Design Portfolio / Caacyk: Data Treatment Works by Jeff Bray

DATA TREATMENT WORKS _ascetic retreat for data cleansing

jeffrey bray

class.architecturalPoetry

Materiality and structure of New Babylon Material Ecology to design holistic materials.

3D printing and growing material to expand existing architectures.

Sketch model of example 3D printed structure

Generative Design to create variety of organic structures.

3D printing, AI and robotics to automate and innovate construction solutions.

class.theGameOfLife

Urban fabric and sprawl of New Babylon AI could be used to assess the needs of the population and design the city according to use and density

This could be used as a model for the colonization of new planets.

Sketch model of urban sprawl as designed by AI

class.theLABYRINTH

Complex labyrinths of New Babylon that connect to create a never ending maze of spaces for one to wonder around

VR could be used to increase the perceived size of spaces utilising the principle of virtual gains and folding

Sketch model of virtual space within the labyrinths

class.neuralNetworks_GAN(face)

What is a Generative Adversarial Network (GAN) - A GAN is a neural network that outputs realistic simulated images based on an original true dataset. GANs are comprised of 2 competing networks to train the AI to be able to output random images what are effective simulations of the original. The generator’s purpose is to generate new data as close to the true dataset as possible to fool the discriminator. The discriminator’s purpose is to evaluate the input of either the true dataset or the fake simulation from the generator, and decide if it is true of false. Over many generations of feedback from the discriminator the generator learns how to trick the discriminator into thinking the fake data is part of the true dataset.

Top: Dataset of 30,000 images of real celebrities used to train GAN to produce images on new celebrities. The images below were created by the GAN, they are not real people. Above: GAN flow chart. Left: Process of GAN learning - resolution over epochs.

class.AIartwork_robbieBarrat

c:\> nudes Barrat trained a neural network from 20,000 images of classic nude paintings to create new nude portraits from AI. The AI failed to learn all of the proper attributes found in the original images and instead it generates surreal blobs of flesh with strange faces and tentacle like arms. “Is this how machines see people?” Barrat_

class.AIartwork_nudeModel

The original dataset was curated by Robbie Barrat and an image was generated by the AI_ The image is then reinterpreted into 3d space by the designer with additional creative input i.e.. deciding the third dimension and extending the figure outside the frame of the image_

class.AIartwork_sofiaCrespo c:\> neuralZoo In this work neural networks are the creator of new biological creatures using a dataset designed by the Crespo. These works show how the AI has reinterpreted the data in a frankenstienian manner with yet no recognisable or distinguishable separate pieces. “Our visual cortex recognizes the textures, but the brain is simultaneously aware that those elements don’t belong to any arrangement of reality that it has access to.” Crespo_

root:\> How can we continue to inspire machines to create art for us as emotional human beings? Crespo_

class.GAN_no.AI.manual‘GAN’

London

}

Belgium

Romania

Spain

human curated height map

Working with a dataset of doors from different European countries I created a collate in an attempt to replicate the output of Sofia Crespo’s GAN, A new frankensteinian door_ The human version of the GAN has a different aesthetic, it is very time intensive and potentially not possible in the way human brains are wired to consider the entire dataset in the creation of one image; as such I chose 3 images from each subset and this is a small dataset of just 120 images. AI is far better at reinterpreting all images of a dataset and can work with 1000s of images_ I then created a greyscale height map as input into an algorithm that interprets 2d information into 3d form to explore how the algorithm ‘sees’ the image_

algorithmic 3d interpretation

human collage

Portugal

class.GAN_no.AI’GAN’model

rendered model

class.GAN_facade(label2pix) Training Dataset of real images

Human Made Label Input

GAN AI Output

Designer Made Label Input

Novel GAN AI Output

class.GAN_floorplan(architecture) Stanislas Chaillou of Harvard University developed this methodology and workflow which uses GANS trained on existing layouts to generate multiple layouts from which the architect can choose. The architect can also interact with the network via the very basic labelling interface to create layouts in a similar fashion as the label2pix GAN detailed in the previous page. Chaillou also trained the network on a variety of historic and contemporary plan styles such as the example shown below of baroque. I find this is an interesting use of the technology however this is very constrained way to apply the rules and constraints of the past into current designs, I imagine this could be used to quickly produce varying layouts for large developments. However this powerful technology should be used to push the field of architecture further into the future. I suggest that this technology can be used to push the aesthetic and style of architecture in a new direction rather than simply using it to replicate the past.

GAN pipeline using architect as selector and AI as generator

GAN trained on Baroque architectural plans takes a contemporary layout and creates a baroque interpretation of it.

class.GAN_3d(from2d) In 2020 Google AI challenged AI developers to use the public 2d data on google earth/maps and google images which are added to by users everyday to develop 3d models of famous landmarks using neural networks, machine learning in a process called structure_from_motion This process recognises features across multiple 2d images and uses triangulation to extrapolate the 3d location of all pixels in the image. This method is already used by google to create the 3d models used for google earth/maps however this is based on carefully curated images taken by high-quality cameras in a specific way by the street view mobile camera units and satellite imagery. The challenging aspect of this task is to train the network to ignore transient feature such as crowds and to make allowance for differing image quality, light levels etc. With neural networks that can take the vast resource of digital photography freely available on the internet it may be possible to categorise and model the global architecture scene in 3d.

Visualisation of network triangulating points in latent space across multiple 2d images of St Paul’s Cathedral from different angles

2d images are collected from google images and google maps users of Trevi Fountain to form a dataset of thousands of images from different angles

reconstructed 3d model of Trevi fountain extrapolated by the network from crowd sourced 2 dimensional images

class.GAN_3d(from3d) Robbie Barrat has developed a GAN that can process a 3d dataset, there are a lot of limitations however this is an exciting branch of research and the field is progressing quickly. I believe the difficulties lie in the amount of processing power that is required to process this amount of information and the limitation of the availability of good quality training data and existing 3d models. However with the advancements in 3d scanning and photogrammetry as the above, there may be in the near future the appropriate data to advance this field of theory. Neural networks that can understand, process and generate 3d digital objects will be a game changing advancement for architecture research. In the below example Robbie Barrat trains a neural network using model database Thingi10K, part of the limitation of the GAN is that it requires the models are converted into 32x32x32 voxelations in order to be processed by the network, which is not very useful. This GAN can output 2d PNG visualisations of the generated models and a manifold OBJ that is able to be 3d printed straight out of the network. If this was able to process much higher resolutions and models such as matterport3d scans of building exterior and interiors, and therefore able to output finished 3d printable architectures, would the role of the architect be similar to the role played by Sofia Crespo as an artist in collaboration with AI neural networks?

3d model from Thingi10K 3d model database

Database contains 10,000 high quality but small OBJ models

2d PNG visualisations output by the network of generated 3d models

3d printed 323 voxel model generated by the network trained on the Thingi10k dataset

Julian Assange // Wikile

aks

def research.ai(massSurveillance):

Public data and AI sold to military

y Five Eyes

Edward Snowden whistle blower on mass surveillance b

Google owns the vast majority of data on the population due to its 5.6 billion searches per day. Google develops AI and sells our data without permission.

In 2003 Edward Snowden with the help of Wikileaks (an organisation that releases leaked classified information) lifted the lid on the use of datasets collected from the public domain and utilised by governments all over the world without consent.

FBI NSA

CIA

Data // game (Watch Dogs, 2014)

Watch Dogs is a game where the main character can access the governments AI system to find information about people via facial recognition. These databases are being used by AI owned by authorities around the world.

MI6 MI5

gnition

n, 2019)

collecting information

vent AI recognition

Masks to cover face to pre

protesters against use of AI use face masks, gas masks, ski masks and umbrellas to avoid their faces being recognised by AI cameras

within the public domain

Cameras are everywhere

Paglen used facebook profile images to understand how the AI recognises people, this image is the proto-face of the profile owner. This is the users average face minus everything common among other users.

AI Faces GAN // neural network (nVidia, 2019)

Facial Recognition // movie (Anon, 2018)

nfuse AI ortrait from training

AI output from training

her genetic li

My 3D model interpretati

Artist Robbie Barrat trained a GAN on a dataset of classic nude portraits, this shows the new generated images of the GAN and my 3D model of this image. I will be collaborating with AI throughout my design processes.

Crossbreed of 2 genetic families

on of AI image

ry to prevent AI recogniti

Mixed with anot

neage

Ewa Nowak // Facial Jewell

There is a movement of people obscuring their faces from AI facial recognition to prevent their real world actions being recorded to their digital profile.

AI understanding of my face

Example classic nude p

The AI engine behind these technologies are Generative Adversarial Networks (GANs) These are able to learn from images and make correlations between datasets and patterns in images. They are able to generate images of faces like the one to the left that are not real. The GAN learns the underlying patterns in images and will learn the structural information that defines facial features, colour schemes, poses and whole faces. These are stored in ‘neurons’ aka layers of algorithms. These are called ‘genes’ and can be altered independently and mixed together.

dataset

CV Dazzle // makeup to co

n with facial reco

Hong Kong Riots // 2019 artist (Trevor Pagle

Invisible Images //

British police va

import artbreeder def process.futureRep(aiGeneticDesign): mixGenes(gene1,gene2,gene3,gene4)

gene1.child3

gene1.child4

gene2.child2

gene1.child5

gene2.child1

gene1.GASMASK

gene1.child2

gene2.SKIMASK

combine.mixGenes(1,2,3,4)

gene1.child1

gene2.child3

gene2.child4

gene2.child5

gene3.child1

gene4.child5

gene4.FACEMASK

gene3.UMBRELLA gene3.child2

gene4.child5 gene3.child5

gene3.child3

gene3.child4

gene4.child1

gene4.child2

gene4.child3

Using the data scraping python script below 795 images were downloaded autonomously from Google images from the search term “monastery”.

Artificial Intelligence_Data_Scraping Below is the uncurated dataset; note that there are inappropriate images, images are of differing sizes & some images are completely irrelevant.

from selenium import webdriver from selenium.webdriver.common.keys import Keys from time import sleep import requests import io import os import PIL import hashlib from PIL import Image driver = webdriver.Chrome('E://dataScraping/chromedriver') def fetch_image_urls(query:str, max_links_to_fetch:int, wd:webdriver, sleep_between_interactions:int=1): def scroll_to_end(wd): wd.execute_script("window.scrollTo(0, document.body.scrollHeight);") sleep(sleep_between_interactions) # build the google query search_url = "https://www.google.com/search?safe=off&site=&tbm=isch&source=hp&q={q}&oq={q}&gs_l=img" # load the page wd.get(search_url.format(q=query)) image_urls = set() image_count = 0 results_start = 0 while image_count < max_links_to_fetch: scroll_to_end(wd) # get all image thumbnail results thumbnail_results = wd.find_elements_by_css_selector("img.Q4LuWd") number_results = len(thumbnail_results) print(f"Found: {number_results} search results. Extracting links from {results_start}:{number_results}") for img in thumbnail_results[results_start:number_results]: # try to click every thumbnail such that we can get the real image behind it try: img.click() sleep(sleep_between_interactions) except Exception: continue # extract image urls actual_images = wd.find_elements_by_css_selector('img.n3VNCb') for actual_image in actual_images: if actual_image.get_attribute('src') and 'http' in actual_image.get_attribute('src'): image_urls.add(actual_image.get_attribute('src')) image_count = len(image_urls) if len(image_urls) >= max_links_to_fetch: print(f"Found: {len(image_urls)} image links, done!") break # move the result startpoint further down results_start = len(image_urls) return image_urls

def persist_image(folder_path:str,url:str): try: image_content = requests.get(url).content except Exception as e: print(f"ERROR - Could not download {url} - {e}") try:

image_file = io.BytesIO(image_content) image = Image.open(image_file).convert('RGB') file_path = os.path.join(folder_path,hashlib.sha1(image_content).hexdigest()[:10] + '.jpg') with open(file_path, 'wb') as f: image.save(f, "JPEG", quality=85) print(f"SUCCESS - saved {url} - as {file_path}") except Exception as e: print(f"ERROR - Could not save {url} - {e}") def snd(search_term:str,number_images=5): target_folder = os.path.join('E://dataScraping/images','_'.join(search_term.lower().split(' '))) if not os.path.exists(target_folder): os.makedirs(target_folder) with driver as wd: res = fetch_image_urls(search_term, number_images, wd=driver, sleep_between_interactions=0.5) for elem in res: persist_image(target_folder,elem) snd("monastery", 700)

To train the AI ‘gene’ to produce images of monasteries the dataset is curated by the designer to include specific images that define the ‘gene’ Images are resized and cropped to the same size and ratio 250px 1:1 in this case to make the training more stable and then sorted into discarded and inappropriate images that may be used Artificial Intelligence_Curating_Datasets for a related ‘gene’. In this case only 32.1% of the image are chosen by the designer to form the training dataset. In this example that is only 255 images, in reality its takes millions of images to train a robust ‘gene’. At this yield a 1 million image training dataset would require the sorting of 3.1 million uncurated images. at 5 seconds per image that is approx 4,000 ‘man hours’ or 500 working ‘man days’_

Discarded Images (64.3% of uncurated dataset)

Images that could be used to train a gene for monastery interiors (3.6% of uncurated dataset)

Images to form the training dataset for the monastery gene (32.1% of uncurated dataset)

Images created by AI trained using the curated dataset

When outputting images from genetic algorithms the process begins with inputting an image of randomly generated noise. This prevents the network always outputting the same image, it adds chaos into the system. The randomness of the input noise can be controlled by using a floating value referred to as chaos. The effects of this value are demonstrated below.

class.Design_evolution(noise_Chaos) 1

0.9

0.8

0.7

chaos

0.6

0.5

0.4

0.3

0.2

0.1

Chaos = 0.92

Chaos = 0.83

Chaos = 0.69

These ‘genes’ are then used in a genetic design process that evolves with the designer making selections and alterations to the genetic make up. In Darwinian terms this process could be thought of as survival of the Artificial Intelligence_Design_Process designers aesthetic vision.

class.architecture_geneticDesign1 Genetic Make Up

Human in the Loop Design Evolution

weighted gene

Reference image

The designer evolves their aesthetic vision with the AI and then designs with this in mind, using human skills and knowledge along with bespoke inspiration developed harmoniously with AI.

Artificial Intelligence_Aesthetic Vision

class.abattoir_model(render)

class.architecture_geneticDesign2 Genetic Make Up

Human in the Loop Design Evolution

weighted gene

Genetic Make Up

weighted gene

class.aquarium_model(render)

class.architecture_geneticDesign3 Genetic Make Up

Human in the Loop Design Evolution

weighted gene

Genetic Make Up

weighted gene

class.arboretum_model(render)

def research.ai(dataHacking): "This is a small, justfor-kicks release of some internal data from senate. gov – is this an act of war, gentlemen? Problem?"

Lulz Security // Hacktivist

fi novel / movie (George O

group (2011)

Nineteen Eighty Four // sci

“We mean you no harm and only want to help you fix your tech issues.”

Hacking // game (Watch Dogs, 2014)

Cambridge analytica used data from individual’s facebook accounts to train AI and target personalised political messages directly to individuals browsers. Personal data was used without consent of the individuals. Our data is being used to control our thoughts. This is reminiscent of the totalitarian propaganda of the novel 1984 by George Orwell (1949).

rwell, 1949) Christopher Wylie data consultant

Corrupt Tunisian Government

"The Internet is the last frontier and we will not let corrupt governments spoil it. We are Anonymous, We are LulzSec, We are People from around the world who are stepping in the name of freedom" Anonymous 2011

AI Propaganda

AI Hacking Population // advertising

up (2004-)

Anonymous // Hacktivist Gro

Brexit campaign

Trump campaign

(Brexit & Trump campaigns, 2016)

Westboro

"Your downfall is underway. Since your one-dimensional thought protocol will conform not to any modern logic, we will not debate, argue, or attempt to reason with you." Anonymous 2012

baptist church Data Hack // glitch art (miri kat, 2019)

Databending // 3D .OBJ data

What if there was a new hacktivist group that could use AI systems to fight government AI systems by corrupting the dataset from which the AI works. If you could hack your own digital dataset and corrupt the AI’s understanding of you. You could disconnect your physical self from your digital self.

hacked in text editor

3D Glitch // pixel sort

rio Klingemann, 2018)

Neural Glitch // AI GAN artist (Ma

itch

3D Glitch // shift & RGB gl

acked in text editor

Databending // image data h

In this artwork Klingemann corrupts the dataset used to teach an AI GAN neural network about faces.

def research.ai(Discrimination):

Person of Interest is a sci fi TV series where detectives use an AI system to predict where and when the next crime will take place. A real system is used in the US called PredPol that does exactly this.

AI Persecution // movie

(Person of Interest, 2011-2016)

AI Persecution // tv series

(Minority Report, 2002)

PredPol AI

Minority Report is a sci fi movie where detectives use a system to predict who will commit the next crime and catch the criminal before the crime is committed. This is frighteningly similar to system used in the real world such as COMPAS in the US.

COMPAS AI Discrimination

Discrimination

PredPol learns from police reports to predict when and where crimes will happen rather than unbiased statistics. If a minority area is targeted due to human bias more reports are filed and a feedback loop is created.

The Digital Poorhouse

COMPAS AI system used in the US to assist in sentencing based on risk of re-offence based on facial features.

nition - gender dis

AI Facial Mis-recog Facial recognition software works very well for white males. But it doesn’t ignore other genders and races it can misrecognises them as white males.

AI systems connected to the biometric system in India did not recognise the finger print of Motka Manjhi as a result his rations were stopped and he starved to death.

Un-curated data sets

crimination ...our gender

Governments are not curating datasets carefully to ensure equality and unbiased AI systems they do not even check to test accuracy. However I do worry about the power they will harness through AI when they learn how to control AI.

AI Oppression becau se of...

...our beliefs AI systems are being used to categorise us and segregate people into sub-datasets to be treated differently, it does this by assessing our faces and linking our digital identity but also by finding arbitrary patterns between something like crime rates and pictures of faces. What would the AI see if we all looked the same.

look

...the way we

Memo Aktin’s Dirty Data research with AI neural networks (GANs) used an algorithm to skim images using the search term ‘Donald Trump’ on Google Images and then trained the AI without curating the dataset at all. The output is less than accurate. This is an allegory of the uncurated data used by governments to make ethical and moral decisions.

looked the same (not like

John Malkovich)

...What if we all

class.ai_dataSet( )

class.aiDataset_training(clickworkers)

Clickworkers perform mundane tasks to label data such as this to create datasets for researchers to teach AI networks how to see and process images as a human

Chihuahua or muffin - easy for humans very hard for algorithms

Google re-captcha has us training AI for free

Mechanical Turk Chess playing Machine

Data labelling factories in 3rd world locations where labour is cheap.

Crowdworking using platforms such as amazon mechanical turk pay poorly for menial tasks

Crowdworking sweatshops are distributed worldwide

class.site_hashimaIsland(History) Forced Labour - Hashima Island 1930s Coal mining

Hashima Island coal mining island and city

Digital Sweatshops - Earth 2020 Data curation

class.typology_asceticRetreat In Evil Paradises: Dreamworlds of Neoliberalism Mike Davis attributes the sudden rise in ascetic retreats in the western world to neoliberal values. He describes in chapter 17, titled Monastery Chic, that “Contemporary Americans... are attracted by the very incompatibility of current and monastic values: the monks’ life of duty, ritual, rules, and selfdenial contrasts with the perceived “emptiness” of modern life, devoid of tradition, structure, and obligation, and mired in meaningless consumption” The monastic retreat provides a sort of self deprivation and an escape from the materialistic, fast paced and tech connected world we live in, however it seems the opulence and materialism is still present in most monastic retreats that offer decidedly unascetic luxuries such as swimming pools, the only thing on offer is a break from technology and fast paced life. This is not as much of a return to traditional monastic asceticism as we would like to believe. I propose a typology where the truly repentant criminals of AI data crimes against humanity might go to devote themselves to a life of true asceticism and penance. A monastic life of data humanisation where they focus on helping to fight the war on corrupt AI training datasets and the resulting oppression of the human race.

Neoliberal indulgence, luxury, transient, distraction, isolation, indulgence, materialism & gratification

Ascetic tourism, transient, isolation, indulgence, distraction, luxury

Monastic asceticism, ritual, community, abstinence, faith, focus

class.site_hashimaIsland(location)

Section through island and mine layout

Plan of Hashima Island

Short sections of Hashima Island

class.site_hashimaIsland(Abandoned)

class.narrative_dataNomads

class.narrative_dataHackers

class.narrative_dataMonks

class.neuralNet_architecture(mineDesign)

The visualisation of the architecture of neural networks will be used to define the physical structure of the mines at a human scale, where each of the layer nodes or neurons will be the location of holoPlatforms where the humanisation of the data will occur via human interaction with holographic visualisation of the data. The links between the layers and neurons will form the physical infrastructure of the network which will be represented by the aesthetic of biological neurons. The mine will be set up in a way that the layers and holoPlatforms can be used in any setup up to accommodate the many different types of artificial neural network, multiple operatives can be set up on each holo_platform to increase the amount of neurons per layer.

This dataset visualisation while not a neural network offers a visualisation that represents how the data mine will spread beneath Caacyk like a root system.

Data Mine_Physical and Neural Network Diagram

//key Pearlescent materials represent physical architecture pink strands represent diagrammatic AI processed data flow green strands represent diagrammatic human processed data flow _

data curation process: incoming / outgoing global network data

Caacyk data_centre requests data from the AI data is collected from Caacyk residents and from global network [internet] data is processed by AI in a virtual network supported by neural computing and the beginings of a dataset is created data then sent down main nerve into first layer of the neural_mine network and connects to a holo_platform where the dataset is visualised as a hologram Data is sorted by humans interacting with the data holograms and sorted data is sent to holoPlatforms on the next layer of the network along with more AI data from the main nerve

data collected from architecture and humans in Caacyk

horizontal tunnels at each layer of the network physically connect holo_ stations in each layer

this repeats through the network until the final holo_platform where the data is considered humanised and is sent up the secondary nerve wrapped around the elevator to the Caacyk data-centre _

holo_platforms display data visualisations as AR holograms for human interaction in curation rituals

location of mine render _ shows AI data nerve, elevator and holo_ platforms vertical data circulation via central ‘nerve’ incoming/outgoing AI curated data to holo_platforms for human interaction throughout the neural network.

vertical human circulation via elevator takes operatives to different levels of the physical infrastructure network and also carries Humanised data back to Caacky via data nerve wrapped around elevator.

class.neurons_network(designInfluence) Artificial neural network architecture (B based, at least theoretically, on neurons natural brain (A & C). As such the design network infrastructure in the neural_mine based on biological neurons and nerves.

& D) is of the of the will be

To understand the aesthetic of these biological items I found an interactive model of a fly brain (largest complete biological neuron map currently modelled) which was developed with the help of artificial neural networks to trace and label individual neural networks through nano slices through the entire brain. Below is one of the slices of the brain with the olfactory neural network highlighted as the neurons pass through, below that is the 3d model of the complete olfactory network.

Biological & artificial neuron similarities

Fly brain MRI slice visualised in Neuroglancer

Fly brain Olfactory neuron system visualised in Neuroglancer

class.data_visualisation(holoPlatforms)

Data visualisation of AI data curation (note free nodes)

Example of AI curated data set

Example of same dataset after human has refined and organised

Holographic GUI for physical human data curation

WAKE UP AND WASH - 4AM

DAILY_SCHEDULE

1_ VR CONTEMPLATION WITHIN CHAMBERS - 5AM

2_ WALK TO DATA_CENTRE - 6AM

7 3_ WORK IN ARCHIVES IN DATA_CENTRE - 7AM

4_ HUMANISE DATA IN NEURAL_MINE - 10AM

5_ DAILY COMMUNAL SUPPER - 2PM

6_ PHYSICAL CURATING TRAINING - 4PM

1 7_ MAN DATA_MAST SECURITY 7PM

SOCIALISE AND SLEEP - 10PM

NEURAL_MINE_SECTION In a more sustainable future AI will be utilised to make the best use of existing topography and geology. The existing mines stretch to 600m below sea level with a main vertical shaft from the island surface. It is from this starting point the AI and swarm robots begin, with the data centre on the surface following the existing vertical shafts to house vertical circulation of the operatives as well as the data infrastructure. horizontal circulation tunnels are burrowed to the outer rings that form the levels of the neural mine the excavated material is repurposed and 3d printed to create the retaining structure. Excavations follow the grain of the softer rock and target natural caverns for holo_station locations to reduce energy usage.

1 Vertical data infrastructure - Organically reinforced fibre data cable with very high bandwidth to carry data from data centre to holo_stations.

2 3 4

LAYER_01

2 Vertical operative circulation to carry operatives to layers of the lift - this is also organically reinforced and is wrapped by data fibre cables carrying humanised data from last layer in the mine to the data_centre. 3 Outer ring in each layer provides horizontal circulation between holo_stations on the same layer of the network. 4 holo_station - this is where the operatives curate datasets collated by AI at the data centre.

2 4 5 horizontal circulated from core vertical shafts to outer ring

LAYER_02

6 data_centre - where data is initially processed by AI, stored in archives, and strategically reimplemented into the global network 7 data_mast - where data is collected from and sent to the global network.

LAYER_03

LAYER_04

NEURAL_MINE_DATAFLOW 1 data is collected from the vr processing machines in the residential chambers of the operatives and from the global network via the data mast. This data is then collated into dataset by AI to be sent into the mine for humanisation. 2 data is sent into the neural_mines via the main AI data nerve and distributed to the holo_stations in each level of the mine. 3 within the holo_stations the data is separated into sub-datasets and distributed to the holo_platforms for human curation. The humanised data is then distributed to the next level of the neural_mine and combined with new dirty AI data to be refined further by the operatives in that level.

4 once the data has been filtered through all the levels of the neural_mine it is sent up the return data nerve that is wrapped around the lift. It is taken back to the data_centre to be archived and redistributed to the global network via the data_ mast.

4 dirty data humanised data

AI OPTIMISED 3D PRINTED WEATHER FACADE EXISTING RUINS FOR RESIDENTIAL CHAMBERS The island is a place for people from the world of corrupt AI systems to come and focus on helping in the endless task of cleaning and humanising dataset to be used for AI systems as a data operative. As such the focus is not on luxury but on functionality and asceticism. The living quarters of the operatives are housed within the existing ruins, autonomous swarm robotics is utilised along with AI structural optimisation to build a new structure to encapsulate the entire building in a weather tight skin that is designed to take advantage of passive heating and cooling in both its physical design and materiality.

Swarms of autonomous robots guided by AI systems 3D print facade structure in a similar way to the exoskeleton of the new buildings. This is designed to provide structure to the facade system and is self supporting by seating itself on the topography at the base of the frame. It is braced by connections to the existing building.

The glazing is printed with the structure of a composite silicon material designed with properties for passive heating and cooling such a to reduce solar gain during the summer when the sun is high and to increase solar gain in the winter when the sun it low. The glazing also allows the space within to breathe with vents controlled by AI systems. The glazing panels are angled downwards to allow natural shading during the summer.

The existing ruins of the island are made of strong concrete frames, however they are not weatherproof. As the island is about reducing energy use the residential chambers for the operatives will be housed within the existing ruins that are suitable. A weatherproofing housing will be built to encapsulate the whole building.

class.render_themes(colourScheme) Reference image

VR CONTEMPLATION

DATA CENTER

DATA ARCHIVES

NEURAL MINES

FOOD COURT

HOLO TRAINING

DATA MAST

Colour scheme replacement image if original scheme undesirable

extracted colour scheme

Colourised reference image

class.Design_evolution(VR_contemplation) Evolution and Crossbreeding a

Genetic input

b e

AI Reference Image

Chaos = 0.73

PERSONAL_DATA_CENTRE 19

The operatives have a VR data machine in the chambers, this is used to collect human data from the operative so that AI can learn to understand the human condition through short programmes run in VR so the AI can test the human response, these programmes are designed and tested using data from the data centre archives.

The machines process large amounts of data using organic neural processing units. 1 Personal data mast collects data sent from the main data_mast to the chambers of the operatives and sends data back to the main data_mast and back to the data_ centre.

2 Lid to the preservation chamber. 3 Preservation chamber glass unit supported by 3d printed structure. 4 Structure to hold the organic material from the mast.

5 Part one of the 3D printed, protective Faraday cage to protect sensitive organic processing unit from electromagnetic interference.

6 Part two of the protective Faraday cage. 7 Organic neural processing unit processed initial data from the mast.

8 VR machine base - receives data from no.7 and from the feedback loop from the VR headset, 10, 11 and 12 back to the base and out to the VR machine. 9 VR headset receives initial programme and is part of a feedback loop that feeds human feedback into the Machine, the AI learns and alters the programme and sends back to the VR headset.

10 Pre GAN organic processing unit to process human response and feedback from the VR programme. 11 Sub GAN layer 1 to process and learn from data gained from VR programme. 12 Sub GAN network layer 2. 13 Pre GAN processing unit Faraday cage.

14 pre GAN processing unit frame and glass preservation chamber.

15 Lid to preservation chamber with power input socket. 16 Preservation tube for Sub GAN layer 1. 17 Preservation tube for Sub GAN layer 2. 18 Power distribution node. 19 Power input cable.

6 12

8 physical connection power flow data flow

class.Design_evolution(data_centre) Evolution and Crossbreeding

Genetic input

designer curated datasets create artificial aesthetic genes for mixing using algorithms

AI Reference Image Chaos = 0.81

AI OPTIMISED 3D PRINTED EXOSKELETON - NEW ARCHITECTURE

Swarms of autonomous robots guided by AI systems 3D print building exoskeleton structure from composite material devised from local surplus material gathered from the excavation of the mines. The AI system optimises the material required by applying more materials where the stresses are greatest in the structure.

Zaha Hadid Architects exoskeleton hotel

The exoskeleton is also printed to house the glazing, the material for which is also sourced locally from the silicone in the excavated rock. The structure is holistic and the glazing contributes to the structural integrity of the exoskeleton. Materials are printed at a molecular level to tailor material attributes

Iris van Herpen’s 3D printed dresses made from composite polymer

The floors are made of an optimised honeycomb lattice that is structural, insulative and lightweight.

Neri Oxman’s material ecology grows bio-materials on 3D printed lattice

Fibre optic data nerves carry data to and from the neural mines and branch from the top down as the dirty data is refined into datasets and from bottom to top as humanised data is organised and disseminated/archived. The data nerves are supported by organic bio-materials that are grown along 3D printed matrix.

Neri Oxman 3d prints columns with swarm fibre robotics

class.Design_evolution(data_archives) Evolution and Crossbreeding

Genetic input

AI Reference Image

Chaos = 0.68

DATA_CENTRE_ARCHIVES The data centre archives are where the pre-processing and post-processing occurs for the neural_mines datasets. Operatives work here with AI systems to organise and disseminate the data. The data_centre is built from the ground up and is built around the data nerves that serve the neural_ mines. 1 External building exoskeleton 2 Internal 3d printed structure in concentric circles is built around the data_nerve as it is built/grown. These layers support the edges of the floor plates that are cut to allow the data_nerves to pass through the building. As the internal structure is printed as a single unit it grows from the ground floor plate, providing a stable foundation.

3 These 6 concentric layers act as a faraday cage protecting the data from external influences such as electromagnetic fields generated by the large amount of servers in the data_archives.

4 Organic bio-material is grown on a 3d printed matrix and provides flexible support to the data_ nerves. 5 A transparent polymer supports the growth of the organic nerve support and prevents the growth damaging the fibre nerves within. 6 Firbe optic cables carry large amounts of data to and from the mines, and to the archive holo_ platforms.

9 6

7 Small holo_casters, similar to those in the neural_ mine holo_stations, receive incoming data and project holographic visualisations to be organised and disseminated to the relevant archives or to specific uses in the global network by the operatives working in the archives. 8 Holo_platforms receive the sorted data information and carry this into the archives in the level below. 9 Data_servers hold all the data to train AI algorithms and the resulting algorithms. Information stored as AI algorithms takes up vastly less space as they build a mathematical understanding of the patterns within the data and store it in a trained code. For example a database of 2.56TB images of faces can be stored as an AI algorithm of just 1MB that can generate images of faces based on the original dataset is just 50% of the levels in the data_centre are devoted to servers.

class.Design_evolution(neural_mines) Evolution and Crossbreeding c

Genetic input

AI Reference Image

Chaos = 0.86

Holo_station.Floorplan The holo_platforms receive AI curated datasets “dirty data” through the network via the data centre. Humans act as a humanising component in the data circuit physically bridging the flow of data between the holo_casters and the holo_platform. Without the human the data cannot physically pass any further through the network, this way no “dirty” data is output from the mines, only humanised data.

01. Raw (uncurated) data collected by the data_mast from the global internet and sorted into datasets by AI at the data_centre. These “dirty datasets” are sent down the mine along a data_nerve delivering pre-humanised datasets to multiple holo_stations at each level of the neural mine network.

02. Data is carried from the data nerve to the holo_caster units above each platform in the holo-station via fibre optic cables. 03. Fibre cables plug into the holo_casters above the platforms, the data is processed by the holo_casters into 3D visualisations so that humans can interface with and understand the data relationships in the complex AI datasets.

04. The holo_casters project weighted holograms to provide a tangible experience for the human data organisers, creating a tactile rather than digital experience. This provides the organisers with a more engaging and healthy task while sorting and humanising the datasets.

05. Humanised where the to refine delivered

data is sent to holo_stations in the next level of the network humanised data is processed together with dirty data by the AI, and gradually humanise the AI algorithms, before it is then to the next human in the loop.

06. Once the data has been through every level of the mine and is humanised it is then sent up to the data_centre archives and redistribution hub to be carefully reintroduced to the global internet for worldwide use. 07. Vertical circulation carries data organisers to each level of the neural mine network. 08. Stairs provide access to each holo_platform in the holo_station.

09. Sea water is used to cool the processing units and is stored around the holo_station platforms, this is continually circulating and cooled via large vertical waterfall wind tunnels.

10. 3D printed concrete retaining wall structures provide support to caverns carved in the mines that house the holo-stations. The designs are structurally optimised by AI and are bespoke for the specific loads acting within each space. 11. Autonomous swarm robots are sent into the mines to carve the caverns and platforms from the rock and build the support structures.

18 17

class.Design_evolution(food_court) Evolution and Crossbreeding

Genetic input

AI Reference Image

Chaos = 0.65

class.Design_evolution(holo_training) Genetic input

Evolution and Crossbreeding c

f e

AI Reference Image Chaos = 0.83

class.Design_evolution(data_mast) Evolution and Crossbreeding

Genetic input

AI Reference Image

Chaos = 0.75