1
CONTINUOUS
INTELLIGENCE
David Colls
ThoughtWorks Live 2018
2
ARTIFICIAL
INTELLIGENCE
AMAZING OPPORTUNITY
There is almost nothing we can think of that
cannot be made new, different, or more valuable
by infusing it with some extra IQ.
The business plans of the next 10,000 start-ups
are easy to forecast: take X and add AI.
Kevin Kelly – The Inevitable
“
“
AND UNPRECEDENTED RESOURCES
2
DEDICATED
PARALLEL
HARDWARE
3
ADVANCED
MODELLING
1
WEB-SCALE
DATA
DATA VOLUMES
DOUBLE EVERY YEAR
MASSIVE ADOPTION OF GPU
(AND TPU, HPU, FPGA)
LATEST RESEARCH
AND TOOLS
NO WONDER AI GENERATES
INTEREST FROM LEADERS
But the big change from last
year is that 81% cited
artificial intelligence and
machine learning as either
very important or extremely
important to their company’s
future, up from just 54% in
2016.
Fortune.com Fortune 500 CEO survey 2017
81% “
“
There is almost nothing we can think of that
cannot be made new, different, or more valuable
by infusing it with some extra IQ.
Kevin Kelly – The Inevitable
“ “
THE HYPE PROBLEM
So how do we think about opportunities?
AI CAN DO ANYTHING
WE WANT (IT’S MAGIC)
There is almost nothing we can think of that
cannot be made new, different, or more valuable
by infusing it with some extra IQ.
Kevin Kelly – The Inevitable
“
AND THE KERNEL OF THE TRUTH
“
AND ABOUT
PEOPLE
AND WE CAN
BUILD
ITERATIVELY
AI IS NARROW
(BUT COMPOSABLE)
There is almost nothing we can think of that
cannot be made new, different, or more valuable
by infusing it with some extra IQ.
Kevin Kelly – The Inevitable
There is almost nothing we can think of that
cannot be made new, different, or more valuable
by infusing it with some extra IQ.
Kevin Kelly – The Inevitable
There is almost nothing we can think of that
cannot be made new, different, or more valuable
by infusing it with some extra IQ.
Kevin Kelly – The Inevitable
8
INTELLIGENT
EMPOWERMENT
AUGMENTATION OVER AUTOMATION
AUGMENTATION OVER AUTOMATION
MACHINES
Wider learners
Scalable thinkers
HUMANS
Faster learners
Flexible thinkers
+
THE INTELLIGENT EMPOWERMENT CHALLENGE
High volume 

judgement tasks are
high cost 

or variable quality
High quality human
judgement is hard to
deliver at speed or
scale
Empower all your
people with the
judgement of your
best performers
Leadership
Lenses
BUSINESS
CUSTOMERS
EMPLOYEES
Outperform with insight
Experience driven by AI
Your people at their best
12
COMPOSABLE
INTELLIGENCE
PROBLEM EXAMPLE
PREDICT A NUMBER
● How much will this house sell for?
● How many shoes will sell this week, based on recent sales?
BINARY PREDICTION
● Will this customer churn this week?
● Is this transaction fraudulent?
PREDICT A CATEGORY
● Is this customer NEW, LOYAL, FICKLE, …?
● Is the t-shirt in this image CREW, V-NECK, POLO, …?
UNDERSTANDING

NATURAL LANGUAGE
● What is the sentiment of this text?
● What action and parameters is a customer requesting?
RECOMMENDATION
● Given your purchases, what else might you buy?
● Given your contacts, who else might you know?
WHAT TYPES OF PROBLEMS?
3
CONSTRUCTION:
Weatherboard
EXIF LOCATION
BEDROOMS
PRICE: $1M
COMPOSABLE SOLUTIONS
BEDROOMS:

3
PHOTO
CBD DISTANCE:
10km
CONSTRUCTION:
Weatherboard
EXIF LOCATION
CBD DISTANCE:
10km
EVOLVABLE COMPOSABLE SOLUTIONS
BEDROOMS:

3
PHOTO
CONDITION:

Good
PRICE: $1.1M
16
CONTINUOUS
INTELLIGENCE
THE NEW NEW PRODUCT
DEVELOPMENT GAME
PRODUCT
RULES

+ Software architecture
CODE
Research
& Analyse
Code
Deploy
Test
Validate
CUSTOMER OBJECTIVES
THE NEW NEW NEW PRODUCT
DEVELOPMENT GAME
Software Engineering Machine Learning
DATA SET

+ Network architecture
MODEL
Research
& Analyse
Code
Deploy
Test Test
Validate
Curate &
Transport
Train
Deploy
RULES

+ Software architecture
CODE
PRODUCT
CUSTOMER OBJECTIVES
THE NEW NEW NEW PRODUCT
DEVELOPMENT GAME
Software Engineering Machine Learning
Validate
RULES
GENERATE
DATA
USAGE
GENERATES
DATA
RULES

+ Software architecture
CODE
PRODUCT
CUSTOMER OBJECTIVES
DATA SET

+ Network architecture
MODEL
DEPLOYMENT
REQUIRES
ENGINEERING
TRANSPORT
REQUIRES
ENGINEERING
THE NEW NEW NEW PRODUCT
DEVELOPMENT GAME
Software Engineering Machine Learning
Validate
RULES

+ Software architecture
CODE
PRODUCT
CUSTOMER OBJECTIVES
DATA SET

+ Network architecture
MODEL
Model
Capacity
Data
Estate
PROGRESS
SWEET ZONE
X
RAPID START BUT FRAGILE AND
LIMITED DIFFERENTIATION
(EG MLAAS)
X
SIMPLE TO DO BUT LEAVES
UNREALISED DATA VALUE
(EG SPREADSHEETS)
ITERATIVE SOLUTIONS
PoC IDEA
TOO MANY PoCs?
MAKE IT
SIMPLER
TEST IN
LAB?
DEPLOY TO
PROD
COLLECT,
EVALUATE
MODEL
ITERATION
DARK
LAUNCH
OR A/B
TEST IN
LAB
ADD VALUE
UNCERTAIN
HOW TO ADD
VALUE
PoC IDEA
CLEAR VALUE
ADD
REPEAT! REPEAT!
23
ML PRODUCT
SQUADS
ML Principals
Product Squads
Scale Partners
ML PRODUCT SQUADS
Self-service data platform
Consumer
responsibilities
Producer
responsibilities
Organisational agility
and learning culture
HCD+ML Diverse Teams
25
ETHICS, RISK 

& GOVERNANCE
RISKS
Bias
Explainability
Attack Susceptibility
! !
TECHNOLOGY RESPONSES
Baseline Model
Asymmetric Costs
Wide & Deep Networks
Fallback on low Confidence
Use Canaries as Early Warning
Black/white list inputs/outputs
In 2018, AI will gain a moral compass
MUSTAFA SULEYMAN, DeepMind
“
“
CORRUPTION OF SOCIAL SPACES
DATA COLLECTION & SECURITY
SOCIAL LICENCE
ETHICS OF ALGORITHMIC DECISION MAKING
AUTONOMOUS WEAPONS
29
GETTING STARTED
We Can Build
Iteratively
AI is Narrow
It’s About People
CONTINUOUS
INTELLIGENCE
COMPOSABLE
INTELLIGENCE
INTELLIGENT
EMPOWERMENT
THE NATURE OF SOLUTIONS
WHERE TO FIND OPPORTUNITY
BUSINESS
CUSTOMERS
EMPLOYEES
VOLUME QUALITY SCALE
VERY FIRST STEPS
! Introduce a step for Machine Learning
in your data pipeline (input == output)
! Look for a binary Yes/No suggestion or a
suggested delta to existing results
! Then iterate
THANKS
David Colls- dcolls@thoughtworks.com

Continuous Intelligence - David Colls (ThoughtWorks Live)

  • 1.
  • 2.
  • 3.
    AMAZING OPPORTUNITY There isalmost nothing we can think of that cannot be made new, different, or more valuable by infusing it with some extra IQ. The business plans of the next 10,000 start-ups are easy to forecast: take X and add AI. Kevin Kelly – The Inevitable “ “
  • 4.
    AND UNPRECEDENTED RESOURCES 2 DEDICATED PARALLEL HARDWARE 3 ADVANCED MODELLING 1 WEB-SCALE DATA DATAVOLUMES DOUBLE EVERY YEAR MASSIVE ADOPTION OF GPU (AND TPU, HPU, FPGA) LATEST RESEARCH AND TOOLS
  • 5.
    NO WONDER AIGENERATES INTEREST FROM LEADERS But the big change from last year is that 81% cited artificial intelligence and machine learning as either very important or extremely important to their company’s future, up from just 54% in 2016. Fortune.com Fortune 500 CEO survey 2017 81% “ “
  • 6.
    There is almostnothing we can think of that cannot be made new, different, or more valuable by infusing it with some extra IQ. Kevin Kelly – The Inevitable “ “ THE HYPE PROBLEM So how do we think about opportunities? AI CAN DO ANYTHING WE WANT (IT’S MAGIC)
  • 7.
    There is almostnothing we can think of that cannot be made new, different, or more valuable by infusing it with some extra IQ. Kevin Kelly – The Inevitable “ AND THE KERNEL OF THE TRUTH “ AND ABOUT PEOPLE AND WE CAN BUILD ITERATIVELY AI IS NARROW (BUT COMPOSABLE) There is almost nothing we can think of that cannot be made new, different, or more valuable by infusing it with some extra IQ. Kevin Kelly – The Inevitable There is almost nothing we can think of that cannot be made new, different, or more valuable by infusing it with some extra IQ. Kevin Kelly – The Inevitable There is almost nothing we can think of that cannot be made new, different, or more valuable by infusing it with some extra IQ. Kevin Kelly – The Inevitable
  • 8.
  • 9.
  • 10.
    AUGMENTATION OVER AUTOMATION MACHINES Widerlearners Scalable thinkers HUMANS Faster learners Flexible thinkers +
  • 11.
    THE INTELLIGENT EMPOWERMENTCHALLENGE High volume 
 judgement tasks are high cost 
 or variable quality High quality human judgement is hard to deliver at speed or scale Empower all your people with the judgement of your best performers Leadership Lenses BUSINESS CUSTOMERS EMPLOYEES Outperform with insight Experience driven by AI Your people at their best
  • 12.
  • 13.
    PROBLEM EXAMPLE PREDICT ANUMBER ● How much will this house sell for? ● How many shoes will sell this week, based on recent sales? BINARY PREDICTION ● Will this customer churn this week? ● Is this transaction fraudulent? PREDICT A CATEGORY ● Is this customer NEW, LOYAL, FICKLE, …? ● Is the t-shirt in this image CREW, V-NECK, POLO, …? UNDERSTANDING
 NATURAL LANGUAGE ● What is the sentiment of this text? ● What action and parameters is a customer requesting? RECOMMENDATION ● Given your purchases, what else might you buy? ● Given your contacts, who else might you know? WHAT TYPES OF PROBLEMS?
  • 14.
    3 CONSTRUCTION: Weatherboard EXIF LOCATION BEDROOMS PRICE: $1M COMPOSABLESOLUTIONS BEDROOMS:
 3 PHOTO CBD DISTANCE: 10km
  • 15.
    CONSTRUCTION: Weatherboard EXIF LOCATION CBD DISTANCE: 10km EVOLVABLECOMPOSABLE SOLUTIONS BEDROOMS:
 3 PHOTO CONDITION:
 Good PRICE: $1.1M
  • 16.
  • 17.
    THE NEW NEWPRODUCT DEVELOPMENT GAME PRODUCT RULES
 + Software architecture CODE Research & Analyse Code Deploy Test Validate CUSTOMER OBJECTIVES
  • 18.
    THE NEW NEWNEW PRODUCT DEVELOPMENT GAME Software Engineering Machine Learning DATA SET
 + Network architecture MODEL Research & Analyse Code Deploy Test Test Validate Curate & Transport Train Deploy RULES
 + Software architecture CODE PRODUCT CUSTOMER OBJECTIVES
  • 19.
    THE NEW NEWNEW PRODUCT DEVELOPMENT GAME Software Engineering Machine Learning Validate RULES GENERATE DATA USAGE GENERATES DATA RULES
 + Software architecture CODE PRODUCT CUSTOMER OBJECTIVES DATA SET
 + Network architecture MODEL
  • 20.
    DEPLOYMENT REQUIRES ENGINEERING TRANSPORT REQUIRES ENGINEERING THE NEW NEWNEW PRODUCT DEVELOPMENT GAME Software Engineering Machine Learning Validate RULES
 + Software architecture CODE PRODUCT CUSTOMER OBJECTIVES DATA SET
 + Network architecture MODEL
  • 21.
    Model Capacity Data Estate PROGRESS SWEET ZONE X RAPID STARTBUT FRAGILE AND LIMITED DIFFERENTIATION (EG MLAAS) X SIMPLE TO DO BUT LEAVES UNREALISED DATA VALUE (EG SPREADSHEETS) ITERATIVE SOLUTIONS
  • 22.
    PoC IDEA TOO MANYPoCs? MAKE IT SIMPLER TEST IN LAB? DEPLOY TO PROD COLLECT, EVALUATE MODEL ITERATION DARK LAUNCH OR A/B TEST IN LAB ADD VALUE UNCERTAIN HOW TO ADD VALUE PoC IDEA CLEAR VALUE ADD REPEAT! REPEAT!
  • 23.
  • 24.
    ML Principals Product Squads ScalePartners ML PRODUCT SQUADS Self-service data platform Consumer responsibilities Producer responsibilities Organisational agility and learning culture HCD+ML Diverse Teams
  • 25.
  • 26.
    RISKS Bias Explainability Attack Susceptibility ! ! TECHNOLOGYRESPONSES Baseline Model Asymmetric Costs Wide & Deep Networks Fallback on low Confidence Use Canaries as Early Warning Black/white list inputs/outputs
  • 27.
    In 2018, AIwill gain a moral compass MUSTAFA SULEYMAN, DeepMind “ “
  • 28.
    CORRUPTION OF SOCIALSPACES DATA COLLECTION & SECURITY SOCIAL LICENCE ETHICS OF ALGORITHMIC DECISION MAKING AUTONOMOUS WEAPONS
  • 29.
  • 30.
    We Can Build Iteratively AIis Narrow It’s About People CONTINUOUS INTELLIGENCE COMPOSABLE INTELLIGENCE INTELLIGENT EMPOWERMENT THE NATURE OF SOLUTIONS
  • 31.
    WHERE TO FINDOPPORTUNITY BUSINESS CUSTOMERS EMPLOYEES VOLUME QUALITY SCALE
  • 32.
    VERY FIRST STEPS !Introduce a step for Machine Learning in your data pipeline (input == output) ! Look for a binary Yes/No suggestion or a suggested delta to existing results ! Then iterate
  • 33.