Professional Documents
Culture Documents
Machine Learning Architecture in The Age of Artificial Intelligence
Machine Learning Architecture in The Age of Artificial Intelligence
of
Age
the
Iin
DOI: 10.4324/9781003297192
ACKNOWLEDGEMENTS
What I did not realise as I started the project was that it would require me
Bob Frew, indulged my interest in using early computers (which were hardly
up to any useful graphic task) for project management. Late in my final
graduate year while on vacation in California, my uncle Edward Bernstein,
always one to have the latest gadget, gave me a week with his brand-new
Apple II Plus, on which was a new piece of software called VisiCalc, the first
spreadsheet. Suddenly, numeric modelling no longer required hard coding
and I surprised Bob with a finished project right after the holiday. Two years
later in San Francisco, the architect Plerb McLaughlin sent me several times to
Palo Alto, to research a new technology called 'expert systems', a provocative
but otherwise completely unrealisable technology. Around the same time, I
bought his firm's first personal computer not for generating drawings, but
-
images, I came to realise, is irresistible, but digital tools have just as much, if
not more, agency in architectural process outside of design itself.
After dozen years in mainstream practice with César Pelli and his managing
a
source of much and insight from the company. Sam Omans, now an
help
industry manager with Autodesk's Architecture/Engineering/Construction
business (and likely one of the few folks working in tech with a PhD in
architectural theory) has helped clarify ideas, locate information with the
Autodesk labyrinth and chase down critical images that illuminate the text.
Grace Liu, from the Autodesk Intellectual
Property team, was invaluable in
completing all the necessary image permissions there.
intelligent machines.
Here at Yale today I came to rely on our able architectural librarian in the Hass
Arts Library, Colwell, without whom I would have been unable to navigate
Tess
the university's vast, but often opaque, resources. Dean Deborah Berke has
been unfailingly supportive as I laboured to complete the manuscript while
we steered the School of Architecture through the global pandemic. And my
editors at RIBA Publishing -
and patience, particularly with my peripatetic schedule. Clare was not fazed a
bit when I asked for my long-time friend and collaborator, the editor Andreas
Muller, to pitch in as an additional set of eyes on the project. Andreas, who
edited my last manuscript, offered regular, clear and very useful advice to
improve both the flow and logic of the argument.
places, replicate the architect's unique blend of formalism and creativity, and be
responsible for the safety and fitness of a building? To address this, Phil's analysis
goes beyond the usual reductionist critique and considers how modern Al could
not only make the overall value chain of architecture faster or more efficient,
but also result in a stronger architectural profession overall. I find this book
tremendously thought-provoking, and I hope you do as well.
INTRODUCTION
Deciding that slog was not for me, I stumbled upon an exotic class in the
same department called 'Natural Language Processing', where, apparently,
we were going to teach computers to understand English. Our avuncular - if
prickly - professor, Roger Schank, explained that he had uncovered one of
the fundamental aspects of human existence by discovering the structure of
language understanding encoded within the mind. Our job was to translate
that theory into computer code. Of course, we were doing so on a 16-bit
predecessor of the IBM PC called a PDP-11 /45, with a whopping 256 kilobytes
of main memory. One afternoon, in the computer lab, with ten or more of us
working on the system, it burst into flames. 1
0.1:
A CATHODE RAV
OSCILLOSCOPE,
G. 1996, NOT
MUCH CHANGED
FROM ITS 1975
PREDECESSOR
(COURTESY
OF MAKEHAVEN
INC.)
Introduction
0.2:
AIM EARLIER
VERSION OF
THE PDP-11
COMPUTER,
NOT ON FIRE 2
My project for the semester was to write a program that would accept input
from what was then called a 'newswire' –
profession believes itself once again under threat. While the University of
Oxford's Richard and Daniel Susskind, exploring the implications of artificial
intelligence on the work of professionals, have suggested that 'we will neither
need nor want professionals to work in the way that they did in the twentieth
4
century and before', there is very little computational intelligence impinging
on the practice of architecture today. Yet we worry. Are we headed to a world
Here is one example. When computers design buildings, the public will lose
out. Why? Because when an architect imagines a building, it is from its base
upwards. What you see is an imagined image of what is being created from bottom
to top. A real building shows us a holistic structure that starts from the very
foundations and works all the way to the top, from foundation to roof, and so on.
If we do not know the foundations of the building, we do not know what we are
being sold, and we will not understand what we are looking at. The bottom line is
that there is little or no aesthetic value in a building that has no soul, and so the
negative view of digital architecture is not justified.
Why? Because when architect imagines a building, it is from its base upwards.
What you imagined image of what is being created from bottom to top.
see is an
A real building shows us a holistic structure that starts from the very foundations
and works all the way to the top, from roof to foundation and so on. If we don't
know the foundations of the building, we don't know what we're being sold, and
we won't understand what we're looking at. The bottom line is that there is little or
no aesthetic value in a building that has no soul and so the negative view of digital
I see two reasons why digital architecture is lacking aesthetics. First, it is solely
a technical process. If you try to make the process of creating a building into
something beautiful, a lot of work will go into that, and you won't save much in
time. And second, digital architecture uses software as the sole element in the
building process, which we know is unreliable.
Software is so far from being easy
to understand, we wouldn't recognise
even software. We can't see a word of
it as
the message it's trying to send. So, it's not easy to tell from looking at a building
whether it was designed by a computer or not. The work of design is not something
we can simply throw into the computer and have it turn out like a BMW.
Digital architects try to make the process of creating buildings into something
beautiful, but ultimately they don't deliver. There are other issues with digital
architecture. To make the building more "real", you need software to manipulate
the material the building will be made from. In the past, materials were never
digitally designed, so when a digital architect took the material out of a computer,
the software had no idea how to work with it, and the result was a "coffin" with
"free" space at the base. Digital architecture only takes you to the model first, and
once it's built, it will only keep it in the model, and create something completely
different with it. Digital architecture can give you better clarity, but at the cost of
aesthetic. It has never produced anything that looks like a real building.
Here's dear reader, that you're still with me after that brief italicised
hoping,
diversion, which was generated entirely (and unedited by me) with a
technology called GPT-3, the third generation of what is called a Generative
Pre-Trained Transformer, an artificial
intelligence (Al) system that conjures
original text in response to a prompt, which in this case was '(w)hen
computers design buildings, the public will lose out'. 5 GPT-3 is my oil tanker
program run in reverse –
The sample output above has an uncanny similarity to coherent English, the
broad framework of an 'original' argument and even cites concrete examples.
It has the maddening quality of bare intelligibility, as the more you think you
understand it, the more obscure it becomes. GPT-3 technology is the current
generation of machine intelligence that'learns' language by ingesting huge
amounts of text from the internet and 'teaches' itself underlying semantic
structures. This is the same strategy that the mostly reliable Google Translate
uses to translate a web page from English to Spanish, but greatly accelerated
by rapidly advancing machine learning (ML), Yet both my early effort and that
of GPT-3, above, lack real coherence, and the more time spent reading the
text above, the less sense it seems to make. GPT-3 is certainly a more efficient
approach than mine of 1976, but without the sweeping philosophical assertions.
Yet, the possibilities here are intriguing. Could a computer design an entire
building well? The computer scientist, Mark Greaves, who contributed the
Foreword to this book, describes the advances in natural language generation
with tools like GPT-3 as having 'fluency and expressivity':
>>
Using modern techniques, machines are starting to successfully
ML
always the first place that architectural theory visits when struggling with a
8
big problem. Unexamined, however, are the implications for the practice,
rather than the result, of architecture. If designers solve, as described by Peter
Rowe (quoting Horst Rittel) 'wicked problems', 9 with open-ended beginnings
and fixed conclusion, competent practice requires heuristics across a
no
PROCESS
ARCHITECTURE
IN THE AGE
OF MACHINE RELATIONSHIPS
INTELLIGENCE
FOREWORD V
INTRODUCTION VI
BIBLIOGRAPHY 173
REFERENCES 176
INDEX /
CREDITS 184
PROCESS
01
DOI: 10.4324/9781003297192-1
TECHNOLOGIES
AND
TOOLS
1
1.
DOI: 10.4324/9781003297192-2
TAXONOMY OF TECHNOLOGIES
the ability to perform complex cognitive tasks in ways that produce results akin to
the human mind is an enabler of the various tools described in this taxonomy.
-
of design. For at least the next 42 centuries architects abstracted their ideas
a
the act of drawing two parallel lines on a piece of paper the architect could
-
1.1.1: GUDEA,
WITH A PLAN
DRAWING IN
HIS LAP, C.
2150 BC 4
Machine Learning
COMPUTER-AIDED DESIGN
object of which was still the production of drawings but more precise. The
-
plan, rather than representing the diagram of a design that otherwise lived 1.1.2 AN
EARLY CAD
in his head, would now be an extraction from full-scale digital replica living
DRAWING
in computer memory just another view of the relevant data. Every member
-
BY PELLI
of the design-to-build team could, in theory, add information to that model CLARKE PELLI
process and outcome, the adversarial nature of building and the centuries-old
allure of drawings have made BIM a tool used largely for production of even
better working drawings than CAD. Its epistemological value as an organising
principle for the informatics of building is almost completely ignored.
The last 10 years have seen explosive digitisation of many aspects of modern
life, and design and construction have been no exception. Powered by the
ubiquitous availability of the massive storage and CPU power of the cloud,
and the ability to deliver those capabilities virtually anywhere through the
internet, the architecture/engineering/construction/operation (AECO) industry
is adopting a variety of computational tools, if peripatetically.
Building owners:
» demand digital documentation of completed projects in lieu of rolls of
post-construction drawings
» assemble data from sensor networks in their building control systems
to optimise building performance.
we are likely to see more years of excitement and confusion as the building
industry wills itself into its digital future at the same time that it gives up
8
the simple 'interoperability' afforded by paper-based, analogue processes.
Desires for some sort of broad theory of global data exchange will remain
The building industry is typically not well-enough organised, nor can it compile
enough market clout, to adopt fresh technologies or innovations soon after
their introduction. It often has to wait until hardware, software or business
models are sufficiently mature for architects, engineers and builders to adopt,
adapt and improve such systems for their use. Such was the case with CAD
platforms, which were originated by the aerospace industry and had to be
ported down to personal computers sufficiently inexpensive to be in reach
of AECO customers. Similarly, modelling platforms such as BIM or high-
resolution rendering eventually appropriated the tools of manufacturing and
movie-making once those technologies were within economic reach.
9
1.1.3:
A LIDAR SCAN
OF CONSTRUCTION
IN PROGRESS,
MAPPED AGAINST
A BIM DATA SET.
THE CONDUITS
(IN GREEN)
AND DUCTING
(IN BLUE)
ARE VIRTUAL
ADDITIONS TO
THE DIGITAL
SCAN FROM THE
MODEL.
Machine learning and artificial intelligence, the most recent tools on our
timeline, likely to follow the typical path in order to reach architects.
are
While major corporations are already absorbing AI/ML capabilities into their
core operating strategies, most artificial intelligence available to architects
11
request a ride to the office. A few promising start-ups and other experiments
are testing the technology on various tasks on the construction site (e.g.
1.1.4:
AUTODESK
SPACEMAKER
AI'S
DESIGN AND
EVALUATION
INTERFACE
_.._____
,,..=--
..,.,.,....,._,.,,,,,..,,..,_.
-
..,.
However, there are indications that this wait may not last much longer.
As I write this chapter in November 2020, Autodesk, my former employer,
announced the acquisition of Spacemaker, an Al-driven design generation tool
that evaluates site and building constraints and generates preliminary design
solutions (see Figure 1.1.4). That tool comprises a combination of design
representation, evaluative analysis and an Al infrastructure that learns best
results by interacting with its human decision-maker/operator.
A TAXONOMY OF USE
In earlier work I have proposed a taxonomy by which the vast array of
digital tools emerging might be categorised, irrespective of their underlying
technologies.13 In that analysis I suggest that the tasks of the building
enterprise, as supported by computation, fall broadly into four categories:
Representation
Drawings, text, images and physical models were the representational tools
of thepre-digital age, followed by CAD and, eventually, BIM and parametric
design. In a world of increasingly digitised data that
might be consumed by
smart machines, mathematical models and other data sets depicting a built
asset like sensor data coming from a building control system should also
- -
Computation is well-suited to
examining data that results from representation
in order to understand and evaluate it. Extended by the putative power of
artificial intelligence, this capability might more accurately be described, as
suggested by Agrawal, Gans and Goldfarb, as prediction: using (representational
and other) data to evaluate the implications of design decisions and predict
outcomes and implications of their underlying logic and decisions. 14
Realisation
of construction process that are mapped into BIM by intelligent systems are
one example is another aspect of digital project realisation.
-
Collaboration
Each of the technologies described above analogue drawing, CAD, BIM, and
-
ultimately AI/ML -
will change the way work is done and the tools deployed to
accomplish it. The intersection of technology types and the evolution of the
the tasks of the architect. Figure 1.1.6 describes some of these possibilities,
mapping the evolution of our four categories of technologies against both the
data richness of the design enterprise (the top curve) and the automation of
design process (the bottom curve).
1.1.6:
HUMAN VS
MACHINE
AUTOMATION
It is suggested here that as tools become enabled, there are parallel
more
virtual 3D models inparallel with parametric BIM families and data sets, and
generative design approaches through mature scripting that will memorialise
processes and procedures once dependent on human intervention. While
scripts within CAD simply manipulated geometry, tools like Dynamo allow a
designer to parametrically manipulate both the components of the design
(like the
size of windows) and their relationship to the overall building (like
the location of the windows within an exterior wall). In the interregnum, we
will see extensive generation of digital data and the episodic automation of
various processes that would be otherwise disconnected.
The transition to broad-scale AI/ML will greatly enhance both the amount and
the value of its precedent digital sources, serving as data lakes for intelligent
machines to learn from. At the same time, computers will start to train
themselves to perform machine-automated tasks. Those same machines will
teach themselves about the relationships of the heterogenous data sets that
comprise projects and create an 'interoperable' constellation of AECO data.
the
unprecedented power of digital simulations, one may surmise that
B ; Given
at some point virtual models may become perfect duplicates of and substitutes
for, the buildings they represent embodying and enacting all and every aspect of
-
them. Their designers could then make a digital model just as builders would once
have made an actual building, and the final translation from model to building
would entail no intellectual (or informational) added value whatsoever. « 15
»
THE IDEA THAT COMPUTERS MIGHT
DRAMATICALLY AUGMENT THE CAPABILITIES
OF HUMANS -
OR POSSIBLY SUPPLANT
US ALTOGETHER -
IS MANY YEARS OLD.
BEFORE SETTING OUT THE OPPORTUNITIES
AND THREATS OF AI FOR ARCHITECTURAL
DOI: 10.4324/9781003297192-3
Machine Learning
The computer scientist John McCarthy is generally credited with coining the
term 'artificial intelligence' in 1956, suggesting that computing machines
could somehow mimic the functions of the human brain. And long before
computing was within reach of mainstream practising architects, Nicholas
Negroponte and others were exploring the idea of digital design. As early
as 1964, Walter Gropius acknowledged that there might be a role for these
1.2.1:
NICHOLAS
NEGROPONTE’S
URBAN 5
SEEK, BY THE
ARCHITECTURE
MACHINE
GROUP, AS
EXHIBITED AT
THE JEWISH
MUSEUM, NEW
YORK, 197θ
1.2 What is Artificial Intel igence (AI)?
At the same time Nicholas Negroponte was exploring more practical questions
in his MIT
lab, 'The Architecture Machine'. An architect by training, Negroponte
experimented widely in the early uses of technology and design, anticipating
our use large screens, video and cameras, machine intelligence and
of tools like
immersive environments. His efforts anticipated early strategies for artificial
intelligence, positing the possibilities of'an intelligent environment that we
would all eventually inhabit and that would eventually surround all of us'. 2
BEYOND PERCEPTRONS
By the 1990s, computers were getting faster, cheaper and more available, and
a different strategy for AI emerged: neural networks and machine learning.
Benefiting from the vast computing power and equally gigantic storage
-
be
told about -
the world.9
» (l)t is a fallacy to suppose that what worked reasonably well for
Marcus argues that the work should consider a return to the original
motivations of the Al field simulating human cognition and combine the
- -
data collection and analysis capabilities of deep learning systems with new
models of perception and inference. While some progress has been made on
this front, such systems do not exist today and will depend on the digitisation
11
of new theories of understanding and, particularly, causality.
» her talents, skills and experience (as certified by, for example,
12
own
» machine learning systems (which might learn from data coming from
theirdesign projects, or even sensors within finished buildings and
provide insight), and
»
ultimately the speculative prospect of cognitive systems that can
reason within context (only seen in science fiction today).
The computers available to architects today are more adept at direct problem-
solving than what Stanford Anderson once called 'problem-worrying', resolving
14
the goals of the problem while simultaneously creating the design, evocative
of both Negroponte's and Peter Rowe's interest in heuristics as a strategy
for solving 'wicked problems'. Hard-coded software single-mindedly solves
15
specific problems; your cost-estimating system will tell you nothing about
the fire exiting required of your design, nor is it capable of learning how to
do so. Emerging AI/ML systems, now being applied to problems of the built
environment, may be able to evaluate or even predict issues in a specific
context, but certainly are nowhere near ready to design entire buildings,
heuristically or otherwise. And, as of this writing, devotees of cognitive
16
systems have spent decades building 'real world knowledge' as the basis of a
MACHINE CAPABILITIES
1.2.4:
MODIFIED
GREAVES’S
MODEL OF AI
CAPABILITIES 23
This is, however, only the current state of affairs and is a function of empiricist
Alsystems that can only 'deduce' based on massive correlations of data. It
is likely that, over time, empiricism will give way to emulation of cognition
as philosophers, neuroscientists, computer scientists and their commercial
counterparts build ever more capable machines that move toward general Al,
what Pedro Domingos has called the Master Algorithm, in the service of what
24
including BIM and emerging deep learning tools into coherent and valuable
practice in anticipation of the day when cognitive platforms are readily
available. By then, one hopes, the profession will have a firm grip on both the
technologies available and the means to direct them.
KNOWLEDGE
AND
INFORMATION
PROFESSIONAL
3
1.
DOI: 10.4324/9781003297192-4
Machine Learning
In The Future of the Professions, Richard and Daniel Susskind's treatise on the
potential demise of the
professions in a world of increasingly capable artificial
intelligence, you can find the following complete, if slightly tongue-in-cheek,
definition of why society has created and empowered a class of professionals,
architectural or otherwise (emphasis added):
There is a social bargain defined here: running our world would seem to
expertise who provide judgement and take personal responsibility for results
of that judgement. In the systems of building delivery, this means that, unlike
almost everyone else (builders, subcontractors, product suppliers, fabricators),
professional architects provide services, not things. And as a result, they take
personal, rather than corporate, responsibility for their actions.
1.3 Professional Information and Knowledge
The philosopher Donald Schön, who studied how professionals learn and
deploy expertise, has suggested that architects and other professionals work
in a way that is distinct from less institutionalised careers by virtue of what
he calls `reflective practice', or the ability to apply insight and make decisions
through the implicit understanding gained with `extraordinary knowledge in
matters of human importance'. What Schön called tacit `knowing in place'
3
However -
the formulation,
design, procurement, construction and
operation of a building is rife with
procedural and data-driven tasks, ranging from calculating quantities
to modulating temperature and humidity. As such, the systems within
which those processes operate are sure to be influenced, if not partially
transformed, by computing. The question for architects is where,
autonomous
how and what will tomorrow's Al-assisted architects really need to know?
job was to graduate students not in order to support the practice of today,
but rather to radically reform it in and for the future. Unwilling to miss an
opportunity to triangulate one of my favourite hobby horses, I asked the
practitioner how the profession simultaneously demand that graduates
can
firm, reaching licensure meant a bump in pay of $1,000, or slightly less than 2%
of what I expect he pays his least experienced staff. 5
ARB Part 3 examination in the UK and the comparable standard in the United
States: the Architectural Registration Exam (ARE) administered by the National
Council of Architectural Registrations Boards (NCARB). 6
and theoretical
platforms of the discipline of architecture.
»
Building Technology seeing and performing the technical
-
experience.
» Urban Design and Landscape -
»
Project Planning and Design -
what do you need to know to design
a complete building that meets all applicable regulations and
requirements?
»
Project Development and Documentation once the design is set, -
rough proxies for what our discipline believes an architect needs to know to
practise. And while the American standard seems to emphasise the synthetic
act of design itself in comparison to its British counterpart, an understanding
of performance, practice and technical issues, as indicated by the emphasis
on environmental systems, building technology and professional practice,
this door swings in the right direction' to as lofty as `try to stir the soul with
this spatial experience'. The knowledge necessary to accomplish the former
requires rudimentary understanding of how doors work, perhaps in the
a
context of a building regulation, while the latter likely demands the collective
insight from a lifetime of work. Either way, each goal shares the common
requirement that the architect reference proper, current and relevant
information and apply her judgement in its use.
aesthetic or technical -
is a task for
A brief demonstration tells this story well. Consider the image in Figure 1.3.2 .
descriptive phrase. The picture here is the result of the phrase `building in a
city', not a difficult or particularly complex challenge. You can judge the result
for yourself and try your hand at Al-created images on the Allen Institute
website. 14
1. Most relevant -
in the United States) Lexus/Nexus, which provides access to the entire SYSTEM IN
THE UNITED
history of American legal cases. Proper medical treatment would be almost STATES.
impossible without the former, nor could common law jurisprudence like in THERE IS NO
the US or UK progress without the latter. COMPARABLE
SYSTEM FOR
2. Perhaps, therefore, the first role of any Al system aimed at the building ARCHITECTURAL
industry could be getting our data, which are increasingly digitised and DATA
sorted.
At the point where the of architectural data, ranging from
numerous sources
In doing so, one of the greatest yet untapped resources available to today's
architects may become available to their successors, to wit, the digital project
data, terabytes of which reside on servers in today's offices across the worid,
that are the artefacts of project work. As a practitioner in the 1980s and 1990s,
despite all
our digital drawings and other data, we relied on memory to inform
1.3.4: us, with only our brains to connect our hard-won experience on previous
DIGITAL 17
INFORMATION
projects with decisions we needed to make on our current jobs. While
SOURCES AS today a human architect might have to scan a multitude of digital models to
SUGGESTED BY determine a best practice or trend illuminated by that data, an Al is well-suited
THE BUILDING
to gathering and evaluating such information from a firm's archives.
VENTURES
INNOVATION
NETWORK
That same principle may apply to another intractable issue of architectural
knowledge, the interoperability of information across digital systems. Figure
1.3.4 describes a potential array of today's digital information that could be
rationalised by Al into new outputs. Al processes are well-suited to indexing,
cross-referencing and correlating such data, and as such could become an
implicit interoperability tool for AEC information, and thereby begin to build a
more coherent informational platform for subsequent systems disciplinary -
epistemological coherence will combine for architects, and the days of disparate
standards, incompatible digital processes and inaccessible insight will end.
That work will likely grow from the priorities of practice, which, as argued
above, are largely concerned with the more practical, procedural and prosaic.
And as other parts of the building delivery process, examples of which include
feasibility studies, precision cost-modelling, construction automation and
autonomous digital building operation, evolve through increasing digitisation,
architects will need to understand how to manage and access information and
deploy it inthe service of the new responsibilities and professional obligations
that will result. Where today's architect relies on passing familiarity with an
DOI: 10.4324/9781003297192-5
My first real job in an architect's office was in the pre-digital era, long before
computers became ubiquitous in the profession. The managing partner of our
small practice in North Carolina returned to the studio one summer afternoon
after a demonstration of a new technology called 'computer-aided drafting'.
While he deeply sceptical
was of the entire idea, and especially the cost, he
noted that watching the plotter create a drawing was mesmerising, 'like a real
1
draftsman working on one part of the drawing and then another'. He also
made it clear that no machine was going to be replacing anyone there laying
down plastic lead on mylar sheets in our office in the foreseeable future.
1.4.1 AN
EARLY PEN
PLOTTER,
C. 198θ
Machine Learning
today are radically different from those of four decades ago, but the
ambivalence justifiably remains.
» Until recently, one assumed that automation would impact only
poorly qualified jobs. This might not be the case. Architecture will be
probably among the most severely hit disciplines. The reason for this
high degree of vulnerability is that architecture is among the most
formalized of all the arts. The mechanical part is stronger than in
other domains, and hence the traditional position of the discipline on
the threshold between art and technology.5 «
There can be a wide gulf between theoretical speculation and the realities
of daily practice, so now is an opportune time to bridge the two, lest the
dismantling begin in earnest. Beyond theorising about the possibilities of this
new technology ignoring in the hope it will pass us by or fighting the inevitable
-
design process and machine intelligence to determine how, if at all, they can at
worst co-exist and at best be mutually complementary.
1.4 Ai and Process Transformation in Design and Beyond
is digitised today. The ease with which the architecture profession shifted
entirely to remote work during the Covid pandemic demonstrates that design
work can be largely digitised and is likely to remain so, permanently.
1.4.4:
A LIDAR
SCAN
1.4.5:
DIGITAL
ANALYSIS AND
SIMULATION
1.4.6:
AUTODESK
B36θ
COST
MANAGEMENT
TRACKING
SYSTEM
1.4.7:
AUTODESK
TANDEM
DIGITAL TWIN
PROCESS TRANSFORMATION
At the centre of the transformation of data streams into new, digitally enabled
outcomes are both existing and new software, platforms and computerised
procedures that work in combination to move the work of architects to its
digital future, comprised of three elements.
The designers, builders and operators of a project then have two types of
tools to either create or support their respective roles in the creation of a
building: those that are 'automated' and those that are 'autonomous'. It is a
distinction that will be important to define, during the balance of this book,
how Al tools offer different opportunities and threats to the architect.
- -
Design task automation: Procedures and protocols that require the direct
intervention of the designer as likely to be autonomous in the future.
While most code checking is a manual process today, that procedure
can be supported by submitting a digital model to a code-checking tool
that uses AI to evaluate code compliance, combining a more traditional
‘architectural’ model with a technoscientific counterpart.
1.4.9:
MANUALLY
CONSTRUCTED
STUDY
MODELS, C.
1993
1.4.1¸:
SCRIPTING
TOOLS TO
GENERATE A
BUILDING
ENCLOSURE
1.4.11
PICTOBOT, AN
AUTONOMOUS
PAINTING
ROBOT, AS
PROPOSED BY
E. ASADI, B.
LI AND I.
CHEN 1 B8;
A VIRTUOUS LOOP
DOI: 10.4324/9781003297192-6
Machine Learning
When Richard and Daniel Susskind argued in their 2015 book, The Future of the
Professions, that artificial
intelligence would eventually replace society's need for
professionals, architects were understandably concerned. Ours is a precarious
profession, smaller, less politically powerful and certainly less remunerated
than our equally threatened brethren in law or medicine, who will certainly put
up a bigger fight before allowing themselves to be automated out of existence.
The Susskinds declared, right at the outset of their treatise, that
machine. Radiologists will give way to algorithms who can more patiently and
accurately read diagnostic images; attorneys will no longer be needed to search
documents for evidence, prepare routine legal arrangements or even represent
clients in disputes; and architects will not be needed to design, document or
help build projects.
2
As I write in 2021, we are six years past this declaration of extinction, with no
which will not be automated because it is impossible to capture such work with
4
rules expressed with logical expressions such as algorithms.
» The temptation is to say that because machines cannot reason like us, they
will never exercise judgement; because they cannot think like us, they will never
exercise creativity; because they cannot feel like us, they will never be empathetic.
And all that may be right. But it fails to recognize that machines still might be able
to carry out tasks that require empathy, judgement, or creativity when done by a
human being -
If today's most capable intelligent machines are based on various strategies for
deep learning, we can evaluate their capabilities by what they can be trained
to do. Mark Greaves's machine capability taxonomy, described in Chapter 1.2 ,
It can be argued that the most successful deep learning systems today -
the
ones that invent new game strategies and thereby annihilate their human
opponents, credibly translate from English to Japanese or even compose
music or paintings have somehow climbed Bloom's pyramid, having gone
-
far past remembering or even testing data to 'creating' new concepts. Within
certain extremely limited contexts, like the specific rules of the game of Go,
for example, or the learned' patterns of thousands of paintings, perhaps this
is true. A machine programmed with a rigorous set of rules, however, can be
1.5.1: ANDERSON/
LEARNING BLOOM KRATHWOHL GREAVES
TAXONOMIES
CD DESIGN
H CREATE FORMULATE
DEVELOP
X
DEFEND
LU EVALUATE JUDGE
Q
X SELECT
o
X
UJ
COMPARE h
ANALYSE DISCRIMINATE <
UJ
CC
TEST o
CHOOSE
H APPLY DEMONSTRATE
SKETCH/SOLVE H
OT
CLASSIFY
UJ UNDERSTAND LOCATE
a
cc TRANSLATE
o
cc
ID
DEFINE
S DUPLICATE
o REMEMBER
MEMORISE
said to have a ton of explicit knowledge, and as such is 'creating' in only a very
limited way, particularly since the measures of success winning the game or
-
So let us assume, at least for the next few pages of this argument, that
Susskind's thesis of task automation is the most likely implication of Al on
architectural practice in the foreseeable future, and that selected tasks within
the services that the profession provides may well be augmented, accelerated
or even replaced by an intelligent computer. Susskind asserts that those tasks
are easy to identify: they serve explicit goals that can be easily measured (to
determine success) and there needs to be a lot of data for the machine to learn
how to achieve the goal. 7 One might argue that a wide array of architectural
tasks might fit this bill, including questions such as: 'Does this project meet the
firesafety code?' or 'Does this ceiling plenum accommodate all the building
services?' With enough data and proper training, could a computer achieve
these goals? Or, even better, not just answer the question but generate the
required design solutions to meet those needs?
Why not? That stuff either fits in the ceiling plenum or not, and those doors
either have the right fire rating and swing in the correct direction, or they do not.
Assuming our deep-learning Al could study enough ceilings and exit corridors, it
should be able to learn right from wrong, and correct from negligent.
If
only it were that simple, we could start building architectural intelligence into
machines right away. However, there is another dimension to task automation,
what Stuart Russell calls hierarchical planning and management:
Intelligent behavior over long time scales requires the ability to plan and
»
from doing a PhD (one trillion actions) to a single motor control command
sent to one finger as a part of typing a single character in the application cover
letter. 8 «
ARCHITECTURAL SERVICES
In general, each scope prescribes a route through a standard set of tasks that can
be described in in categories like Project Definition, Design, Production and so forth.
Each phase of the work is comprised of a series of subtasks that differ by phase and
are modulated based on the expectations, deliverables and
professional standards
that govern the architect's services. These subtasks themselves can be categorised
into general buckets like Practice Management, Project Management, etc, and they
span across the phases of service. A rough mapping of a sample of such tasks,
aligned with service categories, can be found in Figure 1.5.3.
1.5.2: If we look at these task components through a lens that combines Susskind's
RIBA AMD
'clear goals and lots of data' criteria and Russell's task hierarchy, and includes
AIA SCOPE OF
SERVICES the ALM's tasks that require implicit knowledge, we can start sorting the service
work of architectsby likelihood of empiricist automation.
Let us call any task component that can be easily defined with a measurable
goal and executed through explicit logic as 'procedural', those that require an
intelligent integration of procedural tasks to reach a goal, even a measurable
one, as 'integrative' and those that are inherently creative, subjective and/
reliant on implicit knowledge as 'perceptive'. Figure 1.5.3 attempts to categorise
each component on this continuum, from procedural through integrative to
perceptive, depending on the work necessary to complete each task component.
As the coded bars suggest, there is very little that today'sarchitects do, even
at this relatively detailed level of examination, that can be characterised as
easily automatable. In fact, much of the technology of today is procedural
(including every piece of software we use), all of which is deployed in the
service of higher order tasks they accomplish. Eliminating humans from the
architectural equation is going to require an enormous jump in capability,
climbing the Bloom Taxonomy while combining those capabilities to accomplish
hierarchically complex objectives. This suggests that architects would better
spend time strategising which procedural aspects of practice might best benefit
from autonomous processes of Al, rather than worrying that our work will be
replaced wholesale by capable machines.
AI-SUPPORTING SERVICES
What this quick sketch problem suggests is that, at least in the near term, Al
1.5.5:
COMPANY GOAL LOGIC DATA
AI-BASED
START-UPS IN SPACEMAKER AI OPTIMISE MULTI- ADJUST BUILDING SITE PLANS AND
ARCHITECTURE, UNIT BUILDING DIMENSIONAL ANALYTICAL
2021
2θ21 CONFIGURATIONS PARAMETERS TO OUTPUTS
ON A SITE OPTIMISE USE AND OF SPACE USE
CONFIGURATION
As these systems become more capable, collecting data and building complex,
correlative data structures within their neural networks, it is likely that their
logics will expand to a wider range of targeted tasks across the architect's
responsibilities. Russell suggests that new ideas were often attributable to
'the three ineffable I's: intuition, insight, and inspiration'. 11 Procedural Al will
augment these critical (perceptual) components of professional judgement,
making the architect's services increasingly reliant upon, and validated by,
analysis and data. One can imagine a day where the architect, having fully
explored a range of options for the configuration of site including the resulting
-
performance data about rental area, storm water draining, zoning conformance
and even construction cost -
Beyond the near-term future, the architect's services will need to respond to an
Few of the required services that architects will need to provide to address
these needs can be found in the traditional methodologies of today, be it
through the RIBA Plan of Work or the AlA's definition of'Basic Services'. Deeper
analytical insight, deployment of broad data evaluation and coordination of the
data-driven tasks of a design team with varied (and ever-increasing) numbers of
consultants will require architects to integrate the Als that will support this work,
in the same way in which they manage their engineers today. The challenges of
design tomorrow will be best faced and conquered by people, masters of the
ineffable I's, whose ideas will drive the spaces, buildings and cities of tomorrow,
even if we reach the distant goal of Domingos' Master Algorithm.
1.5.6:
SPACEMAKER
AI, RECENTLY
ACQUIRED BY
AUTODESK FOR
$24θ MILLION
METHODS
AND
MEANS
DELIVERY,
6
1.
DOI: 10.4324/9781003297192-7
SYSTEMS OF DELIVERY
The term 'project delivery' comprises two central aspects of making buildings:
1. The constellation of clients (who create demand for buildings and then
operate and use them), architects (who design them) and constructors
(who procure, fabricate, assemble and build them).
2. How those players are arrayed in a set of professional, informational,
financial and legal relationships defined by their respective roles,
responsibilities and ability to manage risk.
1.6.1:
A SELECTION
OF TYPICAL
DELIVERY
MODELS SEEN
TODAY
Over the course of the 20th century, and particularly as architecture evolved
into a bona fide profession, the key players in these delivery models
developed prototypical roles. Clients look to convert capital into a physical
asset, but lack the technical capability to do so, so they hire architects to
define their needs and contractors to translate that definition into a building.
For a variety of reasons that includes the misalignment of interests, these
systems yield unsatisfactory results and as such there has been extensive
2
DESIGN INTENT
It was not always such. In the late 18th century in the UK, and as long as 100 years
later in the US, the architect was wholly responsible for all aspects of construction.
Higgin and Jessop described the project delivery model in Figure 1.6.2 .
1.6.2:
PROJECT
DELIVERY IN
THE UK, 18TH
CENTURY 6
Note the appearance of the quantity surveyor (to manage costs) and the
main contractor (to procure and coordinate the work), both roles that reflect
current practice. As these models evolved, the level of granularity and
resolution of information necessary to build ever increased, and it became
apparent that as the project's interface with the procurement process, the
contractor and her minions – including subcontractors who would fabricate
building systems – was best to 'finalise' the design itself.8
As the building industry has moved from drawings to BIM, the inherent
tensions of such a system were exacerbated rather than calmed
by the
availability of 3D information. Architects that
they had neither the
complained
expertise nor the fee to provide extreme construction detail in their design
intent BIM data, and builders declared that the resulting BIM deliverables
were unsuitable for building. So despite the insertion of a technology designed
Yet there are other forces at play that may, through technology, finally close
this ancient divide. In a market that is increasingly pressed towards more
1.6.3:
PROJECT
DELIVERY IN
THE UK, 19TH
AND 20TH
CENTURIES 9
efficiency and productivity, and dissatisfied with the results of traditional
delivery approaches, the move to digital fabrication and industrial processes
in construction is inevitable. This is particularly true as automation moves the
resulting demand for higher quality digital deliverables to smooth the path.
However, will architects and their consulting engineers deliver them? And can
Al play a part in closing the professional divide?
PROFESSIONAL RELATIONSHIPS
This question must be examined in the context of all the obligations of the
architect, informational and otherwise. To do so, we can turn the traditional
delivery model diagrams found in Figure 1.6.1 inside-out to look at the specific
connections of the architect in any such structure, depicted in Figure 1.6.2 .
With our hypothetical architect in the middle of her relationships, we can see
four distinct roles required (see Figure 1.6.4).
An agent of the owner, who acts as the client's intermediary in the process,
generating the design and stewarding it, armed with descriptions of her
'design intent', through construction and acting as the owner's proxy to assure
the building conforms to that intent.
A leader of the design team, who orchestrates and integrates the work of
various consultancies in the service of creating a coherent, coordinated and
accurate design which will be passed along to the builders.
A guide to the builder, to articulate the goals of the project and help the
construction team to interpret, clarify and ultimately review and approve
the design intent on behalf of the client as it is translated into more detailed
information to support construction. A subcontractor, responsible for a given
building system, will often create very detailed information in support of the
fabrication of that system (shop drawings) but the architect must review and
approve such proposals before the fabricator may begin.
A protector of the public, including both the specific users of the project as
well as those with whom the architect has nospecific contractual obligations
but nonetheless is responsible for the health, safety and welfare of those who
inhabit her design.
1.6.4:
ARCHITECT’S
RELATIONSHIPS
IN DELIVERY
MODELS*
*
For further detail on this diagram,
see Figure 2.2.1.
So perhaps Al will day rationalise the ebbs and flows of digital information
one
that are so good at finding patterns and connections between data points.
But how might Al, in the immediate future, either augment or eliminate the
jobs of architects as defined in these four roles? Some early speculation is
summarised in Table 1.6.5 .
1.6.5:
EXAMPLE
OF AI
IMPLICATIONS
FOR THE
DELIVERY
ROLE OF THE
ARCHITECT
So what is the most useful focus for Al in delivery from the architect's
perspective? Automation algorithms good at memorialising processes (like
are
DOI: 10.4324/9781003297192-8
VALUE
AND
COMPENSATION
ECONOMICS,
1
2.
DOI: 10.4324/9781003297192-9
Architects operate within the larger economic models of the building
industry, which has long struggled to create consistent value propositions
for its participants. Most buildings are produced under an economic dictum:
achieve the end product by optimising for the single variable of lowest
first cost. Clients often select architects by arbitraging fees for the lowest
price; architects then choose their consulting engineers in the same way.
Contractors are often chosen based upon lowest bid, passing that logic down
the entire supply chain to the far reaches of building product manufacturers,
fabricators and suppliers. The value of the resulting project to the participants
in its creation or, ultimately, the client, is not reflected in the economic deals
that actualise it. As artificial intelligence changes the capabilities, obligations
and outcomes of project architects, how might the economic propositions
evolve accordingly?
CANONICAL MODELS
however, that those jobs are more than replaced by new ones necessary to
support the creation and support of the new technologies. So you might have
lost your job working in an architect's office doing zoning and code analysis
and planning studies, but you (or someone else) will surely be hired by one
of the many Al-based companies creating software to do that particular
task. And, in theory at least, the quality of the resulting design work created
3
Machine Learning
jobs. He suggests that technology will continue to spread the gap between
skilled labour (not easily replaced by machines) and unskilled jobs (where
many tasks may be automated). Architects are generally considered skilled
workers, but there are broad swathes of our jobs subject to such automation,
as suggested in Figure 1.5.3 .
The challenge, as always, is converting that resulting potency into actual value
that is reflected in the economics of practice. Technology notwithstanding,
it has always been difficult for architects to both improve the quality of our
services and the amount of money we are paid to provide them.
clients an hourly fee for 'CAD services' that was treated as a project expense
not unlike travel or blueprinting. Over time, however, clients got wise to
this idea and refused to pay these charges, arguing that the benefits of the
computer's precision and efficiency accrued to the architect, not themselves.
Of course, this point of view did not reflect the greater accuracy or consistency
of CAD-generated deliverables, nor the increased complexity of design
solutions they were able to create, and architects (once again) failed to convert
the improvement in the quality of their services to an increase in their fees,
which continued to be pressured by lowest-first-cost competition.
2.1.1:
A HAPPY CAD
OPERATOR AT
HIS STATION
IN THE
OFFICES OF
CESAR PELLI
& ASSOCIATES
[NOW PELLI
CLARKE PELLI
ARCHITECTS],
C. 2000
recovery. If net revenue is arough proxy for work produced, then that same
amount of effort was produced by 16,000 fewer employees (about 11 %) at the
same time that BIM adoption rose almost four-fold:
2.1.2:
COMPARING
REVENUE, BIM
ADOPTION AND
EMPLOYMENT IN
TWO RECENT
YEARS 5
What the data seems to suggest is that these architects increased their
productive capacity, presumably by use of newer technology, by over 10% by
producing the same amount of work with far fewer people. Anecdotal data
from practitioners indicate that while work volume increased steadily after the
Crisis, fee multiples stayed depressed, suggesting that these numbers may
not completely reflect the productivity gains of BIM technology. A more careful
analysis mapping profits, BIM adoption and a proxy for productivity (net fee
revenue per employee) in Figure 2.1.3 indicates that productivity accelerates
PRODUCTIVITY REDEFINED
process, exactly how long does it take to have a good idea, and then produce
it? If design is a process of solving Horst Rittel's 'wicked problems', then the
'wicked' nature of the process itself makes it difficult to precisely answer
this question, making the resulting projections of time, effort and expense
similarly intractable.
Once this
system has learned to find all the pieces of the building's existing
components, it can move to 'Part 2' of this process applying evaluative
-
design. Do the doors swing in the right direction? Do they have the proper
fire ratings? Are the stairs properly configured, within protective enclosures,
large enough to handle the number of occupants? This is the intersection
of automation, analysis and evaluation in Greaves's terms, indicated in the
diagrams as 'Part 1' .
an
overlay of fire
spread, smoke distribution and occupancy behaviour it can -
» If it works well, it saves time, effort and brainpower for the architect,
who can then, in theory, either convert the resulting work cycles into
profit (by simply cashing them) or by applying them to improve the
'wicked' characteristics of the design itself. Either way, this architect
has converted the capabilities of her new AI-enabled design assistant
into value, economic or otherwise. 8
data, sometimes
is notorious for lack of data standards, repositories or means for shared
a
require enormous amounts of data, it is more likely that architectural data per
se would be collected among the profession itself. Beyond that, training and
reference data for machine learning platforms could be aggregated at project
level, say by building type,or even across the design-to-build marketplace.
Eventually, all such data could exist in the context of the overall building
industry, including design, construction, procurement and operational
information, ultimately referenceable across all the players implementing AI.
2.1.5:
DATA
REALMS
WITH
EXCHANGE
VALUE IN
AECO
There is a tremendous economic value in such a proposed data reserve,
both as a source of
insight and
employment, Training systems requires
as AI
coherent, consistent, 'clean' data and it has been argued that conforming data
requires as much, if not more, effort than building the AI software itself.
10
As the 'canonical model' suggests, curating and managing this data is a likely
source of employment for architects,
new roles emerging from the resulting
creative destruction.
Such data should not, however, be simply gifted to the industry, as it has
tremendous inherent value as well as potential risk of misuse. The creation
If Susskind is right, the next decade will see an array of architectural tasks
augmented, and in many supplanted,
cases by AI systems. Whether entire
jobs will be replaced as a result is the
subject of some debate. Mastering
a game like chess or Go with very specific rules, precedents and a highly
-
DOI: 10.4324/9781003297192-10
While we await a wave of
intelligent machines to change the architectural
profession, AI/ML technology is already in wide use to set insurance rates,
evaluate radiographic images, determine eligibility for loans and other
government benefits, even capture suspected criminals with facial recognition.
And given the inscrutable nature of these algorithms, which 'teach themselves'
to generate results from pools of data (like hundreds of thousands of
insurance claims, or even millions of portraits of potential criminals scraped
from the internet), it is impossible to unpack the underlying logic by which
they make decisions. Worse, that logic is barely comprehensible to the
professionals that deploy it.
» Not until they were standing in the courtroom in the middle of
a hearing did the witness representing the state reveal that the
government had just adopted a new algorithm. The witness, a nurse,
couldn't explain anything about it. 'Of course not they bought it off
-
the shelf,' Gilman says. 'At least she's a nurse, not a computer scientist.
She couldn't answer what factors go into it. How is it weighted? What
This vignette elegantly captures the range of challenges that the regulators,
the public, the courts, architects and their clients face as the work of
professionals is automated: who or what is responsible for the implications of
decisions made by machines, and can they be sufficiently understood to
ever
The United Kingdom and United States share a common law tradition
establishing the competency standard for architects (emphasis added):
licensed architect must design any building for human habitation larger than
approximately 275 sqm. 5
2.2 Laws, Policy and Risk
1. Business risk: the possibility that the obligations of service will require
more resources than available in the contract, particularly fees.
2. Professional liability: the possibility that an error in judgement will result
in an assertion of professional negligence in violation of the duty of care.
The business risks of machine intelligence in architecture are more existential,
and these are addressed in a later section in this
chapter. Of more direct,
practical consequence are questions of professional responsibility, duty of
care and the implications of machines making complex decisions either in
obscure -
it will be impossible to
determine why a given decision has been made.
In fact, the polar opposite is the case. Before you are allowed to access any
9
from the software you have licensed.
lot of architects, engineers and contractors. At the risk of either inspiring you
to turn the page in boredom or abandon the chapter entirely, I quote at length
below two relevant but nonetheless redacted passages from the standard
Autodesk EULA, with emphasis (bolding) added. While painful to read, they are
illuminating:
5.2 Disclaimer. EXCEPT FOR THE EXPRESS LIMITED WARRANTY PROVIDED
IN SECTION 5.1 (LIMITED WARRANTY), AND TO THE MAXIMUM EXTENT
PERMITTED BY APPLICABLE LAW, AUTODESK AND ITS SUPPLIERS MAKE,
AND LICENSEE RECEIVES, NO WARRANTIES, REPRESENTATIONS, OR
CONDITIONS OF ANY KIND, EXPRESS OR IMPLIED (INCLUDING, WITHOUT
and you are still on the hook. Please make sure tokeep your
subscription current.
FAILURES OF EXECUTION
Failures in the building industry are common, ranging from the more typical
broken schedules and blown budgets to infrequent but calamitous disasters
like the Grenfell Tower fire in west London. Somewhere in between these
extremes lies the responsibility of buildings to be technically, environmentally,
socially andcontextually appropriate. Surely technology can play an important
part in helping building professionals and especially principal designers
-
such as architects do a better job with such outcomes. For the purposes of
-
this particular exploration, the question remains about the extent to which
the responsibilities of managing the resulting risks of project execution are
increased or diminished by Al-driven tools.
Those risks are sketched in Figure 2.2.1 based , on Figure 1.6.5 , that examines
the fundamental risks of failure when Al is enlisted to assist the architect in
each of her four fundamental roles during project execution:
2.2.1:
ROLES AMD
PROFESSIONAL
RISK
Consider these risks in the context of the terrible fire at Grenfell Tower in
2017. In that disaster, a small
appliance fire on the fourth floor of a residential
high-rise spread uncontrolled through the building envelope, and 72 people
perished. Components of that envelope had been replaced during a 2015
refurbishment conducted by the building owners and managers and without
direct involvement of the original principal designers. In fact, the requirement
that every project even involve a principal designer was implemented in 2015,
too late to be relevant during the refurbishment project. The Grenfell disaster
was a result of a confluence of technical decisions and errors made by a
andparticularlythe architect -
POLICY-MAKING
a global scale) and the Grenfell disaster (at a project scale). Understanding,
Two policy initiatives are suggested by this logic. Given that each of the
responsibilities described in Figure 2.2.1 might be easily characterised as
a 'wicked problem', it makes little sense for the resulting obligations to be
delegated strictly to machines, which will play an important but not exclusive
-
14
-
In the short term, certain firms will likely establish a viable but short-lived
competitive advantage by early adoption of Al that will differentiate their
services by capabilities or efficiency. As more firms follow their lead, this
Over the long term, however, architects are likely to face the same questions
ofdisruption and replacement by cognitive automation as other knowledge
15
workers, although the argument above suggests that the timeline of our
destruction may be attenuated. Our demise could be largely eliminated,
however, by using the capabilities of Al technologies to increase the value of
our services and by implication, of the built environment itself and to make
- -
society more dependent on architects and the machines that assist them, in
that order.
PROFESSIONALS
FOR
DEMAND
THE
3
2.
DOI: 10.4324/9781003297192-11
When the UK Government's Health and Safety Executive (HSE) enacted
the requirement that building projects require a principal designer, it
acknowledged that built assets manifest in three distinct phases: planning,
execution and use (followed, presumably, by eventual demolition). Before
2015, when the HSE CDM (Construction, Design and Management) regulations
went into effect, designers were considered desirable, but not necessary,
HSE makes the case for the necessity for architects (as one option) crisply in
their regulation, indicating that principal designers must (emphasis added):
Missing, of course, from this otherwise nifty summary of the need for
designers is anything about the quality of the resulting artefact, including
its suitability for relationship to context, expressive nature, or even
use,
environmental social appropriateness. These are results that clients
or
who hire architects clearly desire, even though they could meet the CDM
requirements with any party willing to assume the role of principal designer.
Even so, consider whether an intelligent machine in the foreseeable future
Machine Learning
might 'plan, manage and coordinate health and safety', 'help and advise
the client', 'eliminate foreseeable risks' or 'ensure everyone communicates
and cooperates'. If these things were even remotely possible, I suspect
construction managers, who perform many of the same tasks during their
phase of the work, will join architects at the unemployment office. However, is
even considering such a future a good idea?
A comparison of the two structures, derived from Figure 1.5.2 can be seen ,
in Figure 2.3.1 in which I have also diagrammed the subtle but important
,
In contrast with the 'bottom up' analysis of services we looked at in Chapter 1.5 ,
let us consider the implications of machine intelligence on these super-stages
and how the work of the architect might be either augmented or replaced.
PROJECTIVE RESPONSIBILITIES
At the heart of the architect's value in creating the built environment is what
I will call, for purposes of this discussion, her 'projective responsibilities' to
generate and instantiate ideas about the future state of the building she is
designing. To do this job, her conceptual skills must range from broad-scale
predictions about the implications of her building in the city, to the minute
choices of finishes in the interior; this is a very broad remit, particularly when
each of these decisions should support an integrated vision of the project.
to follow, like the rules of chess or Go. Their ability to project the future
state of, say, a building design, is a function of past experience (as defined
by data generated by other projects) and whatever rule set they have been
programmed to follow. Completely missing in today's systems is the ability
to reason counter-factually or to understand causality (why something did
or did not happen) versus correlation (something might happen because,
Pearl is convinced that truly intelligent machines are not possible until they
can reason causally and climb what he calls The Ladder of Causation'
(see Figure 2.3.2 ), which has three rungs;
mentioned earlier. Schank believed that the true test of any intelligent
machine was the ability to reason inferentially -
need to explicitly code all the resulting cognitive logic. Schank's thesis faded
with the 'Al Winter' of the 1980s.
manipulate at will to
imagine hypothetical environments for planning
and learning (they have) the ability to create and store a mental
...
prosaic terms, 'Project Definition', is the central value of a good architect and
demands third-rung talents that are unlikely to be achieved by machines
anytime soon. In fact, should machines reach the second rung, we might
achieve tools that help architects speculate on 'what if, the more modern
versions of today's analysis software, that could be hugely helpful to the
human architects occupying the top of Pearl's ladder.
TECHNICAL COMPRESSION
11
own ...
morally responsible agent.' If there is
as a any lesson from Grenfell,
it is that the complexity of the building enterprise -
in all dimensions of
construction, supply chain management, certification and regulation,
design,
and human behaviour is beyond the ken of machines. As Dennett further
-
posits: 'We don't need conscious agents. There is a surfeit of natural conscious
agents, enough to handle whatever tasks should be reserved for such special
and privileged entities. We need intelligent tools.' 12
our building and give our students access to our facilities while working in
studio. Every classroom was evaluated and rearranged so those faculty who
could teach in person might do so. Air systems were evaluated and adjusted,
and a safe occupancy schedule established and enforced. The school year
ended without almost no positive cases among our students.
>>
TECHNOLOGIES, BY THEMSELVES, RARELY
BEND THE ARC OF ARCHITECTURAL EDUCATION
OR CERTIFICATION. NEW TOOLS CHANGE THE
NATURE OF THE ARCHITECTURAL PROCESS,
ALBEIT VERY SLOWLY, AND THE INTELLECTUAL
INFRASTRUCTURE OF THE PROFESSION -
DOI: 10.4324/9781003297192-12
I have argued to this point that while Al systems are likely to both augment
and impinge onthe work of architects, they are unlikely to replace us as
designers, a capability that will require the development of artificial general
intelligence (AGI). The data scientist Herbert Roitblat correlates AGI with just
such an ability to attack the 'wicked' problem of design, suggesting that 'To
have a truly general intelligence, computers will need the capability to define
1
and structure their own problems,' which is an excellent way to characterise
the value of a talented designer.
institutions -
LEARNING OBJECTIVES
We will touch first on the well-trod and contested ground between those who
prepare architects for practice in the academy and those who establish the
criteria for accredited curricula and, eventually, professional certification that
leads to licensure. The uneasy truce between providers of architectural talent,
the certifiers of competence and the consumers of that talent in daily practice
is underpinned by a basic tension: what does it mean to educate a competent
While the standards and structures differ slightly, these arrangements are
largely the same in the UK and US, and compiled in Figure 2.4.1 .
2.4.1:
INTELLECTUAL
DEMANDS
ON THE
ARCHITECT
2.4 Education, Certification and Training
TERMS OF TOOLS
technical degrees for those so inclined.4 Since BIM joined a crowded field
of'representational' instruments (used directly to depict a design), and that
software is deployed largely in the service of form- and image-making, this
conclusion is understandable. At my institution we are careful to say that we
teach principles and theory, not tools, and there are no parts of the curriculum
5
(save one) where learning tools can result in credit toward the degree.
This arm's-length relationship, however, will not serve either students or the
overall professional well in the long term, and the advent of machines that
can do knowledge work is best faced now by educators and other leaders
These two objectives are self-reinforcing; by laying out the proper terms of
intelligent digital competence in the enterprise of design, the academy can set
the direction for their use in the marketplace.
algorithms that design things themselves and those that provide a supporting
role. Subscribing to the earlier argument that we need 'tools, not colleagues', 6
some of the most interesting research today in the architecture/ML nexus looks
This work is important in that it may yield insights into building organisation,
or even optimisation. It allows designers to see problems in a different light,
but it does not solve those same problems. Stuart Russell suggests that'... Al
research has focused on systems that are better at making decisions, but that is
not the same as making better decisions'. 7 Thus, this research is not likely to be
of the most immediate use in a world when structured, scientific and technical
On leaving the academy, our future architect enters the domain of the
profession and its disparate masters: registration authorities and professional
associations and accreditors, both with roles in certifying competence, and
therefore in defining the knowledge and skills that a professional architect
must possess. However, where technology once relieved architects of the
whether in school or otherwise. Even in 2012, the year the last analysis was
completed (see Figure 2.4.3 ), there was strong agreement that technological
skill was necessary to be deemed competent, if only towards various
representational ends (like drawings).
2.4.3:
NCARB 2012
PRACTICE While we wait for the 2020 analysis, it is safe to presume that the 2032
ANALYSIS OF
analysis will include knowledge and skills of Al applications, and the overall
KNOWLEDGE/
SKILLS summary of competencies will reflect the idea that certain functions of today's
RELATED TO architects, particularly those related to building science, will be performed by
TECHNOLOGY 8 machines and managed and integrated into project process by architects.
The lessons of the UK National Level 2 BIM Standard, which stipulates both
the information outputs and performance levels of the design process when
powered by BIM, may be instructive. 9 It was created by an industry consortium
and eventually evolved from a UK-only template (PAS 1192) to an international
standard (ISO 19650). While some practitioners in the UK may avoid working in
BIM, any government-funded project requires it for the large number of industry
projects they fund, and it is a matter of time before BIM techniques and data
strategies will be instantiated into the standard of professional care, expected of
competent practitioners.
The Level 2 standard was based larger industry strategy established by the
on a
This suggests that the most important role of certifiers in establishing the
use of machine intelligence should not stem from determining or driving the
Professional associations, like RIBA or AIA, have more vague ideas about
certification for membership, as such considerations are primarily designed
to assure that members have credibility with the
marketplace. Continuing
professional development requirements ask only that member architects
be regularly exposed to a broad spectrum of technical and professional
concerns, stipulating general categories (such as 'Health, Safety and Welfare',
for example, in the US) and numbers of hours of attendance. Over the next
decade, as architectural clients rely on AI-based processes in their business
or government operations, it is possible that those clients will ask the same of
their architects, and professional education and CPD certification are sure to
follow the desires of the customer base.
The marketplace is currently pressing the building industry about carbon and
climatechange, and that challenge gives us a good opportunity to speculate
on how the academic and professional platforms of architecture might
In 2032 the Ministry ofJustice issues a request for proposals for a new
update of ISO 19650, the so called'Intelligent Level 3 BIM' mandate, and reflect
best practices in responsible environmental design, including certifications that the
project will be net zero, generate at least 500 construction jobs and be free of any
evidence of modern slavery in labour or material practices.
The SPLC program has been created in concert with a new definition of Principal
Designer established in 2029 by the Health and Safety Executive that includes
environmental and labour equity in the responsibilities for that designation. Many
firms have been experimenting with platforms, smartTALLY (see Figure 2.4.4)
two AI
and buiidFRDM (see Figure 2.4.5), that assess Level 3 BIM schemes for embodied
carbon and forced labour, and collect information about design decisions and
strategies that are contributed to the National Building Data Trust. The HSE has
further stipulated that projects for human habitation larger than 300 sqm must
have an assigned Principal Designer who is a licensed architect, causing some
consternation amongst the country's construction/design managers none of -
This admittedly rosy scenario presupposes that our profession organise itself
in ways as yet unseen to accomplish three ends:
DOI: 10.4324/9781003297192-13
OBJECTIVES
THE
1
3.
DESIGN
OF
DOI: 10.4324/9781003297192-14
When Alberti defined the architect's design of a building as 'conceived in the
mind, made up of lines and angles, and perfected in the learned mind and
imagination', he centred those responsibilities on the abstract projection of
1
»
Designers first need drawings and models to explore, nurture and
develop the idea of the building (those) models should also be used
...
Today's architects use much the same approach, bolstered by various digital
armaments, but what happens when those tools become agents of design?
Let us assume that an architect by, say, the year 2030 has a complement
of Al-enabled tools at her disposal, along with significant advancements
in the resolution, precision and flexibility of modelling platforms that one
hopes would be the logical successors of today's BIM. While there may be a
conceptual breakthrough, some time in the future, in what I have described
as 'cognitive' Al platforms that can reason inferentially about the complex
interactions that comprise a building, let us assume that by 2030 we only are
at the point of useful architectural versions of 'undeniably single-mindedly
successful' platforms like today's language and game-playing software, for
5
Machine Learning
6
example GPT-3 or AlphaGo. Such systems would be tightly tied to design
modelling/representational platforms and their data, and receive training
from other information sources like engineering systems, real-world data
collection about context from LIDAR or GIS, construction management sources
that describe process and results from contractors, and building operations
data from existing projects controlled by sensor-driven building management
and control systems. These systems are likely to be semi-autonomous, cloud-
based agents that operate in the background of the architect's process,
appearing when the architect demands some piece of insight or analysis.
While it is
impossible to accurately predict what will comprise this new set of
Al-empowered tools, Table 3.1.1 summarises a few speculative suggestions
designed to sketch the potential future of autonomous, Al-based tools.
systems that augment (but do not replace) the central role of the designer.
First, as has already been seen in our now data-rich world, the extensive
availability of information in digital form, combined with the predictive and
analytical power of Al systems, will make the role of evidence in supporting
design decisions much more apparent. While, as I have argued previously,
the credibility of design decisions stemmed primarily from the (presumably)
sound judgement and intuition of an experienced architect, those 8
BUILDING CODE EVALUATOR CHECKS EMERGING SCHEME FOR STANDARD CODES AND LOCAL
CONFORMANCE TO BUILDING/LIFE IMPLEMENTATION, EXAMPLES OF
SAFETY PERFORMANCE CONFORMING CONFIGURATIONS
FROM OTHER PROJECTS
SUPPLY CHAIN AVAILABILITY WORKING WITH THE PRODUCT MANUFACTURING AND SUPPLY
PROBE RECOMMENDER, EVALUATES THE STREAM SHIPPING AND
SUPPLY CHAIN CONSTRAINTS OF A MANIFESTS, CERTIFICATION
PRODUCT SELECTION, INCLUDING DATA, LABOUR STANDARDS INPUTS
AVAILABILITY, COST AND FORCED FROM LOCAL CONDITIONS, MODELS
LABOUR ISSUES. OF PAST PROJECTS
3.1.1:
FUTURE
AI-BASED
TOOLS
generate and leverage just such evidence. And since many of today's clients
rely on Al data systems to run their enterprises, architects will be expected
to do the same to substantiate the decisions that form the design.
Alberti has asserted that the architect should produce designs that are
perfectly ready for a builder to enact physically. At the end of the design
process, Carpo interprets that:
The master builder of Brunelleschi's ilk, a central repository of all things design
and construction and the maker of every decision, gives way to the architect,
generator of complete, immutable and clearly depicted ideas.
curtain walls, for example. In doing so, the procedural knowledge of building
that architects are oft accused of lacking will have been instantiated digitally
and will be accessible to them as an evaluative/performative measure of
the efficacy of their design. 11 While sharing the 'mind and imagination' with
an Al, the architect can perfect the design with substantial new, accessible
the design process are expectations that those goals will be achieved, costs
will be met, documents completed on schedule, materials properly specified
and codes conformed. At the same time, the resulting buildings consume
resources and produce carbon, require regular maintenance and staffing,
healing, goods selling. They contribute (or detract from) the environment,
economic health and social fabric of their locales. As Al systems learn from
the data derived from the built environment, and to the extent that these
characteristics model in predictive Al systems, architects get the ability to
'explicitly' design projects towards improved ends, demonstrating a priori, by
virtue of the resulting simulations, that such outcomes are the result of the
design itself. And while the earliest opportunities may be of a more limited
technical nature (as suggested in Table 3.1.1 ), more sophisticated systems will
model and evaluate larger, more complex contexts.
wave of digital tools, those driven by intelligent computation, will follow, but
shade to limit exposure and thereby reduce the size of a cooling system, for
example -
but this is hardly a strategy for the complete design of the building
enclosure. Over time, Al systems will be able to manage multiple variables while
evaluating design representations created by the architects and engineers,
instantiating expertise while simultaneously recommending alternative
solutions that meet design objectives. As designers select solutions, these
systems will come to learn which combined strategies are best, and thereby
improve their performance -
and evaluation of these issues will speed the process and make a successful
solution more likely, improving over time as the Al systems 'learn' what is
best. Our architect can use the additional time made available to resolve the
proportions of the facade.
At the heart of the potential future changes in the objectives of the design
process wrought by Al lie the implications of data and its use. Today's
architects use digital tools to create data, translate it into various forms
like drawings or specifications, and dispense it as evidence of their design
decisions. When machines can consume, create and deploy data to assist
in those decisions, the models they create extend to and are entwined with
the descriptors of the
design itself. The
resulting capabilities can make the
results of design more precise, transparent and predictable. The balance of
this section will explore strategies for enabling these capabilities in ways that
further empower architects accordingly.
3.1.3:
MECHANICAL
INFRA-
STRUCTURE
MODELLED IN
BIM FOR A
MODERN
HOSPITAL
CURATING
AND
CONSUMING
CREATING,
2
3.
DATA
DOI: 10.4324/9781003297192-15
In 1994, my former employers at Autodesk organised about a dozen
companies across the AECO industry in an effort to address a growing
concern. The company's increasingly ubiquitous CAD platform, AutoCAD©, was
becoming the data standard for the building industry with their proprietary
file format, .DWG. At the same time, the company was building a global
ecosystem of third-party developers to create additional functionality on top
of the AutoCAD© platform, and other software companies were looking to
consume AutoCAD© DWGs in their ownsystems. The file format itself was
understandably defended zealously by Autodesk, as much of their intellectual
relevant data between any software that has been written to generate or
consume it. BuildingSMART's efforts have turned from CAD data (DWG) to BIM
past years, creating standard data exchanges and libraries designed
2
in to
make BIM more open.
powered by two divergent realities: the industry was turning to digital tools,
primarily AutoCAD©, but even more importantly, the vector of information
exchange for architects and engineers was still largely drawings, rather
than more robust data. This made the transmission of information via IFC
relatively simple, using common definitions of geometry, lightly dusted with
meta-data about that geometry. Given the explosion of digital tools today, the
problem, however, is much more complex, and accepted non-proprietary data
standards for the AECO industry have not been established. 3
The move to BIM was but one part of the inevitable digitisation of AECO writ
large. Building things is an information-rich enterprise, and architects among
others had to wait until machines and networks were sufficiently powerful to
Machine Learning
broadly in the structure and use of data in the building industry, and especially
for architects. Being careful not to create too much responsibility or risk for
construction, while managing limited fees with which to produce information,
design data is held closely when released at all. Other learned' professions
carefully generate, curate and consume knowledge about their disciplines:
databases of case law, medical research about treatments and outcomes or
legal decision in the history of jurisprudence (see Figure 1.3.3 ). Architects have
access to no such central data. The move towards interoperability was a plea
3.2.1:
AN IMAGINED
FUTURE OF
BIM
INTEROPERABILITY,
IN 2013 7
3.2.2:
PRECONDITIONS OF INTEROPERABLE DATA
COMPONENTS
OF A
The advent of machine learning-based Al systems demands that our industry
POTENTIAL
not just share toys but builds a new sandbox in which to play with them. This
DATA TRUST
FOR THE is the first and most important precondition of moving towards and taking
BUILDING
complete advantage of the power of Al for architects and other players in
INDUSTRY
the building enterprise. The ability to leverage the potential of Al lies in the
profession working closely with industry partners who might also benefit, and
sharing data to do so in responsible ways.
The problem, of course, is the other external factors, not the least being
underlying motivations (or lack thereof) to share data. We will address some
of the structural risk and reward questions in Chapter 3.5 but for purposes of
,
this discussion the issue is diagrammed in Figure 3.2.2 which represents four
,
The concept of data trusts has evolved in the last several years to address
questions of information coherence, privacy and fiduciary responsibility
in circumstances where individuals contribute their personal data that is
then used, for commercial purposes, by third parties. A data trust is an
independent, third party who collects, manages, anonymises and provides
access to such a large-scale collection of data:
»
Typical use casesdata sharing are fraud detection in financial
for
services, getting greater speed and visibility across supply chains,
improving product development and customer experience, and
combining genetics, insurance data, and patient data to develop new
digital health solutions and insights. Indeed, the research has shown
that 66% of companies across all industries are willing to share data.
Nevertheless, sharing sensitive company data, particularly personal
customer data, is subject to strict regulatory oversight and prone to
While it is
unlikely that 66% of architects, or contractors for that matter,
would be willing to share data today, the benefits of access to a central global
repository of project data, properly anaesthetised for attribution, would
be too great to pass up, as both a useful reference tool and the necessary
information infrastructure to begin Al in earnest. And of course, the challenges
presented by European data sharing standards must be overcome.
in the idea that perhaps it is too expensive to invest in Al that would replace
architects as individual contributors, there is clearly benefit in data sources and
Al platforms suited for the entire industry and it is more likely that investment
will be made by vendors to address a broader market of customers.
The data sources created by a project team, including data from modelling,
of tools that will result from the widespread digitisation of the industry will
make this challenging. It is more likely that Al itself may provide the means to
standardise and conform project data from architects and others before it is
contributed to the common cause.
in her being fired from Google, computer scientist Timnit Gebru argues
that the inherent environmental implications of building Al systems
are underappreciated and accrue to the detriment of underprivileged
communities who do not benefit from them. Calling naturallanguage Als like
GPT-3, upon which Google increasingly relies, 'stochastic parrots', Gebru and
her co-authors recommend that Al systems be designed to acknowledge the
enormous contribution to atmospheric carbon they contribute by virtue of
their intensive training, and rigorously curated to design out the inherent bias
of available training data. Given that, for example, much of the content in the
proposed data trust would likely be sourced from Western projects initially,
carefully tending this data to be free of its obvious neo-liberal dynamics will
12
be critically important. Crawford further argues that the damage to the
environment of extractive mining necessary to build the enormous compute
infrastructure of Al is an externality that should be reflected in its cost and
13
development.
»AI
WILL EXTEND THE RANGE OF AUTOMATED
PROCESSES AVAILABLE TO ARCHITECTS AND
CREATE OPPORTUNITIES FOR OTHER ASPECTS
OF OUR WORK TO BE FULLY AUTONOMOUS,
OPERATING IN PARALLEL WITH HUMAN
COUNTERPARTS. SUCH CHANGES ARE LIKELY
TO FIRST OCCUR IN THE TECHNICAL
OBLIGATIONS OF ARCHITECTS, AND MORE
SPECIFICALLY IN THE RELATIONSHIP BETWEEN
DESIGN AND CONSTRUCTION, WHERE AI-ENABLED,
AUTONOMOUS ANALYSIS CAN IMPROVE
THE EFFECTIVENESS AND CREDIBILITY OF
DESIGN AND PAVE THE WAY TO OTHER AI
CAPABILITIES. «
DOI: 10.4324/9781003297192-16
3.3.1:
DOXEL’S
CONSTRUCTION
SITE INSPECTION
ROBOT, THEIR
FIRST ITERATION
OF SCANNING/
AI-BASED FIELD
VERIFICATION 1
Meanwhile, in the
analogue world of humans, standard services contracts
for architectsstipulate the architect's responsibility for construction
administration (emphasis added):
»
Carry out visual site inspections, as stated in item F of the Contract
Details, to review the general progress and quality of the works
as they relate to the architectural design and issue site inspection
reports to the Client. (RIBA Standard Professional Services Contract 2020)
4
The Architect shall visit the site ... to become generally familiar with
progress and quality of the portion of the Work completed, and to
determine, in general, if the Work observed is being performed in
a manner indicating that the Work, when fully completed, will be
While here in America we clearly have a penchant for more turgid contractual
definitions– a function of years of construction foibles and resulting litigation
– the overlap is clear. Doxel's deep learning-enabled system is designed to
either dramatically augment, or eliminate altogether, the need for human
inspection of construction progress. It is a perfect example of the potential
of AI in the design-to-construction process continuum: a Doxel robot uses
computer vision and machine learning from other projects and related BIM,
looking for something very specific ('is that column installed in the right
place?'), and creates analytical results more quickly, cheaply and accurately
than a person walking the construction site twice a day– an obligation
specifically excluded from the AIA's definition of construction observation.
MISPLACED ANXIETIES
The technical emphasis of this system, and many to follow, was anticipated
early in the last
technological transition to BIM by the architect Patrick
MacLeamy, who created what is now known as the 'MacLeamy Curve', as seen
inFigure 3.3.3 .
Originally an argument for the efficacy of BIM for architects, MacLeamy posited
that since the greatest value of the architect's work was early in the design
process – where important decisions have the best chance of positively affecting
results without disrupting progress
–
AIsystems like Doxel's, which autonomously perform technical tasks that once
required humans, is a logical extension of this same argument. However, while
construction progress evaluation will clearly benefit from additional, digitally
enabled help, a construction site is technically, geographically and, to some
extent, in a way that our little tank is unlikely to be able
politically complicated
to fully understand. So human
our architect, continuing to act on behalf of
the client to protect her interests during construction, will continue to have
3.3.3:
THE MACLEAMY
CURVE
a role even here (despite the likely reduction of
human-required production
tasks during the creation of working drawings, argued in Chapter 1.5 ). In
as
any case, since the technical aspects of design and construction– including
document production and coordination, technical evaluation like code
compliance or coordination, and managing and evaluating the information
flow from the construction process are the most likely candidates for AI-driven
-
RECONSIDERING RELATIONSHIPS
Each of these roles requires the innate human ability to understand context,
manage relationships and make trade-offs and judgements; these are
tasks that precisely the opposite capabilities of AIs, and particularly those
are
whose neural networks require training in large, well-curated data sets. The
multi-valent responsibilities of the architect, writ large as in this definition, are
therefore unlikely to be replaced wholesale by computers, hence the argument,
consistent with Susskind, about task replacement: limited tasks, perhaps, but
production in the name of efficiency. And as the demands for more intelligent
supply chain decisions and management increase in an era of climate change,
reduction of toxicity or even attempts to reduce forced labour, the architect's
specification of materials will tie even more closely with an understanding of
the supply itself. Contractors will therefore come to rely on architects for such
7
intelligent decisions. This is an inversion of the relationship of technology to
design that emerged with the blob-makers of the 1990s, when the architect's
digital shape-making tools made possible ever more elaborate forms that
were left to engineers and builders to actualise.
today are more precise, better coordinated, more accurate and useful to
contractors than their CAD predecessors, and BIM is a good platform, as both
a representational schema and a training ground, for AI to continue this trend.
As many of the tasks of the working drawing phase are procedural in nature,
it should be relatively straightforward to train AIs to perform them. Other
objectives of the hand-off between designer and builder will still demand
the judgement and coordination of the architect, who will be supported by
machine intelligence in areas where large data sets, pattern matching and
complex calculations and predictive algorithms could be of most use.
3.3.4: TASK
ANALYSIS
OF THE
TECHNICAL/
PRODUCTION
STAGES
OF THE
ARCHITECT’S
WORK
To that end, consider the problem of construction cost projection, and
particularly the heated negotiations oft held between designers and builders
about conformance to target costs. The dimensions of this complex dance of
data, judgement and computation are sketched in Figure 3.3.5 .
3.3.5:
ANALYTICAL
ELEMENTS OF
CONSTRUCTION
COST
Quantity surveying: Design documents define design intent, and builders
must interpolate precise system characteristics, unspecified materials and
of writing, material prices have jumped as much as 40% (timber) since January
2021, as the US economy recovers from the pandemic; a similar pattern was
seen in the recovery from the 2008 crisis. AI-based analysis and predictions,
based on previous economic models and pricing profiles, could elucidate the
potential implications for pricing, bidding and market conditions that affect
cost models.
Cost of labour and materials: The costs of both labour and materials are
estimated traditionally from historical information, but the actual pricing of
projects generated by builders is a combination of historical data, in particular
competitive advantages or disadvantages based on the builder's capabilities,
as well as larger market forces such as the availability of skilled workers and
CONSTRUCTING AUTOMATION
The painting robot we examined in Chapter 1.4 signals important changes for
designers, not the least of which is the likelihood that their projects will be
festooned with more precisely applied colours and textures. As Negroponte
suggested in his early explorations of the automation of processes by
machines, a first step towards incorporating digital technology is to use
it to replicate an analogue process, and surely this is where our painting
robot will begin. However, Negroponte further speculated that once these
processes are fully integrated and understood, their capabilities will expand
10
far beyond what their originators could have anticipated. Our painting robot
will improve its technique by repeatedly painting surfaces and sharing the
lessons of its success and failures with other robots doing the same on other
3.3.6: projects. Further insights will be supplied by data coming from the construction
MATERIAL
supervision robots made by Doxel and their inevitable competitors. Eventually,
PRICE
the robotic painter will have AI-automated, autonomous colleagues installing
VOLATILITY
IN THE US, and assembling other aspects of the construction project, and the procedures
11
2006-21 and protocols they generate could combine into an accessible, evolving source
of construction insight that could truly modernise building.
PARTNERS
USING A BOSTON
DYNAMICS
ROBOT FOR
CONSTRUCTION
PROGRESS
ANALYSIS 11
DESIGN
OF
LABOUR
4
3.
DOI: 10.4324/9781003297192-17
My first full-time job in an architect's office was in 1979, the pre-CAD era, when
our work wastaped large
to drafting tables and prepared with plastic lead on
giant sheets of mylar. The small office in Charlotte, North Carolina was known
as a solid, if stolid, practice that did complete working drawings that resulted
in routine buildings which did not leak or miss theirbudgets. I was assigned to
help prepare construction documents for a bland shopping mall in Tennessee
that still haunts my dreams.
studio, as there was so little design to speak of going on at least two or three
-
of whom were architectural draughtsmen (no women). These were older guys,
some with architectural degrees, none with licenses, who drew and lettered
beautifully, knew a lot of about how to put a building together, could never be in
front of a client and were constant sources of knowledge to the younger, better
educated but far less experienced architects-in-waiting like myself. In that era,
many offices of any size had folk like this, whose main job was to draw, leaving
all other architectural responsibilities to others. Late in my 15-month stint in this
firm, the office manager began researching a new idea called 'computer-aided
drafting', which the draughtsmen dismissed entirely as a gimmick.
Two decades later, these sorts of wise but unregistered drafters were largely
missing from practices, unable to make the transition to CAD. Their jobs were
replaced by young, digitally enabled CAD operators who were unafraid of the
computer and had the skills to use it to draw. Almost no one in an architecture
firm was trained or hired at this point as a drafter; young designers in training
filled these roles. Just as today's secretaries no longer do much typing but are
more general support staff, architectural jobs were no longer differentiated by
3.4.1:
THE DRAFTING
ROOM AT
SKIDMORE,
OWINGS &
MERRILL’S
OFFICE,
CHICAGO,
1958
Machine Learning
production tasks (drafting) but more around roles and responsibilities (design,
construction administration, specifications).
As firms are completely reliant on technology of all sorts these days, there
are specialists in network management or even coordination of BIM data, but
such roles are limited and certainly do not contribute to billable work. And even
if yesterday's 'CAD monkeys' are today's 'BIM monkeys', BIM work is not routine
drafting, as using a BIM tool requires a strong understanding of how a building
goes together and how to properly represent it. But even today, while larger
firms are hiring specialists in data management or, in some cases, software
development, practice is largely devoid of technological specialists.
The advent of machine learning tools will reverse this trend. It will create
demand for different sorts of architectural workers. As I have argued (see
Chapter 2.4 ) that artificial
general intelligence (AGI) is far in the future -
particularly AGI that can step into the multi-faceted role of a proper architect
-
support the broader enterprise of design. Those systems will require specialised
understanding of inputs, outputs, data demands and relationships of the Al
system to the broader infrastructure of design information. These are skills that
architects trained as generalists are unlikely to understand, nor, frankly, have
much engagement with: the outputs of such systems will be of great interest;
the process by which they are generated, not so much. While it would be nice
to simply ask the 2030 version of Alexa, 'How much carbon is embodied in my
I have argued up to this point that Al systems are likely to automate the
more routine aspects of technical drawing production, data and document
However, there will be new jobs. Architecture will need experts who can
manage these systems in production, particularly in relation to affiliated Albased
processes that will translate knowledge and insight from construction
back to design. Complex technical analysis and building performance
evaluation will be a necessary element of design generation. The dilemma
3.4 Labour of Design
for practices will be mapping demand for the specialists to handle such work
and the available supply of talent. If an Al-based system is helping optimise
the carbon footprint or forced-labour-free supply chain procurement of your
building in the phase of conceptual design, that work is episodic at best, and
probably punctuated with interactions with other intelligent systems. The data
trust proposed in Chapter 3.2 will provide ubiquitous data, but using it well
will be challenge.
for data scientists and machine learning experts, many of whom earn starting
salaries well above those of their staff architects in UK practices. 2 While
expertise in Al will grow in the next decade, matching supply and demand will
be a challenge.
Automating this work, at least under current business models, means fewer
workers. The so-called 'canonical model' described by Susskind -
where jobs
destroyed by innovation are replaced by the jobs required to create the
new technologies has given way to a new thesis, the so-called 'Autor-Levy-Murnane
-
(ALM) Hypothesis', which declares that the routine tasks of work will
be eliminated by computers, as those tasks are the easiest to teach machine
learning systems to replicate. Skilled workers, like my early drafting colleagues,
4
are eventually replaced, resulting in fewer jobs that do not reappear.
Or at least fewer jobs will reappear than disappear, given the current demand
for architects. However, there is another scenario, where Al technology
empowers architects to the extent that demand for professionals even those
-
doing different jobs, like their counterparts in the early days of CAD will be -
much higher. The history of BIM in the UK suggests that this might be the
case, as demonstrated in Figure 3.4.2 .
3.4.2: Despite the inherent efficiencies introduced by BIM and its adoption driven by
DEMAND FOR
the Level 2 requirements, which in theory should have decreased architectural
UK ARCHITECTS
AND RELATED positions, and even despite a drop in designed construction value in 2020,
CONSTRUCTION the number of architectural positions in the UK has steadily increased for a
VOLUME decade. Perhaps better work begets more employment.
DESIGNED 5
Practice in the era of Al will therefore entailarbitraging the value of expert labour,
machine production and data. The profession will need to find a way to access
this talent across many scales of business, particularly since most firms, world-wide,
are relatively small, averaging less than ten staff each. While larger firms,
6
with more human and financial resources, may get to Al capabilities first, for real
change to be possible the wider profession must be able to access Al assets.
An intelligent design of the industry data trust described in Chapter 3.2 would
entail a central marketplace for such talent and the infrastructure to make
it accessible. The technologies of the so-called gig economy, which match
demand with capacity, are the template for such a system. Like ride- and
apartment-sharing services today, these systems use artificial intelligence
themselves to align unused resources with those who might need them. Given
that Al enactment in design practice will be task-oriented, perhaps the data
trust might manage such a platform as part of the value proposition it puts
forth to the global design industry.
3.4.3:
BUSINESS
MODEL OF
AIRBNB,
MATCHING
HOUSING
CAPACITY WITH
TRAVELLER
DEMAND
require employers to limit hours and provide minimum pay for overtime,
characterising them as 'professional workers'. Doing rote CAD work, not so
match workers with jobs have been demonstrated to instantiate the biases of
the data set (like résumés of employed workers), denying qualified applicants
opportunities generated by machine-based hiring systems. Workers in the
extractive industries that drive modern computation are often exploited and
In her ill-fated paper for Google, Timnit Gebru makes the case that Al
development must occur in parallel with an understanding of its broad
implications. She suggests:
our clients and our workers, and the profession working in concert with
-
providers and the academy, would best heed her advice and begin plotting
the route to equitable Al today.
The ALM hypothesis (as discussed in Chapter 1.5 ) proposes that highly skilled
workers -
like the ones that will create, develop and manage Al will remain in -
demand once machines carry out knowledge work. Very low-skilled but highly
localised jobs, like those in personal services or dining, will also remain. In
Susskind's 'massacre of the Dilberts', many jobs in between will be eliminated.
Most architectural jobs in this scenario are probably safe, but the society for
which we design buildings is likely to be dramatically affected. Like other issues
of social equity, it is best to add this question to the list that the profession must
address to responsibly design the built environment of the future.
PROPOSITIONS
BUSINESS
VALUE
5
3.
AND
MODELS
DOI: 10.4324/9781003297192-18
Machine Learning
» We're paid on a percentage of the cost, but the capable architect
is the one who keeps down the cost. Therefore, by doing his best he
reduces his compensation... The client wants to keep the cost down, and
his architect must help him in this, but the less the cost of a particular
job, the less the compensation and the less likely to be the beauty of its
1
execution from which the architect obtains his reputation. «
The idea that an architect should be paid in some proportion to the cost of
construction seems to have originated centuries ago in Europe. Johnston
explores this trajectory and quotes Benjamin Latrobe, the so-called 'first
architect' of the United States, European-trained, British immigrant to America:
COMMODITY RESISTANCE
3.5.1:
KENT
ROCKWELL’S
'FEES: A
REDUCTIO
ABSURDUM’ FROM
ARCHITECTURE
AND BUILDING,
46, 1914 6
There is no question that those technologies improved the processes and,
in some ways, the results of architectural services. CAD made drafting more
accurate and efficient, while allowing architects to depict projects that were
technically and formally more complex. BIM has allowed all members of
the delivery team to generate, organise, integrate and exchange design
information at much higher levels of resolution and transparency. It also,
in some minimal way, begin to bridge the information gulf between design
and construction; builders who saw the value of 3D data began to request
it to assist their work. Other digital technologies have improved information
the real value of design work lies early in the delivery process, despite the
relatively small degree of effort entailed there compared to production and
delivery stages. Perhaps AI will begin the value shift.
For the last several years we have offered a course at Yale called 'Exploring
New Value Propositions of Design Practice', where our students are asked
to interrogate the business models of architecture with the assumption that
better jobs can be done
designing them. Each semester, teams of students
create what they believe to be the most provocative new models for practice,
in response to the question: 'Where can the value of architectural services be
best translated into a business model?' The projects are required to conform
tojust two conditions: the proposal must be based on something that a
competent architect is capable of providing, and the compensation strategy
must completely abandon any vestige of commoditised pricing, so no fixed or
After few years of teaching this class, several consistent strategic themes
a
emerged from the students' research, which range across the opportunities of
services, organisation and products. These ideas were congruent in that they
7
challenges, or to clients as a validation service for other designs. An architect with ADDITION,
ELEVATION
deep expertise in a given discipline, say healthcare design, may have made
a significant investment to develop AI-based analysis systems to evaluate or
university or retail operator has designed numerous buildings for them. Mapping
-
design model data with data streams from building control operations systems,
3.5.4: she has collected a large enough data base of this client's building base that,
TALLY©
combined with data from the trust, has allowed the firm to create an AI-based
CARBON
ASSESSMENT optimisation tool that can apply to future designs as well as tuning the operation
TOOL of current assets. The
resulting contract, extending in blocks of five years
post-occupancy for every building in the portfolio, requires the architect to
manage and evaluate the data streams from projects, evaluate operational
optimisation and make recommendations to the client. In addition to an
annual service contract fee, the architect is also paid a small percentage of the
operational savings in energy, maintenance and staffing resulting from these
services. Today, architects
EskewDumezRipple reserve 2% of fees to provide
post-occupancy clients, discovering strategies for improving future
services to
These short vignettes are meant to stimulate thought about what alternative
value propositions for architects, underpinned by the new capabilities of AI, might
look like in the future. More important, however, are considerations of how, if at
all, the basic services of architects at the core of our value directly designing the
-
3.5.5:
UNCERTAINTY
built environment -
may evolve when we share the job with intelligent machines. FACTORS IN
It is naïve to believe that architects can immunise themselves completely from
PROJECTS,
the pressures of productivity improvements and knowledge work replacement ACCORDING TO
A SURVEY OF
that these systems will inevitably bring, and further that a profession that has
US DELIVERY
operated with essentially the same business model since before the invention of PARTICIPANTS 9
electricity can, within a generation, turn to fundamentally new busines strategies.
We are stuck with what we have, but can we fix it?
in
advance. A dashboard is shown in Figure 3.5.7 .
Note that neither of these companies is using AI to replace the work of human
construction coordinators, risk managers or safety leaders, but rather augment
their capabilities and allow them to significantly improve their performance.
There is no reason why such an approach, applied to architectural design
3.5.6: process, could not go right to the heart of client and contractor uncertainty
WHEN in design process, be it checking for properly coordinated construction
PROJECTS
documents, cost prediction, lifecycle modelling of materials for durability/price
MEET
EXPECTATIONS, trade-offs, and even missing information from the documents that comprise the
OR NOT 10 contract for construction.
3.5.7:
PROBLEM
DASHBOARD
FROM BIM
360 IQ
The definition of good design must move, at least in part, into the numerative,
environmental and social impact and the proper design and implementation
-
of AI systems is the key to this change. Numerate and talented designers and
the results they can create will have the dual benefits of increasing the credibility
of architects and firmly anchoring our value in the building supply chain. And
perhaps building an AI-assisted numeracy in setting performance objectives and -
getting paid for them will achieve the alignment of value and business models
-
DOI: 10.4324/9781003297192-19
Our profession is, at its core, a creative enterprise that is valued for our ability
to both define and solve problems in unique, appropriate and beautiful ways.
Using Al should extend, rather than exterminate, that obligation to our clients
and the public at large.
While it may take some time until it reaches the far corners of the design
and construction professions, eventually the pattern analysis, autonomous
processing and data evaluation capabilities of AI/ML will appear in the
architectural landscape. We can declare, a priori, that the technology
represents an existential threat to architects, or we can use our problem-solving
skills and design a strategy that determines their ultimate destiny and
use. And in doing so, we can increase the influence and credibility of architects
with clients, mend decades of broken relationships with our construction
collaborators and maybe even break the chain of our commoditised value
propositions and stunted fee structures.
1. Explicitly guide the definition and creation of technologies that will frame
future practice. Given that the next generation of technology may well define
the future of architectural practice, the profession must establish means to
declare its needs and direction in a way that does not defer to the business
of design in its traditional sense, but rather expand it, while simultaneously
enlarging the effectiveness of building and the credibility of architects.
3. Create the data infrastructure that can serve as platforms for design.
Today's designers have created tens of thousands of digital models,
mostly through BIM, of projects that represent an enormous resource
for Al-generated insight. Contractors are doing the same with drone
scans, computer-vision analysis of construction and digital construction
management tools. Building control systems are generating huge lakes
of digital information about systems performance. The potential of these
resources is wholly unrealised without a strategy to organise and access
If the necessary data can be collected, and the systems validated, architecture
can expand its remit dramatically. In a recent article addressing the fraught
between architecture and incarceration, Garrett Jacobs and
relationship
Deanna Van Buren, leaders of the non-profit design alliance, Designing Justice
+
Designing Spaces, argue that architects must apply their skills to 'end the
racism that is embedded in the built environment'. To create prototypes for
their 'Alternatives to Incarceration Plan' for Los
Angeles, they have
implemented
adesign process deeply dependent on complex, interrelated data sources,
explaining that 'We are partnering with data visualization, mapping and
research organizations to understand how various systems– such as health
care, first response, pre-arrest diversion, housing, post-incarceration re-entry,
and more – interact at a district scale.' 10 In a pre-digital era, even collecting this
data would have been impossible, much less evaluating it or deploying it as the
basis for new design. At this juncture, when empiricist systems are coming to
the fore, but cognition is largely missing, human designers play an irreplaceable
role to direct data and marshal it. Only then will strategies for computation and
design combined accomplish completely new results that humans create to
improve our condition. This is the optimal outcome of the highest and best uses
of artificial intelligence in architecture.
BIBLIOGRAPHY
The Simple Economics of Artificial Intelligence Harvard , Cost & Carbon Improvement: A Report for The Government
Business Review Press Boston, MA 2018 , , . Construction Client Group UK Government Department of ,
–
The Architect's Handbook of Professional Practice: 15th Brynjolfsson E. and McAfee A. The Second , ,
'
Lectures given at Architectural Association London March Do? Workforce Implications', Science Vol. 358 w, https://doi.- ,
, ,
1966 and ASCA Cranbrook, 5 June 1966. org/10.1126/science.aap8062 2017 (accessed 27 December ,
2020).
Architects Council of Europe with Nacey Research The .
Practice 2017 , .
'
–
The Digital Turn in Architecture 1992-2012 , Wiley , Chichester ,
Cambridge, MA 2017 , .
'
Press New York 2009 . , .
–
Canonical Models of Architecture ', LOG: Model Behavior Vol. ,
Deamer P. The Architect as Worker: ,
'
Bernstein P. Intention to Artifact' in Digital Workflows
, Engineering Industry Study: Trends and Benchmarks in EMEA and
in Architecture: Design–Assembly–Industry ed. Scott Marble , ,
APAC Deltek 2021, , .
'
–
The Distractions of Disruptions: Technical Supply in an Era Enough for General Al. BDTechTalks.com https://bdtechtalks.- .
'
–
Innovative Tools in the Movement Toward Slave-Free –
'
Folkens B. Chihuahua or Muffin? Cloudsight https://-
, , , Jacobs G. and Van Buren Deanna Decarceration
, , .
London 2020 .
the Dangers of Stochastic Parrots: Can Language Models Be ,
'
Too Big?'FaccT'21 Virtual Canada 2021 , , , . Leach N. Design in The Age of Artificial Intelligence ', Landscape
,
' '
Foundation 2021 , .
–
York 2010 .
China 2020 , .
,
Hao K. The Coming War on the Hidden Algorithms that Trap Marcus G. and Davis E. Rebooting AI: Building Artificial
, ,
,
https://www.technologyreview.com/2020/12/04/1013068/- Marcus G. The Next Decade in AI: Four Steps Towards Robust
,
'
–
These Weird, Unsettling Photos Show That AI Is Getting –Artificial Intelligence Is Stuck. Here's How To Move It
Smarter, MIT Technology Review https://www.technology-- , Forward ', The New York Times 30 July 2017 , .
–Inside The Fight to Reclaim AI from Big Tech's Control, MIT Metz C. One Genius' Lonely Crusade to Teach a Computer
,
com/2021/05/27/1025453/artificial-intelligence-learning--
create-itself-agi/ 27 May 2021 (accessed 6 June 2021). ,
Mokyr J. Vickers C. and Ziebarth N. L. 'The History of
, , , ,
'
Moretti E. and ProQuest The New Geography ofJobs
, . ,
NH 1977
, .
mckinsey.com/business-functions/mckinsey-analytics/-
our-insights/an-executives-guide-to-ai 2020 (accessed 12 ,
Registration Examination® (ARE®) 5.0 Handbook NCARB , ,
– The Next Normal in Construction: How Disruption is Reshaping Negroponte N. The Architecture Machine MIT Press , , ,
Cambridge, MA 1970 .
the World's Largest Ecosystem 2020 , .
,
'
Parman J. Is Architectural Licensing Necessary?' Common Susskind D. A World Without Work: Technology, Automation,
,
Edge https://commonedge.org/is-architectural-licensing--
,
And How We Should Respond Metropolitan Books/Henry Holt ,
necessary/ 2020 (accessed 31 October 2020). , & Company New York 2020 , , .
Rittel H. and Webber M. Dilemmas in a General Theory of P. Bernstein Princeton Architectural Press New York 2010, , , .
, ,
'
Architectural Design ',Journal of Architectural Education Vol. 36 , , UK Building Information Alliance Information Management .
issue 1 1982 pp 18 23 , ,
-
.
According to BS EN ISO 19650 Guidance Part 1: Concepts –
.
' gov.uk/construction/areyou/principal-designer.htm ,
–
Standard Professional Services Contract 2020: Architectural 2016 (accessed 6 June 2021).
Services', RIBA,London 2020 p. 48 , , .
'
–
in Action Basic Books New York 1983 Witt A. Shadow Plays: Models, Drawings, Cognitions, LOG: Model
, , , .
,
timnit-gebru-ai-what-really-happened/ 2021 (accessed 10 , Intelligence Selected Papers from the 1st International Conference
June 2021). on Computational Design and Robotic Fabrication (CDRF
,
2019) https://yale.idm.oclc.org/login?URL=https://doi.-
,
2019 (accessed 6 June 2021). The Ethics of Artificial Intelligence Cambridge Handbook ,
A. Ganzoni D. and Carpo M. The Working Drawing: of Artificial Intelligence, Cambridge University Press
Spiro , , , ,
,
F. and Kent R. Architec-tonics: The Tales of Tom Zarkadakis G. "Data Trusts" Could Be the Key to Better
Squires , ,
,
York 1914 , .
could-be-the-key-to-better-ai 2020 (accessed 11 June 2021). ,
,
'
Staff T. A Wave of Venture Capital Is Pouring Into Zweig Group 2019 Fee & Billing Survey Zweig Group . .
https://therealdeal.com/national/2019/07/03/a-wave-of--
venture-capital-is-pouring-into-construction-tech-sector/ ,
Cambridge, MA 2017 , .
REFERENCES
laptop I am writing on at this moment Resource NY 14 Ajay Agrawal Joshua Gans and
has 16GB of internal memory, or , . ,
PDP-11. It also has yet to catch fire. The Simple Economics of Artificial
Carpo, The Alphabet and the Algorithm .
2 https://en.wikipedia.org/wiki/PDP-- 4 Architecte
Intelligence Harvard Business Review
,
1.2
Drawing: Architecture in the Age of 5 Leon Battista Alberti On the Art ,
5 This 2017 p 18
particular text was created by a Cambridge, MA 1970 , .
, .
try yourself at https://appinferkit.com/- information exchange is that 3 Noam Chomsky at MIT, Terry
generate (first accessed 21 April 2021). it requires no special software nor Winograd at Stanford University and
6 Personal email exchange with Dr data standards, beyond graphic Roger Schank at Yale University (with
Mark Greaves of Pacific Northåwest convention, to transmit or translate it whom I studied in the late 1970s)
National Laboratory 27 November , between originator and consumer. each put forth competing theories of
2020 .
9 The aerospace, automotive and cognition from which various early AI
7 Erik Brynjolfsson and Andrew efforts developed.
manufacturing design disciplines had
Mcafee The Second Machine Age: Work,
, deployed modelling tools decades 4 Gary Marcus and Ernest Davis ,
Progress, and Prosperity in a Time of before architects. Similarly, the Rebooting AI: Building Artificial
Brilliant Technologies W.W. Norton & , form-making and rendering tools of movies Intelligence We Can Trust Pantheon ,
Company New York 1st edn, 2014 , , . and games (such as Autodesk's Maya) Books New York, 2019 p 41
, , .
'
,
example, '
see
Architecture', MIT Press Cambridge, , 24 November 2020). systems were trained to study millions
MA 2011 xi and Mario Carpo The of pictures of cats (readily available
, , , ,
13 Phillip G. Bernstein Architecture, ,
on the internet). Eventually, the
Second Digital Turn: Design Beyond
'
data sets, like all the text ever posted How Technology Will Transform the
17 There are several examples of
online. It is not always successful, Work of Human Experts Oxford ,
10 See Pedro Domingos The Master , process failures (such as delays or 3 Donald A. Schön The Reflective ,
Algorithm: How the Quest for the change orders); BuildTech start-up Practitioner: How Professionals Think in
Ultimate Learning Machine Will Remake Smartvid.io uses computer vision Action Basic Books New York 1983
, , , ,
'
Decade in AI: Four Steps Towards 18 Cade Metz One Genius' Lonely , Theory of Planning', Policy Sciences 4 ,
p 51 < https://arxiv.org/ftp/arxiv/-
, Common Sense ', Wired 24 March ,
5 For statistics on the generally
papers/2002/2002.06177.pdf> 2016 .
dismal state of architectural
(accessed 19 December 2020). 19 Marcus op. cit , . ,
p3 .
graduates see Kermit Baker et ,
'
12 For an explanation of a theory 20 Ibid al. Compensation Report 2019 ',
p 27
,
., .
Judea Pearl and Dana Mackenzie The , 22 See Niall McNulty Introduction to ,
comparison is the exchange
Book of Why: The New Science of Cause Bloom's Revised Taxonomy < https://-
programme that has existed between
,
org/wiki/Cyc > (accessed 31 December project DeepMind, said: 'It doesn't program cannot use their degree to
2020). play like a human and it doesn't play qualify for licensure in the opposite
'
like a program. It plays in a third, country.
15 See Stanford Anderson Problem-Solving ,
almost alien way.' Cade Metz, 'In '
problem-solving and problem-worrying, by McKinsey Consulting that used 8 National Council of Architectural
to explain the possibility of AI in
as Anderson defined it, Registrations Boards (NCARB) ,
Holt & Company New York 2002, very widespread adoption did
'Writing Architecture', MIT Press ,
2020 p 129
, ,
Digital Buildings? paper presented at reported in its 2020 Firm Survey that
at the Allen Institute for AI < https://- ,
,
realistic feline visages but could not How Technology Will Transform the 7 Susskind op. cit , . ,
p 78 .
Architecture, Design, Data: Practice Models, Drawings, Cognitions', LOG: 9 Institute of British Architects
Royal
Competency in the Era of Computation ,
Model Behavior 50 2021 '
ask everyone: 'Remember the second Section 3.3.15 (accessed 1 February 1 Typical models vary by the role
version of the core that we suggested of the contractor, the relationship
2021).
for the first tower at Canary Wharf? between the designer and the builder,
That would work here. 'Those of 10 This robot is designed to paint the number of contracts between
us who were younger, and far less
an interior wall in concert with a various participants and the client,
human painter, see B. Li I. Chen
experienced, found this ability both '
,
and the timing of ascertaining the
and E. Asadi Pictobot: A Cooperative
extremely useful and incredibly cost of construction. Examples include
at least in the UK. See Sir Roger drawings are omitted and the January 2018.
Egan, 'Rethinking Construction: The design itself is documented in detail 6 Economic data on the performance
Report of the Construction Task Force' in shop drawings prepared by of the US architectural industry is
(1998). However, the systematic the subcontractor who is going to
exceedingly difficult to collect, as
exploration of these questions may fabricate and install the system. This the best potential source, the AIA,
be seen to have begun much earlier approach acknowledges that the is constrained, as described below,
by the National Joint Consultative actual, useful value of 'design intent' from collecting and evaluating it.
Committee of Architects, Quantity construction documents for such As such, this chart was derived
Surveyors and Builders in a study systems is minimal, and therefore from a combination of AIA sources,
prepared by the Tavistock Institute of they can be abandoned. specifically the work of Kermit
Human Relations; see Gurth Higgin ,
9 Higgin Jessop
, and Tavistock Baker cited in n. 5, plus additional
William Neil Jessop and the Tavistock Institute of Human Relations London information provided by the AIA
,
step in the delivery process: design, 12 Andrew Rabeneck The ' Place of ,
are extrapolated from available
fabrication, and construction. That Architecture in the New Economy', data on revenue (indicated by the
inefficiency is leading to modern Industries of Architecture eds. Katie ,
solid blue bars). Sources: personal
attempts to integrate and streamline Lloyd Thomas et al Routledge , , correspondence with Kermit Baker,
the process. For example, the list of Abingdon, Oxon 2016 p 192 , , . 15 May 2021; AIA Firm Surveys 'The
companies that prefabricate large 13 Ibid ., p 192 .
Business of Architecture', 2002 2005 , ,
2.1
each fiscal quarter: Bryden Wood, Zweig Group, '2019 Fee & Billing
'
is a collaborator on a study of such destruction > (accessed 30 April 2021 ). Will Swearingen 2019, p 208
, .
4 Higgin, Jessop and the Tavistock conversation with the Chief Economist
Work: Technology, Automation and
Institute of Human Relations London ,
How We Should Respond Metropolitan of the American Institute of Architects,
,
op. cit p 77
. , .
data that enables AI and its ethical own that software, you are merely
hse.gov.uk/construction/cdm/2015/-
and economic implications in the era principal-designers.htm > (accessed 7
licensing its use for your purposes, June 2021).
of big tech, see Kate Crawford Atlas ,
and those purposes had best conform
of AI: Power, Politics, and the Planetary to the extensive terms and conditions
3 For a detailed explanation of
Costs of Artificial Intelligence Yale ,
of the EULA, which you are very American phases of work, see
University Press New Haven, CT , ,
unlikely to have bothered to read.
American Institute of Architects The ,
2021 .
'
Architect's Handbook of Professional
10 See Autodesk LICENSE AND Practice: 15th Edition Wiley Hoboken,
10 Private conversation between SERVICES AGREEMENT', < https://-
, ,
download.autodesk.com/us/FY17/-
National Laboratory and the author, Suites/LSA/en-us/lsa.html > (accessed 4 It is instructive that there is no
16 April 2021. 6 June 2021). British counterpart to the American
11 See Anouk Ruhaak How Data ,
'
4 Health and Safety Executive of HM refurbishment work should not have '
uk/construction/areyou/principal-- Tower fire: Inquiry told firms "deny Architectural Intelligence, Selected
designer.htm > (accessed 6 June 2021). Papers from the 1st International
responsibility'" at < https://www.bbc.-
5 National Council of Architectural com/news/uk-51256738 > (accessed 6 Conference on Computational
'
Report 2020 ', NBS London 2020 p 9 , , , , Books New York 2018
, , .
'
Do in Brockman p 52 , .
13 See Grace Farms Foundation , in visualisation, where theoretical or learning, where the system is
'
Design for Freedom ', Grace Farms historical material is connected to programmed to try alternative
Foundation 2021 < https://www.-
, .
drawing or digital tools. One summer strategies and instantiate successes
designforfreedom.org/wp-content/- course of three credits is dedicated and reject failures; they are the
uploads/2020/10/DesignforFreedom_- directly to BIM, in connection current rage among AI programmers
FullReport_L.pdf> (accessed 7 June with a building construction and in 2021, as described in Ben Dickson,
2021). documentation requirement. 'DeepMind Scientists: Reinforcement
14 Dennett in Brockman p 46 , . 6 Dennett in John Brockman Possible ,
Learning Is Enough for General AI',
Minds: Twenty-Five Ways of Looking at BDTechTalks.com 2021 < https://-
, ,
2.4 bdtechtalks.com/2021/06/07/-
AI Penguin Press New York 2019
, , , ,
p 46 .
deepmind-artificial-intelligence--
AI can't solve unknown problems ', reward-maximization/ > (accessed 8
7 Russell in ibid p 32 .
2 For two examples of early but of this writing, NCARB was completing a fast-food restaurant or an Apple
unrequited enthusiasm for BIM, see the 2020 Practice Analysis. Store might not be suitable for AI
–
CT,: 2011 ; Phillip G. Bernstein and and ISO 19650 ; see < https://www.- difficult to imagine an AI sophisticated
alone. The building industry adopts 11 UK BIM Alliance ,' Information Computation Birkhäuser Basel 2018
, , , .
new technologies and processes at Management According to Bs En Iso 9 Cost modelling in the early stages
a glacial pace, and curricula evolve 19650 Guidance Part 1: Concepts',
–
of a project are parametric (area
even more slowly, even when there buildingSMART London 2019 , , . × cost/unit of area), where later in
2020' (accessed 11 June 2021). unfortunate behaviour first learned Avi Goldfarb Prediction Machines:
,
4 See in school and then transported to the The Simple Economics of Artificial
Liverpool's Building Information
office studio. Intelligence Harvard Business Review
Modelling and Digital Transformation ,
'
Politics, and the Planetary Costs of <
https://architosh.com/2018/01/-
1 See BuildingSMART', < https://- Yale University
Artificial Intelligence , facebooks-vc-backs-new-doxel-ai--
en.wikipedia.org/wiki/BuildingSMART> Press New Haven, CT 2021 p 12
, , , .
and-computer-vision-to-disrupt--
(accessed 11 June 2021). construction-industry/ > (accessed 16
10 'Veridical'
meaning 'truthful,
2 During my time at Autodesk, I September 2021).
veracious, genuine' < https://www.- ,
in COBIE (Construction Operations Margaret Mitchell On the Dangers , 9 While these conclusions are now
of Stochastic Parrots: Can Language being substantiated by firms outside
Building Information Exchange) '
which is, in essence, a spreadsheet-based Models Be Too Big? in FaccT'21 , the building industry like McKinsey,
extraction of non-geometric Virtual Canada 2021 , , .
many of these concerns were defined
data designed to expose certain 13 Crawford op, cit , . ,
Chapter One by Roger Egan in Sir Roger Egan,
characteristics of the underlying 'Earth', pp 23 ff. 'Rethinking Construction: The Report
of the Construction Task Force',
design. This is a complicated set 3.3
of standards and interactions. A 1998. In the US, statistics about
conformance to budget and schedule
good place to start to understand 1 This robot was Doxel's first iteration
them can be found at https://www.- for such a process. In subsequent range from 30% to 60% of projects
designingbuildings.co.uk/wiki/BIM_- version, the company developed failing to meet these goals.
level_2 (accessed 12 June 2021). a technology called VSLAM (vision-based 10 Nicholas Negroponte The ,
4 This move can come none too soon, simultaneous localisation and Architecture Machine MIT Press , ,
as construction is considered one of mapping) that is much less expensive Cambridge, MA 1970 , .
the least digitised (and productive) and easier for construction managers 11
'
Foster
+ Partners and Boston
to deploy. Email correspondence
industries globally. See McKinsey monitor construction with
Dynamics
Global Institute ,'Reinventing between the author and Kevin
'Spot'the robot dog ', Design Boom ,
5 https://www.buildingsmart.org/- built asset marketplace. See What 1 Deltek Clarity with Longitude,
'
about/vision/ (accessed 11 June 2021). does BuildTech mean?', < https://www.- Deltek Clarity Architecture and
archdaily.com/924827/what-does-- Engineering Industry Study: Trends
6 Michael P. Gallaher Alan C. ,
buildtech-mean (accessed 11 June and Benchmarks in Emea and Apac in
'
'
O'Connor et al. Cost Analysis of Annual Comprehensive Reports Deltek
,
2021). , ,
of Using Bim: A Cost Consultant's Contract 2020: Architectural Services', 3 Construction Documents is typically
RIBA London 2020 p 41 35% of the basic fee, and Construction
Perspective', 49th Annual Associates , , , .
Schools of Construction Conference 5 American Institute of Architects Contract Administration another 20%.
, ,
the 1980s and 1990s, where standard Machine: The World of AI Powered
Books/Henry Holt & Company New ,
5 The
' Architectural Profession in
Europe 2020 '. anti-competitive. The AIA has therefore 2 Recently, a collective of UK architects
removed itself entirely from any has expressed strong misgivings about
6 According to the Architects' Council discussions whatsoever of fees. Autodesk's lagging development of the
of Europe, 97% of firms in Europe have
4 Nicolas G. Carr It Doesn't Matter in
' '
flagship BIM platform, Revit. See Daniel
10 or fewer staff. ,
'
Davis Architects Versus Autodesk
'
'
Harvard Business Review 2003 < https://- , ,
,
"
interns shouldn't be given prestigious 4 Issues of forced labour and modern
commissions says designer who
" 6 Ibid ., p 117 .
'
5 See Phillip G. Bernstein A Way ,
Worker: Immaterial Labor, the Creative 8 Lance Hosey Going Beyond the Forward? Integrated Project Delivery
,
',
Class, and the Politics of Design , Punchlist: Why Architects Should Harvard Design Magazine Spring 2010 , .
October 2021) ed. Harvey M. Bernstein McGraw Hill 7 See Phillip G Bernstein Architecture .
,
'
2018 .
11 Hosey op cit , .
Violence at the Threshold of Detectability
Models Be Too Big?' in FacT'21 Virtual , ,
,
Canada 2021 .
Zone Books; The MIT Press Brooklyn, ,
,
3.5
9 This issue (and the example cited) is
1 George Barnett Johnston , explored at length in Brian Christian,
Assembling the Architect: The History The Alignment Problem: Machine Learning
and Theory of Professional Practice , and Human Values 1st edn, W.W. ,
Bloomsbury Visual Arts London 2020 , , , Norton & Company New York 2020 , , ,
York 1914
, .
2 Ibid ., p 119 .
Page numbers in bold indicate figures and tables. business models see value propositions
business risk 85 91 ,
Airbnb 153
Alberti, Leon Battista 4 95 117 118 120 121 , , , , , CAD see computer-aided design (CAD)
algorithmic systems 17 18 18 39 40 -
, ,
-
canonical economic model 73 74 81 151 -
, ,
,
-
, ,
Amara, Roy 33 34 -
American Institute of Architects (AIA) 84 104 110 137 , , , certification/registration 24 25 26 104 106 110 108 , , , ,
-
'Basic Services' 49 50 51 52 53 57 94 95 94 95 ,
-
,
-
, ,
-
,
-
Chaillou, Stanislas 106 107 ,
, , , , ,
Architects Registration Board (ARB) 24 25 26 85 104 , , , , compensation and value 72 81 155 162 160 171 -
,
-
, ,
, ,
architecture curricula 23 26 26 -
, ,
Aristotle 27 productivity 76 79 78 -
, , ,
-
, , , ,
capabilities of 19 20 19 47 -49 48 57 77 79 -
, , , , ,
-
127 149 150
,
-
, , ,
-
construction cost projection 143 145 143 146 147 -
, ,
-
, , , ,
,
-
146 147
,
future AI-based tools 110 111 112 113 117 118 119 -
,
-
,
-
history of 14 17 14 16 -
, , , ,
-
, ,
-
, counterfactual reasoning 96 97 ,
task automation xi 40 41 41 45 57 52 53 54 55 56 74 ,
-
, ,
-
,
-
,
-
, , , creative destruction 73 74 81 -
technology types 17 18 18 -
,
,
, ,
association 96 97 ,
interoperability 127 133 129 130 132 -
, , ,
,
, , , , , , , , ,
,
Deamer, Peggy 153
automation deep learning see empiricist/deep learning systems
demand for professionals 92 99 97 150 154 152 - -
, , ,
- , , ,
, ,
-
,
-
,
-
, , ,
-
,
Dennett, Daniel C. 98 99 ,
142 143,
design, generative 7 9 12 106 107 122 123 124 125 165, , , , , ,
-
, ,
, ,
, , ,
design labour 148 154 -
, ,
automated processes 39 40 -
, ,
design representation 9 10 11 12 , 36 37 , , , ,
,
, ,
efficiencies 74 76 75 76 139 -
construction cost projection 143 145 143 146 147 -
, ,
-
, , ,
, ,
value of data 79 81 80 -
responsibility and 85 86 ,
- -
taxonomy of use 9 10 ,
, , , ,
embodied carbon assessment tool 110 111 112 113 161 162 - -
, 18 36 , Marcus, Gary 17 18 19 20 -
, ,
capabilities of 20 47 49 48 57 77 79 ,
-
, , ,
-
market conditions 144 146 147 ,
-
task automation 48 51 52 53 54 55 56 57 56 -
,
-
,
-
,
-
, MEDLINE index 29
End User License Agreements (EULA) 86 88 -
,
-
modern slavery assessment tool 110 111 113 -
Murnane, Richard 46
expert systems 15 16 16 124 -
, ,
failures of execution 88 90 89 -
, , , , ,
, , , ,
, ,
, ,
,
, , ,
Picon, Antoine 34
Grenfell Tower fire 88 89 98 99 , ,
-
,
-
,
-
, ,
- precarious workers 153 154 -
heuristics x 18 23, ,
prediction 9 10 77 79 78 121
, ,
-
, ,
Higgin, Gurth 61 62 63 , ,
procedural task automation 50 51 52 53 54 55 56 57 -
,
-
,
-
,
-
, ,
,
process transformation 32 44 35
-
implicit knowledge 46 47 50 -
, , ,
, ,
integrative tasks 49 52 53 ,
-
productivity 76 79 78 -
, , , ,
-
, , ,
-
, ,
-
, ,
, ,
-
, ,
project delivery 58 69 59 -
design intent 61 64 62 63 -
, ,
Ladder of Causation 96 98 97 -
,
professional relationships 64 67 , 65 66 , 88 89 , 89 -
,
-
, ,
Lexus/Nexus system 29 29 ,
public policy 84 90 91 93 94 ,
-
,
-
machine capabilities 19 20 19 47 49 48 57 77 79 -
, ,
-
, , ,
-
realisation 8 9 10
MacLeamy Curve 139 139 158 , ,
, ,
,
Index
, , ,
-
, ,
-
IMAGE CREDITS
representation 8 9 10 11 , , , , 12 36 37 121 123
, , ,
-
, ,
RIBA 84 104 110 137 0.2: Loz CC BY-SA 2.0, via Wikimedia Commons
, , ,
Pycock, BY-SA2.0,
Plan of Work 49 50 51 52 53 57 61 94 95 94 95 ,
-
,
-
, , ,
-
,
-
1.1.1: René-Gabriel
Rene-Gabriel Ojeda. @RMN-Grand Palais
Ojéda. Palais/Art
/ Art
risks 85 88 90 89 91
,
-
, ,
Rittel, Horst x , 23 77 ,
Resource, NY
robotics 1.1.2, 1.4.9 & 2.1.1: Pelli Clarke Pelli Architects
1.1.2,1.4.9
construction automation 9 41 42 43 120 145 146 - -
Roitblat, Herbert 103 1.1.4, 1.4.3, 1.4.4, 1.4.5, 1.4.6, 1.4.7, 1.4.10, 1.5.6, 3.1.2 &
Rosenblatt, Frank 14 3.5.7: Autodesk Inc.
Rowe, Peter x 18 ,
1.2.1: Shunk-Kender ©J.
© J. Paul Getty Trust. Getty Research
Russell, Stuart 27 49 50 56 106 , , , ,
Institute, Los Angeles (2014.R.20)
Schank, Roger vi 96 , 1.2.2: Joseph Paul Cohen, Paul Bertin, and Vincent Frappier
1.2.2:Joseph
Schön, Donald 23 2019
Schumpeter, Joseph 73
1.3.3: Westlaw, By permission of Thomson Reuters
scopes of services 49 51 50 51 52 53 57 94 95 94 95
- - - - -
, , , , ,
scripting 7 11 12 41 43
,
-
, ,
1.3.4:
Building Ventures. 5G_wifi_©uniconlabs;
SHOP Architects 159 161 , AdvancedMaterials_advanced_©Wichai.wi; Al_artificial-
Al_artificial--
SmartVid.io 164 166 167 ,
-
intelligence_©Umeicon; Autonomy_stability_©Eucalyp;
Spacemaker 8 8 18 56 57 , , , ,
BIM_planning_©justicon; Blockchain_blockchain_©iconnut;
spanning strategies 159 160 , Com putationDesign_com puter_©xni mrodx;
ComputationDesign_computer_©xnimrodx;
specialisation, flexible 67 68 69 -
,
DigitalTwins_-
ComputerVision_vision_©xnimrodx; DigitalTwins_
Squires, Frederick 156 Drone_drone_©Freepik;
simulation_©Wichai.wi; Drone_drone_g)Freepik;
standard services contracts 84 137
structural engineering software 36 85 86 134
,
-
ElecCar_electric-car_©Freepik; GIS_address_©Cuputo;
, ,
IOT_internet-of-things_©photo3idea_studio; Mobile_
Mobile_-
supply chain conditions 144 smartphone_©Made by Made Premium; ModularPrefab_
ModularPrefab_-
supporting strategies 159 160 ,
RealityCapture_motion--
building_©Smashicons; RealityCapture_motion-
Susskind, Daniel viii xi 19 22 34 46 47 48 50 73 74 140 , , , , ,
-
, , , , , ,
151 154 ,
capture_©Freepik; Robotics_robotic-arm_©smalllikeart;
Susskind, Richard viii 22 46 , ,
TeX_lessee_©Eucalyp; VR_vr-glasses_©Pixel perfect; 3D_3d-
3D_3d--
systems performance information 38 38 ,
printing_©surang; BMS_system_©Freepik
1.4.1: Florian Schaffer,
Schäffer, CC BY-SA 4.0, via Wikimedia
Tally carbon assessment tool
111 112 113 161 162 ,
-
, ,
Commons
task automation xi 40 41 41 45 57 52 53 54 55 56 74 ,
-
, ,
-
,
-
,
-
, , ,
, ,
1.4.8: UpCodes
taxonomy of use 8 10 10 -
,
Li and I. Chen
1.4.11: E. Asadi, B. L.i
technologies and tools 2 12 -
,
- Harel/Judah Pearl/Dana Mackenzie
evolution of processes and 10 12 10 11 -
, , 2.3.3: Apicella
Apicelia + Bunton Architects
history of 3 8 3 5 7 8 -
, , , ,
2.4.2: S. Chaillou
taxonomy of use 8 10 10 -
see also artificial intelligence/machine learning; building 2.4.4 & 3.5.4: Kieran Timberlake Architects
information modelling (BIM); computer-aided design 2.4.5: FRDM.co
(CAD)
3.1.3: DPR Construction
Tombesi, Paolo 67 68 69 -
, ,
use information 38 38 ,
3.3.7: Aaron Flargreaves
Hargreaves / Foster + Partners
use, taxonomy of 8 10 10
-
,
3.4.1: © Ezra Stoller / Esto
value propositions 34 73 155 162 160 171 , ,
-
Witt, Andrew 36
Yale School of Architecture 24 26 99 100 101 158 159 160 , , ,
-
,
-