-
Introduction
-
Machine intelligence
-
AI - not really intelligence - but critical component of intelligence
-
AI as prediction technology
- central input to decision making
- economic has framework to understand decision making
- new and poor understand implications of advances in prediction technology
- old an well understood logic of decision theory
- way to understand AI
- tradeoffs
- more speed
- less accuracy
- more autonomy
- less controll
- more data
- less privacy
- Book: method
- evaluate pro and cons of applied AI to particular business
- optimal ai strategy
- optimal ai tools
-
Cheap Changes Everything
-
economic understand business implications of cheaper prediction
-
prediction machines will be used for traditional "prediction" tasks (inventory, demand forecasting, and new problems - like navigation)
-
drop of cost for prediction, will reflect on other costs
- diminished values of substitutes
- human prediction
- increase on value of complements
- data
- judgment
- action
-
organization - exploit prediction machines
- AI insight - reframing problems as prediction problems
- adapting ai tools
- assisting in adapting current strategies
- changing strategies
- opportunities to do what before was not possible
- if Amazon know what customers want - change from shop/ship to ship/shop strategy
- new set of tradeoffs (AI impact society) - based on needs and preferences - different - cultures and countries
- 1. Prediction
- 2. Decision making
- 3. Tools
- 4. Strategy
- 5. Society
-
1. Prediction
-
prediction machine magic
-
filling in missing information
-
taking information you have (data)
- generate information you do not have
- about future
- about present
- about past
-
improvement in accuracy
- deceptive (not intuitive)
- can be transformational
- seems magical
-
usage
-
traditional usage of predictions
- fraud
- credit worthiness
- health insurance
- inventory management
-
emerged
- object identification
- language translation
- drug discovery
- New usage
-
Why it is called intelligence
-
machine learning science has different goals from statistics
-
statistics - being correct on average
- articulated of hypothesis / human intuition in model specification
-
machine learning - operational efficiency
- could have biases
- experiment and improvements
- advantage of rich data
- fast computesr
- no need for specification in advance
- much more complex models with anticipation of interactions
- advanced in machine learning
- advances in AI
- 1. system predicated on this technique learn and improve over time
- 2. produces more accurate predictions then other approaches
- 3. enhanced accuracy - performing new tasks (translation, navigation) - previously exclusive - human intelligence
- prediction is central to intelligence
-
data is new oil
-
types of data
- 1. training data
- 2. input data for predicting
-
3. feedback data
- improving prediction accuracy
-
data collection is costly
-
investmet
- how much is enough
- how intrusive collection process is
- balance data acquisition / enhanced prediction accuracy
- ROI for each type of data
- how much cost to aquire
- how valuable prediction will be
- approaches
- statistical point of view
- data diminished value
- more data - more value
- economic point view
- ambiguous
- more data can reach threshold of new insight, but when threshold is reached - new data is not very useful
- must understand relation between adding more data and accuracy of prediction
-
new division of labor
-
humans - poor prediction in certain conditions
- overweight salient information
- do not account for statistical properties
-
machines and humans distinct strength and weaknesses in context of prediction
- adjusting division of labor
- prediction machines - better in complex interaction among different indicators / rich data
- human accuracy diminishes / relative to machines
- humans is better when understanding data is advantage - especially: tin data
-
Taxonomy of prediction settings
- known/knowns
-
known/unknowns
- machines - do not known
- people - much better
- understand inaccuracy
-
unknown/knowns
- machines makes predict, but wrong
- data form decision
- data comes from decision
- machines do not understand how data is prepared
- the counterfactual problem form machines
- reverse causality
-
unknown/unknowns
- machines can't predict
- people can't predict
-
Scales
-
prediction machines scales
- unit cost per prediction fall as frequency increase
- most predictions predicated on routine, regular data
- rare event - need help from human
-
Human machines do not scale the same way
- use cognitive model how world works
- prediction on small amount of data
- human prediction by exception
-
2. Decision making
-
unpacking decisions
-
Reasons of prediction machines usefulness: prediction
- Cheaper, faster, better predictions
- Key ingredient in uncertainty
- ubiquitous - in economic and social life
-
prediction is not a decision
-
components
- prediction
- impact humans / assets
- value of substitutes (human prediction) - decline
- value of compliments (data collections, judgment, actions) - more valuable
- judgment
- knowledge
- judgment value increase, as prediction become cheaper
- action
- outcome
- 3 types of data
- input
- training
- training - relations between data
- feedback
-
value of judgment
-
prediction machines increase return to judgment
-
lowering cost of prediction - increase value of understanding reward associate with action
- judgment is costly
- uncertanity
- payoff for different actions
- time
- effort
- experimentation
-
predicting judgment
-
humans - 3 types of data machines don't
- human senses
- ultimate arbiters
- privacy
-
machines can learn to predict human judgment
-
learning limits
- lack of data
- little data - humans uses, machines don't
- analogy
- models
- individual preferences
- machines - bed in prediction rear events
-
taming complexity
-
compromising - complexity to predicting
-
human satisfying
- good enough
-
better prediction
- less satisfying
- many if - then
-
full automated decision making
-
machine perform all components
- task is fully automated
- humans removed from the loop
-
automation first - when / where higher return
- 1. other elements are already automated, except prediction
- 2. return of speed of action is high
- autonomous vehicle
- 3. return of reducing predicting time is high
- space exploration
- Risks
- externalities
- regulation - barrier
- law - regulatory issues
-
3. Tools
-
deconstructing workflow
-
ai tools - point solutions
- related to specific task
-
braking down - in set of tasks
-
ROI - building / buying ai
- working top / down in building ai solutions
- immediately test benefit
- increasing productivity
- rethinking
- reingeneering whole workflow
-
decomposing decisions
-
decompose tasks
- estimate benefits / cost of implementation
- rank ROI - highest/lowest
-
implementing AI tools - until ROI makes sence
- use ai canvas for decomposition process
- every task / decision
- data (3 types)
- what have to be predicted
- core prediction
- objective
- action possibilities
- outcome possibilities
-
job redesign
-
job
-
collection of tasks
- some tasks automated
- other tasks change
- new task created
- AI tool
- implications
- augment jobs
- contract jobs
- reconstitution of jobs
- eliminate tasks
- add new tasks
- shift emphasis on skill required
-
4. Strategy
-
AI in see sweet
-
not delegate AI strategy to IT department
- beyond enhanced priductivity
-
changing strategy
- core tradeoff in business model
- shop/ship vs ship/shop
- trade of - uncertanity
- decrease expenses but increase uncertainty in choice
- AI help in handling of the uncertainty
- tool
- implementation in one part - affect other parts of business
- redesign workflow
-
increase value of complements
- judgment
- change in organizational hierarchy
- actions
- competitive advantage
- data
-
AI transforming business
-
human judgment
- more important
- subjective performance evaluation
-
key strategic points
- where business end
- where another business begin
-
boundary
- uncertainty influences choice
- prediction machines reduce uncertainity
- increase ability to write contracts
- increase incentive - contract capital equipment and labour
- data
- prediction
- action
- decrease incentives to write contract
- labour
- judgment
- if well specified - programmed
- do not need humans
- in house enployment arise
- contracting out labour - fall
- increase incentives to own data
- decide
- purchase predictions
- purchase data
-
learning strategy
-
shifting to AI first
-
downgrading previous priorities
- real trade off
- maximizing AI goals
- compromising other goals
-
disruptive
- startups - focused on AI, not based on old priorities
-
products inferior at first
- time to train
- learning and improving
-
Managing AI risk
-
many types of risk
-
predictions lead to discrimination
- liability
-
ineffective if data is sparse
- quality risk
- unknown/known type
- prediction with confidence, but wrong
-
incorrect input data
- fool prediction machines
-
diversity of prediction machines
- individual / system level outcomes trade off
- less diversity
- benefit individual level performance
- increase the risk of massive failure
-
interrogated
- intellectual property thief
-
feedback can be manipulated
- destructive behavior
-
5. Society
-
beyond business
-
many choices
-
productivity / distribution
- enhances productivity
- not evenly distributed welth
- increase competition between humans for the remaining tasks
- lowering vages
- increase value for owner of capital
- skill biased
- benefit highly skilled worker
-
innovation / competition
- scale economies
- increase returns
- better prediction - more users
- more users - more data
- more data - better prediction
- monopolization
-
performance / privacy
- access to more personal data
- reduced privacy
-
Finite type of actions
-
Judgment - transfer to prediction machine
-
reward function ingeneering
- automation of decision
-
to many options
- judgment - transfer to humans