One of AI’s most promising uses is for robots
to replace humans in performing extremely
dangerous tasks: such as shimmying along
narrow cavities to replace damaged wiring or
record material stresses. Czech writer Karel
Čapek first coined the term robot in his 1921 play
R.U.R: Rossum’s Universal Robots – and today’s
humanoid versions, such as Boston Dynamics’
Atlas and Honda’s Asimo, are astonishingly
agile and sophisticated. Non-humanoid robots,
swivelling from fixed bases or traversing across
gantries, are already printing small dwellings
in masonry or powder-resins, or assembling
timberwork as adroitly as a master carpenter.
Artificial intelligence was ignored by
most built environment professionals until
satellite-enabled telecoms caused widespread
apprehensions during the 1990s, systemic
disruption during the 2000s, and now
inevitably, new ways of understanding and
doing things. Today AI brings another wave of
unfamiliar technologies and terms – including
augmented intelligence, where machines are
intended to improve human abilities to decide
and perform. This seems less threatening
than artificial intelligence, where machines
increasingly replace humans, to a tipping point
known as the Singularity (the term popularised
by Ray Kurzweil).
All intelligence, artificial or natural, flows
from competent processing of information.
Most AI researchers have abandoned their early
reliance on pre-programmed rules to solve
problems. Instead they are evolving machine
learning, where computers use algorithms tolearn how to better display or interpret lakes of
streaming data. The more data that computers
are fed, the more capably they seem to crush
complex tasks; partly through their supra-
human powers of pattern recognition.
Machine vision scientists depend on
open-source datasets comprising images of
objects that are classified and labelled to allow
comparisons with new images containing
similar objects. The world’s largest object
dataset, ImageNet, contains more than 14
million crowd-labelled thumbnails, which can
be downloaded to help identify, for example,
different types of natural places, buildings,
rooms, products such as fridges or dishwashers,
furniture, fabrics, clothes, and apparel such
as hats or sunglasses. Vision boffins classify
database images according to whether they
depict ‘things’ (box-frameable objects like
chairs, people or windows) or ‘stuff’ (matter
with no clear boundaries, like a patch of sky,
an office corridor, a wall, a hillside or a street).
Ironically, the image databases now being
assembled to support AI analytics all depend
on the ‘artificial, artificial [natural] intelligence’
of humans working online to label and
cross-check the images uploaded by database
compilers. One busy conduit is Amazon’s
Mechanical Turk (AMT) portal, which matches
employers (such as public research groups)
wanting freelancers to contribute to specific
human intelligence tasks (HITs). One recent
HIT, to assemble and correctly label 328,
thumbnail images of ‘common objects in context’
for the Microsoft COCO dataset, required 70,above Tianjin Library, Singapore.
Photography by Ossip van Duivenbode
beLoW Richard Buckminster Fuller’s 1928
vision of a ‘4D Air-Ocean World Town Plan’.ARchiTecTuRe & DeSign /PeoPLe/ juL-SeP 201919ADQ3_018_021_AI Architecture_V1.indd 19 26/7/19 4:49 pm