on the creative part of their pro-
problem, the way will open up to the efficient and prolific fessional work (Harouk, 2023).
deployment of AI in architecture and AECO broadly as However, the practical deployment of Spacemaker
section (5) discusses. raises doubts: the workflow is the issue. The user enters
Without questioning skills and ingenuity of the authors, the address of the location and the boundaries of the
the results of the Phillip Isola Research Group, Kyle Stein- territory; with a help of an open database like OSM,
feld, the “typical style” road- and circulation patterns and the algorithm ganarates the terrain, existing buildings,
networks delivered by Neural Turtle Graphics may be and structures; the accuracy of the objects generated
interesting outputs of research efforts in computer sci- depends on the data available the quality of which var-
ence, code development, or graphics, however in terms ies from territory to territory. Nevertheless, so far so
of architectural and AECO workflows and solutions they good. Then the user defines the area to solve, and he
are only minor conributions. Similarly, the parametric can add roads - only manually, an import from a CAD
semiology outcomes of Daniel Bolojan or Tom Mayne’s is not available. Buildings can be placed either manually
operational strategies render too speculative to pro- by inserting individual floorplans as objects that can’t
vide some practical analytical starting point. Results of be subject to later adjustments, or the buildings can
DeepHimmelb(l)au and ZHA alike show outputs of hun- be generated automatically based on input parameters
dreds hours of dedicated work of talented multi-expertise entered: width, height, object shape, minimum/maxi-
teams: outputs (in terms of conceptual approach and mum number of floors, and/or by apartments´ sizes
contribution – leaving aside the “video show” that has mix. Then the user can assess the generated options
little to do with architecture) that the principal of the based on gross- and/or netto-floor area totals. He can
studio would sketch by hand within half an hour or so also further modify the chosen option by some of the
- and, opposed to the AI, he would consider the spatial spatial transformations: shift, rotation, ... Spacemaker
and operational concept represented by the sketch. All evaluates the final solution in terms of noise, wind, sun-
this, is it just a situation of a developing field that needs light, daylight, and microclimate. Exports of the valu-
more time and effort to mature and deliver useful results? ation to Excell and of the model designed to Autodesk
Discussion (4) will confront such a perspective with the Revit or to .ifc format are available. All these are valu-
option that it is a dead-end of the state-of-the-art AI in able and efficient functionalities. However, the work-
architecture and AECO. flow - how the design comes to existence, Pavel Shaban
On the other hand, the values of deliverables provided (2023) of MS architekti, Prague/Czech Republic based
by the AI of Spacemaker, Cove.tool, or Creo appear architectural studio claims, may suit a shortsighted
ambitious. Starting from a better organization of the real estate trafficker but is far from a creative and
Sourek Architectural Intelligence (2024) 3:4 Page 10 of 21

Fig. 2 Spacemaker: Microclimate Analysis. https://www.bing.com/videos/search?q=spacemaker&doc.id=603541543342970728&mid=980959C632

3F8F58FE53980959C6323F8F58FE53&view=detail&FORM=VIRE (accessed Aug. 21, 2023)

responsible architect’s workflow: in a nutshell, a compre- further adjustments) quality addresses man when
hensively sustainable built environment develops along browsing the output set gained this way?
a grid-and-grain public space structure and not vice
versa. The public space - streets, squares, parks, places, 4 Discussion
public amenities areas, ... - has to be designed carefully, AI is a super-parrot: it is superb in repeating what it
responsibly, considerably, and poetically first, to adopt has learned, explains Tomas Mikolov in a chat with
particular buildings only after (Sourek, 2014). However, Dan Vavra (2023). As already mentioned, concerning
this is a process that Spacemaker not only does not sup- AI, learning or training is the key word; for an AI per-
port but also does not allow. formance, the magnitude and comprehensiveness of the
AI applications to generate floorplans and apartment training dataset is the starting point, the algorithm is the
layouts emerge to be ambivalent when it comes to the method or, running on an artificial neural network, the
effectivity and practical usability of their outputs. The tool respectively, and the computational performance is
quality and usability of the deliverables seem to be sim- the limit.
ilar to the performance of creative applications such as
DeepHimmelb(l)au: hundreds of options are delivered 4.1 Poiésis: architectural design beyond and within AECO
to ease an architect’s task when taking over a load of ecosystem
mechanical generating various options, only to be con- From architecture and urban design over construction
sidered by him finally (Chaillou, 2022h); nonetheless, a and MEP (Mechanical, Electricity, Plumbing), environ-
vast majority rather all of them appearing prematurely mental, climatic, meteorological, and microclimatic
published if not useless when reviewed. On the other expertise to transportation expertise, economy, demog-
hand, when furnishing with furniture prepared layouts, raphy, and sociology, multiple professions engage in the
more satisfactory results emerge: as a rule, a generative development of the built environment. The background
AI tool shows more capable than a conceptual one. The of some of the fields is natural sciences while, for the oth-
choice of the (most) suitable of the options generated ers, it is social sciences or even arts – poetics or poiésis
by AI shall remain reserved to the human architect as (Heidegger, 2000), as will be explained soon. According to
“the touch of master’s brush” while the automatic and the nature of the contribution provided by the respective
prompt delivery of “all thinkable” options saved his expertise, the design and evaluation approaches range
time and energy. However, what is the factual contribu- from “hard” to “soft” ones, from quantitative and mate-
tion of AI if no result of acceptable (without substantial rial to qualitative and emotional ones. According to such
Sourek Architectural Intelligence (2024) 3:4 Page 11 of 21

an origin and nature, quantitative parameters define the consciousness; not by accident. Next to nature it is
approach as well as the output in some cases, while it is architecture that creates the world of human existence.
(close to) feelings or moods in others. Obviously, feelings In the essay Poetically Dwells Man (Heidegger, 2000),
and moods resist following parametric algorithmization elaborating further his seminal opus Being and Time
as well as entering datasets. Consequently, a software (Heidegger, 2006) and the theme of Dasein - being-there
ecosystem, artificial neural networks not excluded, inevi- or existence in English - after the Second World War
tably fails to elaborate feelings and moods - as opposed in relation to the timely and pressing topic of housing,
to quantitative magnitudes and performances (Barker, architecture by extension, Heidegger coins the con-
2023). cept of das Geviert - the fourfold in English - the union
Approaching architecture as the most significant of the earthly and the heavenly, the human and the
among the built environment creators, let us be clear: it is divine in man’s existence and in the world of his being
not a natural science scheme, algorithm, or calculus that - thus, as we have seen, in architecture. This is not only
is the architecture’s starting point. Moreover, it is not a another strong argument refuting the vision of archi-
linear sequence of signs - opposed to speech or text. On tecture created by an algorithm. It is no coincidence
the other hand, among many other attributes, architec- that materiality manifests itself in both consciousness
ture can be consumer goods, too; and the more a con- and architecture: materiality manifests itself in them in
sumer goods a practical architecture shall be, the more the same way and is a strong link between them. This
a pattern, a calculus, and an algorithm contribute to the recalls the dual nature of architecture - of ideas, emo-
delivery; but even then, the environment, the narrative tions, and experiences on the one hand and material,
of the development, and/or the people passing, enter- physical on the other - that slowly-slowly begins to lead
ing and using the building or the structure „make the to uncovering the feasible way of deployment of AI in
difference“. The theory of public space puts it clearly: As architecture and grasping its prospects.
soon as and only exposed in public space, a construction Dalibor Vesely (2004) featured and reviewed critically
becomes architecture (Sourek, 2014). In theory, architec- another face of architecture’s duality starting in the head-
ture unanimously distinguishes from arts. But even so, ing of the groundbreaking book Architecture in the Age
even when architecture shall not be an art like painting, of Divided Representation The Question of Creativity in
sculpture, drama, dance, or literature, let us not be shy: the Shadow of Production (2004). Creativity never can
It is poetics or poiésis as Martin Heidegger coins in antic be substituted by production; however, the material side
Greek that is the starting point and method of architec- of architecture - its physical properties both in terms of
tural creativity. Poetically dwells man, puts it Heidegger microclimate convenience, durability, security, ergon-
(2000): full of merit, yet poetically dwells a man. Poiésis omy, operational efficiency, and sustainability - deserve
precludes algorithm and vice versa, and similarly, a train- and are keen to enjoy productivity - productivity, that is
ing dataset limits poiésis. By definition and due to practi- parametric and algorithms-inclined by nature.
cal reasons, a dataset can never be comprehensive. Then, So far in the field of AI in architecture, as in the whole
it cannot but limit the creativity for which, inevitably, the AECO field, however, all the time only analogical, para-
training dataset is „the whole world “– there is nothing metric-oriented approaches have been witnessed (the
beyond. differences between diverse neural networks and AI
Also, Encyclopedia Britannica distinguishes and con- algorithms, as outlined in (2) make no difference in this
firms the emotional, social and societal, non-paramet- regard). Tackling data by a computational algorithm can
ric nature of architecture, … the art and technique of provide poetics only by chance and randomly. It is not a
designing and building, as distinguished from the skills question of learning or training; by definition, a poetic
associated with construction. (Gowans et al., 2023) The “output” cannot be trained. Even if bokeh salience offers
characteristics that distinguish a work of architecture a “hallucination”, it’s not poiesis nor a creative act; it’s
from other built structures are (1) the suitability of the just a random interpretation of training data that we only
work to use by human beings in general and the adapt- additionally realize it was misleading. In a conclusion,
ability of it to particular human activities [and needs], the idea of a creative contribution of AI to conceptual
..., and (3) the communication of experience and ideas architectural design is debunked; and together with it the
through its form. Obviously, “use by human beings”, theoretical collateral and all the AI’s outputs in the field
“human activities and needs” as well as “communication so far. On the other hand, debunking the vision of AI or
of experiences and ideas” cannot but resist algorithmiza- an AI’s “superuser” replacing “the architect genius” (Leach,
tion as well as digital parametrization. 2018b) as erroneous should not prevent algorithmizing
Among all types and natures of creations by and machine-generating what fits; and this is the physical
humans, architecture intertwins the most with human aspect of architecture.
Sourek Architectural Intelligence (2024) 3:4 Page 12 of 21

4.2 When AI works so far, is first a drive to set and maintain a comprehen-
Opposed to fine arts, literature, poetry, dance, or drama, sive list of all such parameters and second an approach to
whose production is only consumed, architecture is parametrize, quantify, and evaluate objectively in a com-
always also used. This is not a denial of the poetic essence parable way the hitherto overlooked parameters. Perhaps
of architecture (recalled in previous paragraphs), this is feasible previews and assessment calculation procedures
just a remark on the complex nature of architecture. Then in these respects have been missing so far; however,
inevitably, two realms of architectural design and a plan equipped with the knowledge of AI, paradigms of its
to build a building can be identified and distinguished: deployment, and its potential in terms of data quantities
The first one comprises properties and performances and their processing both the inadequacy admitted and
that concern (even though not exclusively) the (mate- the two shortcomings can be overcome.
rial) use, while the other delivers poiésis, poetry, mood, Distinguished explicitly from architecture, the con-
excitement, or experience. The interface between the struction is another story: parametric, “mathematic”
two realms does not match the interface between archi- by nature, the mimetic, imitative creativity of design-
tecture, as characterized by concerns to use by humans, ing constructions welcomes algorithms and parametric
to human activities and needs, and the communication patterns. Such is the starting point for the excellence of
of experiences and ideas by its form, and construction generative AI software systems, their leading computa-
that materializes the architecture. Set by architecture, tional approaches being optimization and optioneering,
the spatial structure of a building gives the ergonomy analysis and simulation. The examples of most advanced
and efficiency of movement within the building: it is the aplications are ETABS, SAP 2000, STAAD PRO, RAPT
architectural design, not the construction solution that structural engineering software, SCIA Engineer (2023),
determines these material, quantitative parameters of the or Tribby3D (Salehi & Burgueňo, 2023; Tribby3D, 2023).
building. The structural design community, surprisingly at the first
Anything material can be parameterized, anything par- sight, compared to many other expert fields, restrains
ametric can be quantified, and anything quantitative can from (over) using AI. As shown in (2), the interpret-
be compared and evaluated objectively - or at least (very) ability issue is a natural reason. Generative AI may be
close to objectively. And this is the case with a large part an approach deployed in multiple structural engineer-
of an architectural design, a proposal of a building or an ing software tools “from time immemorial”, however,
enclave of the built environment. Parametric properties a black box must not have a final say when it comes to
of architecture are subject to creativity as well, but, to responsibility, such as in the case of structural design. So
another, one can say the pettier nature of creativity: to the far, rule-based models are proving indispensable in this
mimetics or mimesis, creativity stemming from imitation, regard. The vital reliability of design tools for structural
as opposed to the primal creating of poiésis (Heidegger, engineering is evidence of awareness of the risks that AI
2006). The parameters-oriented approach certainly does algorithms hide and the knowledge of how to tackle them
not equal AI; however, this is the field of AI - a perfor- without giving up the wide possibilities and fundamen-
mance of a machine conditioned by training, which is tal opportunities of applying AI in parametric generative
nothing else than mimetic elaborating of templates. This, design and solution optimization.
naturally, can be the realm of feasible deployment of AI Mimetic, too, is urban design, its supportive disciplines
in architecture and AECO in general: not a poiétic crea- being parametric by nature. Examples of application of
tion by which a computer could replace an architect. AI in the field have been overviewed in (3) zooming in on
Concerning the quantitative, objective and compa- tools such as Spacemaker, Creo, or Cove.tool – in terms
rable assessment of the complex of diverse phsical per- of both succesfull and contributing use of AI and of mis-
formances of buildings´ and built environment’s designs conceptions. The approach represented by Spacemaker
– such as operational and energy efficiency, acoustics, to design development of a tool shows “an embodiment”
ergonomics, daylighting, and other physical benefits that of a problem that tends to become general and affect
architecture provides to man, community, and society – many AI tools in the development of an initially viable
the state-of-the-art performs mature tools related to par- and promising concept. The problem turns out to be an
ticular parameters. This is what software applications like opening of the scissors between the IT line and the user,
Cove.tool, Creo, Giraffe, Spacemaker, the applications i.e. the designer line of the tool development resulting
used by MVRDV as outlined further in this chapter, and in deficiencies concerning the workflow, starting points,
many other tools already introduced and proven in archi- and principles of designing.
tectural and planning practice deliver, though not always At this point, Spacemaker “got spanked” for many
distinguishing between physical architectural respects other AEC software tools that are parametric and
and respects of the construction. What has been lacking shorthand imitative by nature and yet they are pushed
Sourek Architectural Intelligence (2024) 3:4 Page 13 of 21

to architects as creative tools, which is not so rare in Awarded the best skyscraper of 2022 in the Empo-
today’s practice. Fortunately, better cases have been ris Award competition, the MVRDV Valley at the South
witnessed - also in the deployment of Generative AI. Axis, the central business district of Amsterdam, is a
MVRDV, Dutch by origin, today a global architectural showcase of successful AI technologies deployment
studio, shows up as a successful pathfinder in terms alongside authentic architectural creativity. Machine
of AI use and development. In response to the need analyses allowed for developing a rich, truly sculptural
to push the limits of technological possibilities for form and maximizing the efficiency of the land’s and
the sake of innovative architecture, MVRDV NEXT - space’s use while ensuring generous sunlighting and day-
shorthand for New Experimental Technologies - was lighting of all apartments and providing views and liv-
founded in the 2010s as an internal startup. Headed able garden terraces to them. In planning the project, a
by one of the studio’s partners Sanne van den Burgh, a Grasshopper script optimized the architectural form and
group of in-house specialists develops and implements detail to make the construction economical and efficient
computational workflows and new technologies. Through and to provide for sustainability thus. Alongside the com-
a mixture of project-based work and standalone com- prehensive use of information technology to analyze the
putational research, they rationalize designs and setup tasks and the opportunities and to support and stream-
configurations, unlock potentials on an urban and par- line the creative design process, the rigorous avoidance
ticular buildings´ scale, optimize workflows, speed up of the terms AI and ML in the studio’s communication is
processes, and make projects more efficient and adapt- notable (van den Burgh, 2022).
able in the face of change. Represented by projects such
as HouseMaker, VillageMaker, The Vertical Village, 4.3 Generating by patterns
Barba, Space Fighter, or Porocity, and site specifically Opposed to the “sky is the limit” architectures whose
Rotterdam Rooftops or FAR MAX for Bordeaux, their form is often pre-defined neither by existing neighboring
methods allow the studio to explore a future that is structures nor by short-term financial perspectives and
equitable, data-driven, and green (MVRDV, 2023). Spe- approaches, there are architectures – buildings designed
cific, in-house developed applications allow for Design and constructed according to given spatial conditions,
Optioneering using generative analysis and automated terms of future usage, and strict economic templates.
form-finding to expand options and drive design possi- In fact, this is the case with the vast majority of archi-
bilities (as Fig. 3 illustrates), Rationalization standard- tectures - which, nevertheless, neither diminishes their
izing complex geometries to maximize design ambitions importance nor makes the role of their architects less
and feasibility, Performance Evaluation employing sim- responsible and demanding. The vast majority of what is
ulations and analysis tools to evaluate the impact and being designed, planned, and eventually built to saturate
minimize risks, or Advanced Geometries managing com- the needs of a growing population and living standards in
plex modeling tasks across design tools to facilitate effi- terms of dwelling – it is residential buildings - work, and
cient, fast-paced processes. production – from office buildings to production objects

Fig. 3 Solarscape by MVRDV NEXT: an in-house developed Grasshopper/Raytracer-based simulation and negotiation tool to find high-rise areas,
sunspots potential development opportunities; case study Rotterdam. https://www.bing.com/videos/search?q=mvrdv+next+solarscape&qpvt=
mvrdv+next+solarscape&view=detail&mid=35848FD4186FEE4F94A035848FD4186FEE4F94A0&&FORM=VRDGAR&ru=%2Fvideos%2Fsearch%3Fq%
3Dmvrdv%2Bnext%2Bsolarscape%26qpvt%3Dmvrdv%2Bnext%2Bsolarscape%26FORM%3DVDRE (accessed Aug. 21, 2023)
Sourek Architectural Intelligence (2024) 3:4 Page 14 of 21

- transport and logistics – among others logistic com- compared to the intuition of the (human) architect.
plexes and storage facilities - and many other buildings´ Others, like Immanuel Goh or Andrea Banzi searched
typologies falls into the category of mass production and, for explicit rules-scripting-based design generators
kind of, consumer goods. Such a categorization does working with inferred rules drawn from the dataset of
not challenge the contribution of the respective authors, samples. Not a patterns´ assessment and appropriation,
designers, and planners in terms of “creativity used”, craft, but recognizing the internal logic of the pattern, and
and efforts. Many such architectures launch their way to then extrapolating a broader design based on that logic
existence in architectural competitions - formal and non- that could potentially continue forever ... (Leach, 2018d)
formal - and not a few of them get their “five minutes of in reality failing to contribute to the design practice
fame” in architectural websites, magazines, and exhibi- eventually.
tions; nevertheless, they remain a “stardust”, a sort of Most recently XKool, an AI startup in Shenzhen,
“no name” (except for specialist history scholars or local China, developed a web-based platform for using
patriots); not to make anybody offended, let us label them AI across a range of tasks from architecture to urban
“production [ones]” - production architectures and pro- design (Crunchbase, 2022). Though not-so-easy to be
duction architects. However, such architectures make the used practically or to be tested by non-Chinese resi-
complex performance of the built environment in a con- dents (XKool, 2022), the approach of the studio and
sequence: more than 90% of the performance in terms of results achieved so far by the application awaken hope
environmental impacts, sustainability, and macro econ- to overcome the lack of attention to the immense rich-
omy, but also the majority of the performance in social, ness of patterns provided by the existing building stock
cultural, and economic terms. It is not the architectural and design representations. XKool appears the most
icons but these production architectures that the entire efficient of all AI applications for architectural design,
population is exposed to on a day-to-day basis – at home, streamlining the design process and making it more
at work, at school, at leisure and social activities, at com- efficient in terms of both analyzing a vast range of pos-
muting, at going in for sports and recreational activities, sibilities and generating designs (or pre-designs, more
and at walking pets as well as at tackling the household accurately said) based on samples – to evaluate and
budgets. In terms of design and planning, the obvious return the most suitable outcomes, and, moreover, to
richness of examples and models may balance the com- develop them further according to the given constraints
plexity of multiple limits and constraints that intervene (Leach, 2018e). The way of working is revolutionary -
in the related design processes; however, most often, it and no worry that the outcomes do not look very novel
does not make a creative architectural approach redun- as a rule: the core is it copes with the “consumer goods”
dant or expendable. And it demonstrates the importance characteristics of the design category.
and potential contribution of comprehensive research The mission to challenge “the architect genius” that,
and analysis of the huge volume of existing samples and so far, has been the motivation behind the efforts to
inspiration - which, in reality, is far from being carried develop an AI-based design tool (almost) as a rule, shows
out comprehensively, if at all. At the same time, even if debunked by XKool approach and results achieved.
proceeding only from the knowledge of AI and its possi- In general, a new approach emerges consisting of AI
bilities and limitations provided by this paper, it is no less “designing” by - first - delivering pre-designs, it is solu-
evident that such research and analysis fits AI as much as tions close to set parameters - as close as the available
possible. patterns allow, and then - second - “assisting” the human
Little is more overlooked by recent and current designer in adapting the pre-designs, customizing the
efforts in the field of AI than this opportunity and chal- final, specific solution; the nature of the “assisting” is
lenge - both in architectural and planning practice quantitative, parametric assessment, feedback concern-
and in research and tools and processing standards ing the goals and result assessment including finding the
development. system of criteria, specification of the particular crite-
The “almost consumer goods” characteristic evokes ria, and evaluation criteria sets that AI can develop and
approaches deploying algorithms (what else should be complete continuously. The patterns-oriented approach,
more attractive for AI?!), parameters pre-definition, and when confirmed and developed, and developed the pat-
patterns in the design process. One of the first authors terns stock – libraries of parametric examples, represen-
and researchers active in this field was Makoto Sei tations of solutions existing so far - promises to bring a
Watanabe already in the 1990s (Leach, 2018c); however, paradigm change within AI in AEC that, consequently,
having focused on machine-aided design rather than on could find the path to an AECO paradigm change needed
analyses and the use of patterns, he remained unsatis- both by the AECO as well as the community of its clients
fied with what AI was able to deliver in terms of design (which includes the whole population in the end).
Sourek Architectural Intelligence (2024) 3:4 Page 15 of 21

4.4 Advice whispering the results of their previous work costs them nothing;
Not a layout creation but “sampling” of generative pat- the whole field makes a profit developing quicklier and
terns = already existing solutions, selected by AI as the better, a more efficient way based on the joint efforts of
most suitable not only in terms of floor-plan or/and spa- all memebers of the community. The perspective of the
tial solutions but in terms of structural solutions, too, benefit of free approach to the existing solutions – in par-
appears the key. Based on the given goal parameters and ticular parametric representations of architectures both
constraints, an adaptation (human, though AI supported) built and only designed - appears an incentive for recon-
of selected patterns follows. sideration current approaches in terms of architectural
An ability to infer the properties of the solution to design – and whole AECO, too.
which the design development is heading – whether
led by a human or AI – stemming from the experience 4.5 R&D anew
gained in learning on a set of solutions is natural both to After a decade of “challenging the human architect,” the
GANs and VAEs. The predictive inference can be avail- true potential of AI in AECO has only just revealed; R&D
able starting from the earliest phases of design – from the at the threshold begins to specify problems and solve
first sketch in terms of how a human drafts and develops tasks. The results achieved so far by Stanislas Chaillou,
a design. AI can go conveying the inference continuously Nvidia, and others, mentioned in (3), show that the so-
in a way we can call whispering, providing the designer – far-ruling principle of lossy compression and subsequent
human as well as AI – with comprehensive feedback on “creative” decompression within the supervised (or unsu-
his or its design decisions and heading of the design. This pervised) learning has exhausted its possibilities without
way, the design will be optimized not in the mode try – being able to deliver truly usable results.
error – correction – another error – another correction I believe a statistical approach to design conception
– and so forth till the designer is satisfied with the feed- will shape AI’s potential for architecture. … Pix2Pix uses
back parameters, or too tired to continue trying, which is a conditional generative adversarial network (cGAN) to
the state-of-the-art today, but continuously. The effect in learn a mapping from an input image to an output image.
terms of time and cost spent, and quality of the solution … to learn image mappings which lets our models learn
achieved is obvious and huge; AI can never beat a human topological features and space organization directly from
when it comes to true creativity – but no human intuition floor plan images. We control the type of information
and experience combined compares to AI when it comes that the model learns by formatting images. As an exam-
to parametric quantitative assessing and review. Here we ple, just showing our model the shape of a parcel and its
go to the future of AECO. In essence, it is about utilizing associated building footprint yields a model able to create
the relevant knowledge, talent, and efforts of the entire typical building footprints given a parcel’s shape, Chail-
community of architects and the computational force of lou (2023) describes the supervised-learning strategy that
AI combined. Naturally, both the selection and adapta- reduces the essentially comprehensive architectural task
tion processes interweave with outcomes and adaptation to image processing. Leaving aside hundreds of thou-
solutions evaluation in terms of microclimate qualities – sands of images, each of them labeled by humans, that
daylight and sunshine, or temperature stability – energy are a precondition for the “statistics” to work properly,
efficiency and consumption, acoustics, as well as area the strategy has proven a no-go in terms both of the com-
capacity and other qualities of the solution in the process petent AECO workflows and computing power. When
that addresses sustainability comprehensively in terms of trying for a sustainable and feasible AECO design, images
both environment, social aspects, and economy. of pattern layouts cannot do; the parameters of the spa-
Nonetheless, challenges remain: first (and above all), tial structure of the proposed building, the parameters
how to access the immense sum of the preceeding archi- of the physical properties of its constructions, and finally
tectures records when a paradigm of protecting „the the parameters of the internal environment in the object
know-how “by hiding the representations of the archi- must come into „consideration“. Given state-of-the-art
tecture designed exists. In this respect, the approach of ML, the size of a comprehensive training dataset for the
the AECO community contrasts the approach of the IT presented approach should (significantly) exceed the Nth
developers community. IT developers are used to provid- power of two, where N is the number of parameters to
ing each other with their achievements in widely shared specify the ML task: thousands rather than hundreds of
libraries; Github (2023), Gitlab (2023), or Patternforge parameters when it comes to the comprehensive para-
(2023), and many others are the platforms. Who makes metric and physical structure that materializes architec-
the profit are not only particular IT developers that can ture - a building. Even if it were “only” lower hundreds,
fulfill their tasks and achieve goals more quickly, with the number has a hundred and more zeros - a googol: the
less effort, and for lower cost, whilst making available question of computing power - or rather the optimization
Sourek Architectural Intelligence (2024) 3:4 Page 16 of 21

of the structure of the parameters - is immediately raised chess or Go on the masters level: deployment of analogi-
when googol exceeds the estimate of the number of ele- cal algorithms deserves to be studied thoroughly first.
mentary particles in the known universe.
First, fundamentally new types of learning networks are 5 Conclusions
needed, recent supervised – and reinforcement schemes Diverse current and recent attempts and successes to
have proven no more suitable both in terms of the approach the deployment of AI in architecture and, more
required performance and in terms of the working prin- broadly, in AECO have been discussed in previous chap-
ciple. Instead of „good old GANs“, imitation-based learn- ters. The question of the particular field of AI’s deploy-
ing, self-learning, and knowledge-seeking agent schemes ment that would be not only an interesting thesis for an
that focus on the (design) process - “how things come to academic- or allowance-trying but a helpful and feasible
existence“ - instead of output - “how things shall be” shall tool or working environment comes clear as the key; at
be surveyed and developed concerning ML-aided AECO. the very core, queries and issues render.
As mentioned in (2), dataset aggregation, behavior clon-
ing, inverse reinforcement learning, soft Q-learning, 5.1 Can AI be truly creative?
inverse Q-learning, self-imitation and transcendence, Unless computers gain consciousness, there is an une-
and others represent the field of hopefully eventually effi- quivocal answer to the question: no. Several reasons have
cient R&D of ML in AECO. emerged in previous chapters. In this place, the inevitably
Second, training datasets - predicted open source only mimetic aspect of the way a computer is able to work
platforms pose questions on materials assembly, mate- can perhaps close the recent and present attempts even-
rials quality, and size. A pragmatic optimization of the tually that cannot be but futile. Creativity, to be authen-
involved parameters structure appears a key task. The tic and true, cannot be but poetic (Heidegger, 2006). The
question of the data format with which the algorithm will poetic principle requests consciousness together with
work is crucial: it seems obvious that it should be one of intention: only consciousness together with intention is
the BIM formats. The basis of the algorithm structure able to deliver poiésis. In terms of architecture and built
could be - it seems - a pair of mutually interfering loops: environment, consciousness is reserved for a man, or,
a generative loop and an advice-whispering loop, or there more precisely, to Dasein, as Heidegger coined a proved.
can be more advice-whispering loops particularized An algorithm, however complex and sophisticated is the
according to the diverse natures of the parameters, which artificial network it works on, can deliver only based on
will be “switched-on” only in a cascade. In the beginning, the principle of equality (or similarity, which, however, is
a suitable pattern will be selected from the database, only a deficient mode of equality) or by random choice.
which will be tested and optimized due to the specified Face-to-face to new solutions, advance knowledge is the
outlines and concerning a benchmark of independent prerequisite. Prior knowledge is another aptitude reserved
parameters. for consciousness - to a human, not to a machine, and
Human-in-the-loop can be expected to be fundamental not to an algorithm. No consciousness, no own will,
together with an unprecedented streamlined generative and no true creativity, but algorithms and immense data
nature of the algorithm. New classes of imitation-based searched through, assessed, and prioritized according to
learning, learning a behavior policy from demonstra- the defined criteria are the attributes of today’s AI. And
tion, and self-learning paradigms zooming in on the even the state-of-the-art theory does not show a vision of
design-development processes instead of the results (to how machines could overcome the shortcoming.
be) achieved must be welcome to “customize” the most
suitable pattern to the requested final proposal. Starting 5.2 More openly articulated the question: can AI
from following the human-in-the-loop in the phase of contribute directly to how authentic architecture
“customizing”, the algorithm shall learn by self-training comes to existence?
to master the design process better than the man in the Yes; the more poetic creativity is excluded, the better
end - hopefully. To take up such a challenge in the archi- the mimetic, imitative AI approaches fit the parametric
tectural design and building planning realm, lessons from nature of the physical and quantitative aspects of archi-
the in (2) reminded trailblazers such as AlexNet, Deep tecture, not to mention construction and other features
Blue, and especially AlphaGo Zero have to be learned. of buildings and development of the built environment
The evolutionary algorithms approaches and genetic such as energy efficiency, construction cost, environmen-
programming have to undergo a deep survey to be sub- tal footprint, durability, and others, all discussed in (4).
sequently considered as an option. Such a concept ought Adding parameters of using the built environment by
not to be refused or underestimated pointing out the humans such as economy and efficiency of layout, ergo-
gap between designing production buildings and playing nomics, and others, the parametric realm represnting the
Sourek Architectural Intelligence (2024) 3:4 Page 17 of 21

physical side of architecture becomes complete that can projects – first of all, the spatial layouts – in sufficient
be regarded as a domain of AI. When it comes to AECO, numbers, with a richness of volume and quality, and
AI can assess, quantitatively evaluate and compare, and accessibility and transparency for search engines – is an
(pre)design, too, everything except for the sphere of obvious prerequisite for the generative, pattern-based
poiésis, poetry, mood, excitement, or experience – as AI-aided design in AECO.
discussed in (4) as well. As a principle, the performance Second, however great is the so far neglected benefit
of AI in this regard is able to outperform any relevant of learning from patterns, inviting new learning para-
human performance in terms of complexity, scope, accu- digms such as in (4) reminded imitation-based learn-
racy, pace - and cost, of course. ing, learning a behavior policy from demonstration, and
self-learning into the design process emerges the ulti-
5.3 AI in architectural design: the workflow and its mate and most promising challenge. After having learned
execution from learning pioneers such as Deep Blue or AlphaGo
Shown in Fig. 4, the workflow of an AI-aided AECO Zero and having „studied “how the man-in-the-loop
design comprises two phases (as discussed already in has adapted the best fit pattern to a desired new solu-
(4): first, processing generative patterns to a pre-design, tion, computers may eventually take over the routine
a solution as close to the set parameters as the stock of of “mechanical” design- and design-analyze tasks and
generative patterns allows, and second, final adapting deliver room for the architect’s creativity. Obviously,
the pre-designs, customizing the final, specific solution. designing production buildings is not as challenging as
For the first phase, patterns, their stock in the form of playing chess or Go on the masters level. However, it
open-source platforms, and AI-driven search algorithms does not put contrary to deploying analogical algorithms;
to identify (the most) suitable cases/patterns by inferring the advantage, in addition, can be that the imitation-
what fits (better and what less) are the keys, in the sec- based- or self-training would not need to be so intense
ond phase it is (AI-driven) design development support to start to deliver useful results in designing production
in a form of advice whispering and continuous and com- architecture.
plex assessing and “feedback” as outlined in the previous Along with the development of AI’s deployment in
paragraph. AECO design and planning - inherently comprehensive
The formation of open-source platforms of generative disciplines developed in teamwork, as a rule, today - new
patterns – parametric representations of essential fea- roles within the design- and planning teams will emerge:
tures of existing solutions – existing buildings or mature among others, „superuser “tackling the AI, an architect

Fig. 4 AI-aided architectural design workflow

Sourek Architectural Intelligence (2024) 3:4 Page 18 of 21

with a strong IT background or an IT expert with a has proven able to extrapolate rather successfully the past
strong AECO background. However, the „superuser “will processes into the future – to predict the future - not 100
replace the leading architect neither in his conceptual % but still with unprecedented success. Thus, we need
role nor in aesthetic respects. The „superuser “will econ- to deploy AI to first extrapolate a training pool based on
omize the leading architect’s efforts and energy in favor history and current knowledge, for AI-driven specifica-
of indispensable creativity. The role will comprise com- tion and quantification of threats “in the second loop.”
petencies and responsibilities that render to be not so far
(though undoubtedly distinct) from today’s BIM (Build-
ing Information Management) Coordinator. 5.5 Changing the game
A reminder of the lagging of AECO behind global societal
5.4 Design reviews, evaluations of solutions and economic development opened this paper. Previous
and the security issue chapters and paragraphs reveal the potential of feasible
Comparing to design processes, AI-led design reviews deployment of ML in these fields: an immense potential
and evaluations of solutions show to be a sort of business of diverse but interconnected ways of deployment that
as usual, no more a (basic) research and experimental not only contrasts with the so far wavering approach to
development. In (3), several existing applications of this the new technology but can provide essential contribu-
nature have been listed - Cove.tool, Creo, Spacemaker, tions to responding to existing social, cultural and eco-
...; and many others exist. However, the field is far from nomic, and environmental challenges.
being covered. Together with addressing others, so far The first to change is the paradigm of designing archi-
sidelined attributes of the design solutions, the qual- tecture and the built environment, and the whole AECO.
ity of outputs delivered will be welcome to have raised. The new paradigm will replace the “traditional” infinite
Another question is the comprehensiveness of the assess- spiral of proposal and error - assessment - a new proposal
ing and reviewing. The branch can develop separately - as and error - and so forth. Once the ML tools - algorithms,
is most often the rule so far - or can integrate into the AI- models, platforms - as outlined in previous chapters
aided-design environment - in a form of in (4) outlined apply appropriately and as intensely as it deserves in the
advice whispering. parametric realm of architecture and development of the
However, given the issue of (lack of ) interpretability built environment, the process of designing - creativ-
(addressed in (2)) and the state-of-the-art of the field, it ity unrestrained, on the contrary, unleashed - becomes
is not AI that may have the last say. An ongoing advice not only significantly more efficient and productive and
whispering that leads the designer to a benchmark - and less time- and cost-consuming, but consequential and
sometimes “hallucinates” him based on the training guid- objective, too. The impacts on a comprehensive quality
ance in the saliency heat map - is one thing, and another of architecture and the built environment show obvious,
is the final inspection that establishes liability. notable, and positive. The shift in terms of the quality of
Comprehensiveness in assessing and reviewing is the service and deliverables within the AECO professions
another issue. Accidents, crashes, and even catastro- gives a clear perspective of optimization of the outputs,
phes, which could have been prevented if only the rel- which is likely to significantly exceed not only today’s
evant aspects had been taken into consideration, are practice but also any expectation. Poetic creativity that is
lining the history of the development of the built envi- essential for authentic architecture will enjoy not a direct
ronment – and mankind. To minimize the occurrence contribution but a subsidy through creative energy, atten-
and impacts of such events in the future, potentially tion, and capacities released by the deployment of ML’s
threatening aspects shall be identified before they occur. mimetic capabilities in the parametric realm.
Lessons taken from health care (Christian, 2020e), AI The perspective is realistic; it is not a task for the dis-
shows ready to give a hand in this regard, too: the task tant future - the implementation can start now and the
is - as usual - first, to set up the training dataset, second, MVPs (minimum viable products) can be there in les
to apply suitable algorithms to specify the risks and their than half a decade. The motivation is the economy in
parameters, and third, to develop and apply adequate terms of the efficiency of development of the built envi-
technical measures. Today’s engineers are up to the third ronment, a business opportunity in terms of filling a mar-
challenge, and computer scientists and developers are up ket niche, comprehensive quality of our lives, and our
to the second. The history of the built environment and sustainable future that will become boldly resilient - sus-
mankind’s (and the biosphere’s, too) survival provides for tainable beyond today’s state-of-the-art efforts.
the training dataset. However, only for a part of the train- Acknowledgements
ing stock; as a matter of fact, taking lessons from the past The author expresses no acknowledgment related to the manuscript
elaboration.
is what repeatedly shows insufficient. Today, however, AI
Sourek Architectural Intelligence (2024) 3:4 Page 19 of 21

Authors’ contributions Campbell, M., Hoane, A. J., Jr., & Hsu, F. (2002). Deep Blue. Artificial Intelligence.
Michal Sourek is the only author author of the article. Elsevier, 134(1–2), 57–83. https://doi.org/10.1016/S0004-3702(01)
00129-1
Authors’ information Chai, L., Wulff, J., & Isola, P. (2023). Using latent space regression to analyze and
Michal Sourek, professor of architecture and civil engineering at the Faculty leverage compositionality in GANs. Github. https://chail.github.io/latent-
of Civil Engineering, Czech Technical University in Prague, Czech Republic, composition/ (accessed Feb. 12, 2023)
and practicing architect – principal of Prague, Czech Republic-based studio Chaillou, S. (2022a). Artificial intelligence and architecture from research to prac-
MS architekti. Author of multiple implemented buildings, urban- and land tice. Birkhauser Verlag GmbH. ISBN 978–3–0356-2400-7, pp. 112–146
development projects, and awarded competition proposals. He focuses on Chaillou, S. (2022b). Artificial intelligence and architecture from research to prac-
public space and prospects for deploying information and communication tice. Birkhauser Verlag GmbH. ISBN 978–3–0356-2400-7, pp. 80, 86–89
technologies in architecture and the development of the built environment. Chaillou, S. (2022c). Artificial intelligence and architecture from research to prac-
Author of multiple papers on the theory of the fields of his concern published tice. Birkhauser Verlag GmbH. ISBN 978–3–0356-2400-7, p. 90
in international journals. Chaillou, S. (2022e). Artificial intelligence and architecture from research to
practice (pp. 23–48). Birkhauser Verlag GmbH. ISBN 978–3–0356-2400-7.
Funding ibid. f, p. 98
Elaborating on the manuscript was supported by no funding. Chaillou, S. (2022f ). Artificial intelligence and architecture from research to prac-
tice. Birkhauser Verlag GmbH. ISBN 978–3–0356-2400-7, pp. 83–86
Availability of data and materials Chaillou, S. (2022g). Artificial intelligence and architecture from research to prac-
The author does not analyse or generate any datasets, because his work tice. Birkhauser Verlag GmbH. ISBN 978–3–0356-2400-7, pp. 103–109
proceeds within a theoretical approach. Chaillou, S. (2022h). Artificial intelligence and architecture from research to prac-
tice. Birkhauser Verlag GmbH. ISBN 978–3–0356-2400-7, pp. 80, 86–92
Chaillou, S. (2023). ArchiGAN: a Generative Stack for Apartment Building Design.
Declarations NVIDIA DEVELOPER Technical Blog. Jul. 17, 2019. https://developer.
nvidia.com/blog/archigan-generative-stack-apartment-building-
Ethics approval and consent to participate design/. Accessed 29 Mar 2023
This manuscript has no associated data. Christian, B. (2020a). The Alignment Problem: Machine Learning and Human
Values. 1st ed. W.W.Norton. ISBN 9780393635829, p. 206
Competing interests Christian, B. (2020b). The Alignment Problem: Machine Learning and Human
There is a single author of the manuscript. Values. 1st ed. W.W.Norton. ISBN 9780393635829, p. 209
Christian, B. (2020c). The Alignment Problem: Machine Learning and Human
Values. 1st ed. W.W.Norton. ISBN 9780393635829, p. 222
Received: 12 June 2023 Accepted: 19 December 2023 Christian, B. (2020d). The Alignment Problem: Machine Learning and Human
Values. 1st ed. W.W.Norton. ISBN 9780393635829, pp. 240–243
Christian, B. (2020e). The Alignment Problem: Machine Learning and Human
Values. 1st ed. W.W.Norton. ISBN 9780393635829, pp. 28–64
Christian, B. (2020f ). The Alignment Problem: Machine Learning and Human
References Values. 1st ed. W.W.Norton. ISBN 9780393635829, p. 126
AlphaGo. (2023a). Google DeepMind. https://www.deepmind.com/research/ Christian, B. (2020g). The Alignment Problem: Machine Learning and Human
highlighted-research/alphago. Accessed 21 Jan 2023 Values. 1st ed. W.W.Norton. ISBN 9780393635829, pp. 241–243
AlphaGo. (2023b). Zero: Starting from scratch. Google DeepMind. https:// Christian, B. (2020h). The Alignment Problem: Machine Learning and Human
www.deepmind.com/blog/alphago-zero-starting-from-scratch Values. 1st ed. W.W.Norton. ISBN 9780393635829, p. 109
(accessed Feb. 1, 2023) Christian, B. (2020i). The Alignment Problem: Machine Learning and Human
Ayush, T. (2023). Understanding Reinforcement Learning from Human Feed- Values. 1st ed. W.W.Norton. ISBN 9780393635829, p. 114
back (RLHF): Part 1. Weights & Biases.https://wandb.ai/ayush-thakur/ Christian, B. (2020j). The Alignment Problem: Machine Learning and Human
RLHF/repor ts/Understanding-Reinforcement-Learning-from-Human- Values. 1st ed. W.W.Norton. ISBN 9780393635829, p. 113
Feedback-RLHF-Part-1--VmlldzoyODk5MTIx. Accessed 3 Nov 2022 Christian, B. (2020k). The Alignment Problem: Machine Learning and Human
Balestriero, R., Ibrahim, M., Sobal, V., Morcos, A., Shekhar, S., Goldstein, T., Values. 1st ed. W.W.Norton. ISBN 9780393635829, pp. 17–80
Bordes, F., Bardes, A., Mialon, G., Tian, Y., Schwarzschild, A., Gordon A., Christian, B. (2020l). The Alignment Problem: Machine Learning and Human
Geiping, W. J., Garrido, Q., Fernandez, P., Bar, A., Pirsiavash, H., LeCun, Values. 1st ed. W.W.Norton. ISBN 9780393635829, pp. 52–92
H., Goldblum, M. (2023). A Cookbook of Self-Supervised Learning. Coop Himmelb(l)au, Meet DeepHimmelb(l)au. (2023). Our algorithm learns
arXiv. Cornell University. arXiv:2304.12210. https://doi.org/10.48550/ CHBL’s semantic characteristics to generate new interpretations and new
arXiv.2304.12210. Accessed 29 Apr 2023 worlds. Facebook Watch. https://www.facebook.com/watch/?v=14733
Barker, N. (2023). Architects may become a thing of the past, says ChatGPT. 00746151435 (accessed Mar. 3, 2023)
Dezeen. https://www.dezeen.com/2023/02/13/chat- gpt-ai-archi Cove.tool. (2022). Supercharge your building design capabilities. Cove.tool.
tecture/?utm_medium=email&utm_campaign=Daily%20Dezeen& https://www.cove.tools/ (accessed Oct. 1, 2022)
utmcontent=Daily%20Dezeen+CID_6238f123470e57eff1127de07 Creo: Design. (2022). The way it should be. ptc. https://www.ptc.com/en/produ
aa0e2b5&utm_source=Dezeen%20Mail&utm_term=Read%20more cts/creo (accessed Oct. 2, 2022)
(accessed Mar. 3, 2023) Crunchbase. (2022). XKool Technolgy. Crunchbase. https://www.crunchbase.
Barreto, S. (2022). Differences between backpropagation and feedforward com/organization/xkool-technology (accessed Dec. 4, 2022)
networks. Baeldung. https://www.baeldung.com/cs/neural-networks- Datamind. (2022). ARTIFICIAL INTELLIGENCE, MACHINE LEARNING, NEURAL NET-
backprop-vs-feedforward (accessed Nov. 5, 2022) WORKS. Datamind. https://www.datamind.cz/en/services/Artificial_intel
Bishop, C. M. (2020). Pattern recognition and machine learning. Springer. ISBN ligence_machine_learning_deep_neural_networks?msclkid=83db1
0-387-31073-8 4a2aa13173a1045708f78459aa3 (accessed Nov. 3, 2022)
Bolojan, D. (2022). Coop Himmelb(l)au,: Deep HIMMELB(L)AU. YouTube. del Campo, M., Manninger, S. 2019 2022 Imaginary Plans. Proceedings of the
https://www.youtube.com/watch?v=8G3jpGQMvWE (accessed Oct. 2019 ACADIA Conference – Ubiquity and Autonomy. https://www.
5, 2022) academia.edu/40735096/Imaginary_Plans (accessed Nov. 11, 2022)
Burke, J. (2022). Successful generative AI examples woth noting. Tech Accelerator. FinancesOnline. 2022 Cove.tool Reviev. FiancesOnline. https://reviews.finan
https://www.techtarget.com/searchenterpriseai/tip/Successful-gener cesonline.com/p/cove-tool/ (accessed Oct. 4, 2022)
ative-AI-examples-worth-noting (accessed Sept. 15, 2022) Firth-Butterfield, K. (2022). These were the biggest AI developments in 2022. Now
we must decide how to use them. World Economic Forum. https://www.
Sourek Architectural Intelligence (2024) 3:4 Page 20 of 21

weforum.org/agenda/2023/01/davos23-biggest-ai-developments- Leach, N. (2018a). Design in the age of artificial intelligence. Landscape archi-
how-to-use-them (accessed Feb. 19, 2023) tecture Frontiers. Higher Education Press, Peking University, 6(2), 115–118.
Foster+Partners. (2022). Applied R+D. Foster+Partners. https://www.fosterandp ISSN 2096-336X
artners.com/people/teams/applied-rplusd (accessed Nov. 11, 2022) Leach, N. (2018b). Design in the age of artificial intelligence. Landscape archi-
Fukushima, K. (1980). Neocognitron: A self-organizing neural network model tecture Frontiers. Higher Education Press, Peking University, 6(2), 115–118.
for a mechanism of pattern recognition unaffected by shift in position. ISSN 2096-336X. ibid. b, p. 118
Biological Cybernetics. Springer-Verlag., 36(9), 193–202. https://www.rctn. Leach, N. (2018c). Design in the age of artificial intelligence. Landscape archi-
org/bruno/public/papers/Fukushima1980.pdf (accessed Dec. 2, 2022) tecture Frontiers. Higher Education Press, Peking University, 6(2), 115–118.
Gershgorn, D. (2017). IT’S NOT ABOUT THE ALGORITHM THE data that trans- ISSN 2096-336X. ibid. c, pp. 113–114
formed AI research—And possibly the world. Quartz. July 26, 2017. https:// Leach, N. (2018d). Design in the age of artificial intelligence. Landscape archi-
qz.com/1034972/the-data-that-changed-the-direction-of-ai-research- tecture Frontiers. Higher Education Press, Peking University, 6(2), 115–118.
and-possibly-the-world (accessed Nov. 22, 2022) ISSN 2096-336X. ibid. d, p. 114
Giraffe. (2022). The platform for utban analytics. Giraffe. https://www.giraffe. Leach, N. (2018e). Design in the age of artificial intelligence. Landscape archi-
build/ (accessed Oct. 2, 2022) tecture Frontiers. Higher Education Press, Peking University, 6(2), 115–118.
Github: Let’s build from here. Github. https://github.com (accessed Mar. 24, ISSN 2096-336X. ibid. e, p. 117
2023) Leach, N. (2022a). Architecture in the age of artificial intelligence an introduc-
Gitlab: Software. Faster. GitLab. https://gitlab.com/ (accessed Mar. 24, 2023) tion to AI for architects (1st ed., p. 104). Bloomsbury Visual Arts. ISBN
Gonfalonieri, A.: Inverse Reinforcement Learning Introduction and Main Issues. 978-1-3501-6551-9
TDS. https://towardsdatascience.com/inverse-reinforcement-learning- Leach, N. (2022b). Architecture in the age of artificial intelligence an introduction
6453b7cdc90d (acces. Apr. 7, 2023) to AI for architects (1st ed., p. 104). Bloomsbury Visual Arts. ISBN 978-1-
Google Scholar https://scholar.google.com/citations?view_op=view_citat 3501-6551-9. ibid. b, pp. 26–64
ion&hl=en&user=xezhJcAAAAJ&citation_for_view=xegzhJcAAAAJ: Leach, N. (2022c). Architecture in the age of artificial intelligence an introduc-
u5HHmVD_uO8C (accessed Feb 28, 2023) tion to AI for architects (1st ed., p. 104). Bloomsbury Visual Arts. ISBN
Gowans, A., Ackerman, J. S., Scruton, R., Collins, P.: Architecture. Encyclopedia 978-1-3501-6551-9. ibid. c, pp. 62–74
Britannica https://www.britannica.com/topic/architecture (accessed Leach, N. (2022d). Architecture in the age of artificial intelligence an introduction
Jan. 15, 2023) to AI for architects (1st ed., p. 104). Bloomsbury Visual Arts. ISBN 978-1-
Gozallo-Birzuela, R., & Garrido-Merchan, E. C. (2023). GPT is not all you need. A 3501-6551-9. ibid. d, p. 109
state of the art review of large generative AI models. arXiv. Cornell Univer- Leach, N. (2022e). Architecture in the age of artificial intelligence an introduction
sity. https://arxiv.org/abs/2301.04655 (accessed Jan. 29, 2023) to AI for architects (1st ed., p. 104). Bloomsbury Visual Arts. ISBN 978-1-
Harari, Y. N., Musk, E., Wozniak, S., Bengio, Y., & Russel, S. (2023). Among other 3501-6551-9. ibid. e, p. 106
multiple authors: Pause Giant AI Experiments: An Open Letter. future of Liu, Z., Feng, Y., Black, M. J., Nowrouzezahrai, D., Paull, L., Liu, W.: MeshDiffusion:
life. https://futureoflife.org/open-letter/pause-giant-ai-experiments/ Score-based Generative 3D Mesh Modelling. arXiv. Cornell University.
(accessed Apr. 3, 2023) arXiv:2303.08133; arXiv:2303.08133v2; https://doi.org/10.48550/arXiv.
Harouk, Ch.: Spacemaker Proposes AI-Powered Generative Design to Create 2303.08133 (accessed Apr. 21, 2023)
More Sustainable Spaces and Cities. Archdaily. https://www.archdaily. Martin, B.: Investors didn’t think much of Baidu’s ChatGPT – competitor debut,
com/952850/spacemaker-proposes-ai-powered-generative-design-to- wiping $3 billion off the tech giant’s value – Fortune. Inferse. https://
create-more-sustainablespaces-and-cities (accessed Jan. 2, 2023) www.inferse.com/484983/investors-didnt-think-much-of-baidus-chatg
Heidegger, M. (2000). Voll Verdienst, doch dichterisch wohnet / Der Mensch auf pt-competitor-debut-wiping-3-billion-off-the-tech-giants-value-fortu
dieser Erde. Heidegger und Hölderlin, herausgegeben von Peter Trawny. ne/ (acces-sed Apr. 16, 2023)
Vittorio Klostermann. ISBN 978-3-465-03084-3 McCulloch, W., Pitts, W. (1943). A Logical Calculus of Ideas Immanent in Nerv-
Heidegger, M. (2006). Sein und Zeit (pp. 146–216). Max Niemeier Verlag, ous Activity. Bulletin of Mathematical Biophysics, 5(4), 115–133. https://
Tubingen. ISBN 3-484-70153-6. In English being and time. Translated doi.org/10.1007/BF02478259
by Macquarrie, J., Robinson, E. 1st ed., Blackwell publishers, 1962. ISBN Mittal, A. Neural Radiance Fields: Past, Present, and Future. arXiv. Cornell
0-631-19770-2 University. arXiv:2304.10050. https://doi.org/10.48550/arXiv.2304.10050.
IBM: What is supervised learning? IBM. https://www.ibm.com/topics/super Accessed 21 Apr 2023
vised-learning (accessed Oct. 30, 2022a) Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D.,
IBM: What is unsupervised learning? IBM. https://www.ibm.com/topics/unsup Riedmiller, M.: Plaing Atari with Deep Reinforcement Learning. Papers
ervised-learning (accessed Oct. 30, 2022b) with Code. paperswithcode.com/paper/ playing-atari-with-deep-
IMAGENET. https://www.image-net.org (accessed Nov. 5, 2022) reinforcement (accessed Nov. 3, 2022)
Isola, P: Phillip Isola. MIT. http://web.mit.edu/phillipi/ (accessed Feb. 23, 2023) MVRDV: NEW EXPERIMENTAL TECHNOLOGIES. MVRDV. https://www.mvrdv.
Jagtap, R., 2022, Understanding the Makarov Decision Process (MDP). built com/themes/15/next. (accessed Feb. 2, 2023)
in. https://builtin.com/machine-learning/markov-decision-process Nguyen, A., Karampatziakis, N., Chen, W.: Meet in the Middle: A New Pre-
(accessed Nov. 3, 2022) training Paradigm. arXiv. Cornell University. https://arxiv.org/abs/2303.
Kerr, J., Min Kim, Ch., Goldberg, K., Kanazawa, A., Tancik, M. LERF: Language 07295; arXiv:2303.07295v1; https://doi.org/10.48550/arXiv.2303.07295
Embedded Radi-ance Fields. arXiv. Cornell University. arXiv:2303.09553; (accessed Mar. 29, 2023)
arXiv:2303.09553v1. https://doi.org/10.48550/arXiv.2303.09553. Pandey, P.: TCAV: Interpretability Beyon Feature Attribution. Towards Data
Accessed 29 Mar 2023 Science. https://towards datascience.com/tcav-interpretability-beyond-
Kilian, K.: Čína má odpověď na ChatGPT. A ne jednu, další chatbot ukazuje feature-attribution-79b4d3610b4d (accessed Nov. 30, 2022)
Alibaba. Živě. https://www.zive.cz/clanky/cina-ma-odpoved-na-chatgp- Papers with Code: Imitation Learning. Papers with Code. https://paperswith
a-ne-jednu-dalsi-ai-chatbot-ukazujealibaba/sc-3-a-221685/default.aspx code.com/task/imitation-learning (accessed Apr. 12, 2023)
(accessed Apr. 16, 2023) Patternforge: What is PatternForge? PatternForge. https://patternforge.org
Krizhnevsky, A., Sutskever, I., & Hinton, G. (2012). ImageNet classification with (accessed Mar. 24, 2023)
deep convolutional neural networks. Computer Science, Communica- Pichai, S., 2023 An important next step on your AI journey. Google The Key-
tions of the ACM, 3. https://doi.org/10.1145/3065386 https://www. word. https://blog-google/technology/ai/bard-google-ai-search-updat
semanticscholar.org/paper/ImageNet-classification-with-deep-convo es/ (accessed Apr. 2, 2023)
lutional-Krizhevsky-Sutskever/abd1c342495432171beb7ca8fd9551 Pix2Pix. (2023). Machine Learning 4 Artists Github. https://ml4a.github.io/
ef13cbd0ff (accessed Feb 23, 2023) guides/Pix2Pix/ (accessed Jan. 6, 2023)
Kyle, R.: Spacemaker: Merging AI technology with Urban Planning and Design. PTC. (2022). What’s in Creo9. ptc. https://www.ptc.com/en/blogs/cad/whats-in-
HBS Digital Initiative. https://d3.harvard.edu/platform-digit/submission/ creo9 (accessed Oct. 4, 2022)
spacemaker-merging-ai-technology-with-urban-planning-and-design/ Qi, Ch., Cun, X., Zhang, Y., Lei, Ch., Wang, X., Shan, Y., Chen, Q.: FateZero:
(accessed Nov. 11, 2022) Fusing Attentions for Zero-shot Text-based Video Editing. arXiv.
Sourek Architectural Intelligence (2024) 3:4 Page 21 of 21

Cornell University. arXiv:2303.09535; arXiv:2303.09535v2; https://doi. Tian, R.: Suggestive Site Planning with Conditional GAN and Urban Data. Yuan,
org/10.48550/arXiv.2303.09535 (accessed Apr. 18, 2023) P.F., Yao, J., Yan, C., Wang, X., Leach, N. (eds), Proceedings of the 2020
Quintana, M., Schiavon, S., Wai Tham, K., & Miller, C. (2020). Balancing thermal DigitalFUTURES. CDRF 2020. Springer, Singapore. doi: 10.1007/978-981-
comfort datasets: We GAN, but should we? Buildsys ´20: Proceedings of the 33-4400-610. https://link.springer.com/chapter/10.1007/978-981-33-
7th ACM International Conference on Systems for Energy-Eficient Buildings, 4400-6_10 (accessed Nov. 11, 2022)
Cities, and Transportation (pp. 120–129). https://doi.org/10.1145/34083 Tribby3D. (2023). Structural Engineering Software to Learn in 2023: Top 10 Best.
08.3427612 Tribby3d. https://tribby3d.com/blog/structural-engineering-software/
Rijmenam Van, M. (2023). What is generative AI, and how will it disrupt society? (accessed Jan. 30, 2023)
The Digital Speaker. https://www.thedigitalspeaker.com/what-is-gener Turing, A. M. (1937). On Computable Numbers, with an Application to the
ative-ai-how-disrupt-society/ (accessed Feb. 24, 2023) Entscheidungsproblem. Proceedings of the London Mathematical Society,
Ross, S., Gordon, G., Bagnell, D. (2011). A Reduction of Imitation Learning and 42(1), 230–265. https://doi.org/10.1112/plms/s2-42.1.230
Structured Prediction to No-Regret Online Learning. Proceedings of the Urban, P. (2023). Amazon má továrnu na AI a chatboty. Na umělé inteligenci u něj
Fourteenth International Conference on Artificial Intelligence and Sta- pracuje víc lidí než u Microsoftu nebo Googlu. Connect! https://connect.
tistics. Proceedings of Machine Learning Research. vol. 15. pp. 627-635. zive.cz/clanky/amazon-bedrock-titan-generativni-ai/sc-320-a-221729/
https://proceedings.mlr.press/v15/ross11a.html. Accessed 22 Jan 2023 default.aspx (accessed Apr. 16, 2023)
Salehi, H., & Burgueňo, R. (2023). Emerging artificial intelligence methods in struc- van den Burgh, S. (2022). MVRDV next. AECtech. https://www.youtube.com/
tural engineering. ScienceDirect. https://www.sciencedirect.com/scien watch?v=ZIza0cM5wBM (accessed Feb. 3, 2023)
ce/article/abs/pii/S0141029617335526 (accessed Feb. 7, 2023) Vavra, D. (2023). Jak chytra je umela inteligence? PROTI VETRU s Danem Vavrou.
Schmidt, M. (2022). 3 AI predictions for 2023 and beyond, according to an AI https://youtube/vYQ33kP7BdYhttps://youtu.be/vYQ33kP7BdY
expert. World Economic Forum. https://www.weforum.org/agenda/ (accessed Feb. 27, 2023)
2023/01/key-ai-predictions-for-2023-and-beyond/?utm_source=sfmc& Vesely, D. (2004). Architecture in the age of divided representation: The ques-
utm_medium=email&utm_campaign=2794672_WeeklyAgenda3Fe tion of creativity in the shadow of production (1st ed.). MIT Press. ISBN
bruary2023GeorgSigning&utm_term=&emailType=Agenda%20Weekly 0–262–22067-9
(accessed Sept. 19, 2022) Wearrecho. (2023). Embrace new technology and stay ahead of the game: Wear-
SCIA ENGINEER. (2023). Scia. https://www.scia.net/en/scia-engineer (accessed recho will take you to the top. https://wearrecho.space (accessed Apr. 3,
Feb. 7, 2023) 2023)
Shaban, P. (2023). MS architekti. https://www.wemakespaces.archi/cs/lide Weng, Z., Yang, X., Li, A., Wu, Z., & Jiang, Y. (2023). Transforming CLIP to an
(accessed Jan. 23, 2023) Open-vocabulary Video Model via Interpolated Weight Optimization. arXiv.
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., arXiv:2302.00624; https://doi.org/10.48550/arXiv.2302.00624 (accessed
Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Fan, H., Sifre, L., Apr. 18, 2023)
Driessche, G. V. D., Graepel, T., & Hassabis, D. (2017). Mastering the game Wikipedia, a. (2022). Reinforcement learning. Wikipedia. https://en.wikipedia.
of Go without human knowledge (PDF). Nature, 550(7676), 354–359. org/wiki/Reinforcement_learning (accessed Oct. 30, 2022)
https://doi.org/10.1038/nature24270. Bibcode:2017Natur.550.354S. Wikipedia, b. (2022). Convolutional neural network. Wikipedia. https://en.wikip
ISSN 0028-0836. PMID 29052630. S2CID 205261034. closed access edia.org/wiki/Convolutional_neural_network (accessed Nov. 5, 2022)
Sourek, M. (2014). From functional areas towards metropolitans structure: Wikipedia, c. (2022). Recursive neural network. Wikipedia. https://en.wikipedia.
Public space in sustainable development context. Advanced Engineering org/wiki/Recursive_neural_network (accessed Nov. 7, 2022)
Forum. Trans Tech Publications, 12, 176–180. ISBN 978-3-03835-323-2 Wikipedia, d. (2022). Recurrent neural network. Wikipedia. https://en.wikipedia.
Sourek, M. (2022). Virtual twins of architecture: The singularity of the profession org/wiki/Recurrent_neural_network (accessed Nov. 7, 2022)
and the field. International Journal of Architecture and Planning, 2(2), Wikipedia, e. (2022). Variational autoencoder. Wikipedia. https://en.wikipedia.
1–26. https://doi.org/10.51483/IJARP.2.2.2022.1-26 org/wiki/Variational_autoencoder (accessed Sept. 15, 2022)
Spacemaker. (2022). Early-stage planning, reimagined. Autodesk. https://www. XKool. https://www.xkool.ai (accessed Dec. 4, 2022)
autodesk.com/products/spacemaker/overview (accessed Oct. 1, 2022) Zaha Hadid Studio. (2023). Antic Style Architectur. Bing.com Videos. https://
StackExchange. (2022). What are the fundamental differences between VAE and www.bing.com/videos/search?q=zaha+ai&docid=608008305523768
GAN for image generation? StackExchange. https://ai.stackexchange. 389&mid=399E3EA3BEEECAC53B89399E3EA3BEEECAC53B89&view=
com/questions/25601/what-are-the-fundamental-differences-betwe detail&FORM=VIRE (accessed Feb. 13, 2023)
en-vae-and-gan-for-image-generation (accessed Sept. 15, 2022) Zheng, Y., Jiang, Ch., Mao, J., Han, J., Ye, Ch., Huang, Q., Yeung, D., Yang, Z.,
Steinfeld, K. (2023a). Berkeley College of Environmental Design. https://ced.berke Liang, X., Xu, H.. 2023: CLIP2: Cantrastive Language-Image-Point
ley.edu/wpcontent/uploads/2022/02/pix2pixgrid01.jpg. Accessed 6 Prtraining from Real-World Point Cloud Data. arXiv. Cornell Univer-
Jan 2023 sity. arXiv:2303.12417; arXiv:2303.12417v2; https://doi.org/10.48550/
Steinfeld, K. (2023b). Sketch2Pix. CDRF. http://blah.ksteinfe.com/200625/cdrf_ arXiv.2303.12417 (accessed Apr. 18, 2023)
2020.html. Accessed 6 Jan 2023 Zorloni, L. (2023). Il Garante della privaci blocca ChatGPT in Italia. Wired. https://
Steinfeld, K. (2023c). Steinfeld, K. Berkeley College of Environmental Design. www.wired.it/article/chatgpt-blocco-italia-garante-privacy/ (accessed
https://ced.berkeley.edu/people/kyle-steinfeld. Accessed 6 Jan 2023 Apr. 2, 2023)
Stewart, M. (2022). GANs vs, autoencoders: Comparison of deep generative
models. Towards Data Science. https://towardsdatascience.com/gans-
vs-autoencoders-comparison-of-deep-generative-models-985cf15936 Publisher’s Note
ea (accessed Nov. 5, 2022) Springer Nature remains neutral with regard to jurisdictional claims in pub-
Storm, J. (2023). This week alone, more than 200 new AI tools were released. lished maps and institutional affiliations.
LinkedIn. https://www.linkedin.com/posts/joergstorm_drstorm-techn
ology-automotive-activity-7045322482-660036609-CBeT?utm_source=
share&utm_medium=member_android (accessed Apr. 2, 2023)
Stylemania.it. Před Chat GPT: zde jsou průkopníci, kteří připravili půdu pro umělou
inteligenci. Style Mania https://www.msn.com/cs-cz/zpravy/other/před-
chat-gpt-zde-jsou-průkop-níci-kteří-připravili-půdu-pro-umělou-intel
igenci/ss-AA18T6Hs?ocid=msedgdhp&pc=U531&cvid=b90b042fad
6c4cada69739b51d885451&ei=59#image=11 (accessed Mar. 26, 2023)
TechCrunch. (2022). Spacemaker: AI software for urban development is acquired
by Autodesk for 240M. TechCrunch. https://techcrunch.com/2020/11/17/
spacemaker-ai-software-for-urban-development-is-acquired-by-autod
esk-for-240mEC (accessed Oct. 2, 2022)