Summary Nuggets Flashcards
AI use cases:
· Correct Typos/Grammar Spelling
· Understand Spoken Commands
· Aisles in Kroger are optimized
Read handwriting
Supervised Models
Cat vs dog that were tagged by a human
Unsupervised Models
no labels, looks for natural groupings [cluster analysis; associate rule]
Reinforcement Learning Models
tradeoff between exploration and exploitation = using the existing model to drive decisions. [multiarmed bandits]
Four Principles of Transformation
- One strategy
- Architecture clarity
- Agile Product focused organization
- Clear multidisciplinary governance
Starting point for IT Security
- Know your assets
- Protect your assets
- Detect
- Recover
Speed of Trust Formula
Strategy X Execution X Trust = RESULTS
Network Effect
First Fax machine by itself wasn’t very helpful; need a network to communicate to
AirBNB very well clustered and builds upon itself
Uber not as much as its focused on geographical regions
Multihoming
ability to be on two networks/across platforms. Alexa wants you on their network and no other. iPhone and android similar [opposite of walled garden]
Millionaire Mindset
Integrity is key, having passion for your customer and set your leader standard work
People Development –> Drive Line
You should pay enough to take the issue of money off the table –> then the h care about:
1) Autonomy
2) Mastery
3) Purpose
Zero Trust
· is a framework, a single product will not solve your problems
· will replace VPN users request and are validated at the application level vs network level
· No more walled garden, microsegment the applications
· Start with highest value/highest risk apps
Use AI/ML to know what normal good traffic looks like
Verizon DBIR
(Data Breach Investigation Report)
Incident
(1 million in 2022) a security event that is a POTENTIAL EXPOSURE compromises integrity of an asset
Breach
(.25 million in 2022) an incident that results in a confirm disclosure of data to an authorized party
83% of all incidents are external actors
Actors [security]
who is behind the event?
Actions [security]
what tactic or actions did they do?
Confidentiality [security
data exposed to an unauthorized actor
Integrity [security]
data that is changed from original state
Phishing [security]
send dubious attachment
Social Engineering [security]
responding to email asking to update banking info (with PRETEXT)
Why MFA
MFA is critical, need to do the basic things.
CIS Critical Security Controls
more focused on cyber than NIST
Vanity Metrics
sound useful but don’t actually measure product performance (app downloads vs orders)
If a product isn’t continuously used
it’s easily replaceable
Acquisition
how the users come to your products
Activation
users first visit to your product in their experience
Retention
the user liked your product to use it again
Referral
the user liked the product enough to refer someone
Revenue
the user finds the product valuable enough to pay for it
Understanding who is out there…5Cs:
- Company- why does the company exists?
- Customer-who are their customers?
- Collaborators - what external people make this product possible (operators, distributors etc)
- Competitions - who is competing for your customers money?
Climate - what is the macro environment factors
Kano Model
- Basic Feature - what customers expect and will be unhappy if they don’t have it (toilet paper in AirBNB, gas pedal in car)
- Performance Feature (satisfiers) - elements that are not absolutely necessary but increase customer enjoyment (HAVING SUPER FAST FREE INTERNET AT YOUR HOTEL)
Excitement features (delighters) - unexpected wow features that become product differentiators IF YOU ARE MISSING THE CUSTOMER STAY NUETRAL
- Performance Feature (satisfiers) - elements that are not absolutely necessary but increase customer enjoyment (HAVING SUPER FAST FREE INTERNET AT YOUR HOTEL)
PPI
Pay per impression (PPI) - pay for the add whenever its shown (impression = someone saw the ad)
PPC
Pay per click (PPC) - pay when someone clicks on it
PPA
Pay per action (PPA) - pay when a final action is complete, aka downloaded the app
CTR
Click through rate (CTR) - percentage of people who click through on your add
CPI and CPM
Cost per impression (CPI) or cost per 1000 impressions (CPM) - how much you made to have your add shown once
CPC
- Cost per click (CPC) - this is the actual price you pay for each click in PPC advertising
CLV
Customer lifetime value (CLV)- how much money do you expect to make from this customer over products lifetime
CDO Summit - GENAI Takeaways
- We are at a trailblaze moment of GenAI. Need to close the gap in data available to knowledge.
- Doing your own GenAI LLM is too cost prohibitive right now, in context learning and passing variables
- This isn’t an efficiency play, this is a creativity play
- ML is hard to explain, GENAI is even harder
- 4 safety layers [you can’t drive a car without breaks]
- Model safety
- Data safety
- Prompt safety (if the model was trained with sensitive data, a prompt can get it out)
- Regulatory safety
70% of all data is unused/untapped
Data Mesh vs Data Fabric
- Mesh (decentralized)
* Domains own their data as a product
* Engineering team focuses on the infra and processes- Fabric (centralized)
Publishing and unifying across Meshes
- Fabric (centralized)
Data Literacy
context of the data
Getting Value out of Data
- Getting new customers
- More value out of existing customers
- Operational efficiencies, aka Tech Support
- Need to benchmark and show progress and value of data
- Self Service and training will uplift your users (RDF training)
- If you want AI (and all its value) you need clean data as the solid foundation
Steps of ML OPs
1) Explore data
2) Data Prep
3) Model Training and Tuning
4) Model Review and Governance - need to understand lineage of model
5) Model Inference and Serving - Running model in prod and serving up data
6) Model Deployment and Monitoring
7) Automated model retraining
Promotion Order
Dev to Staging to Prod
Staging has prod data
Deploying/Promoting Model or Code
Code will output a model.
Pros of code
If you promote the code you can auto retrain the model and handle data discrepancies across environments.
Pros of model
No need to deal expensive retraining or model changing greatly.
ML Ops Guiding Principles
1) Automated
2) Secure but self service
3) Repeatability
4) Fast feedback loop
5) Model evaluation
6) Platform and cloud agnostics
Steps for a Strong AI Foundation
1) Create strong and scalable data foundation
2) Attack with use cases and tool specific to your industry (architecture and models trained for your industry)
3) Have transparency in the data and the models and the right governance model
Jag Nubank Product Podcast Takeaway
Fundamentally different vs incremental
Customer obsession.
Hire smart people
Empowered cross functional teams with critical metrics.
Churn and Sean Ellis Test. How would you feel if you could no longer use this product.
The responses are:
1. Very disappointed
2. Somewhat disappointed
3. Not disappointed
The goal is to reach a stage where at least 40% of your users would answer “Very disappointed” if they could no longer use your product.
Mock press release
What are the key principles for decisions? How do we put the rules on paper? This will help with empowerment
Radical Candor
Care Personally and Challenge Directly
Gen AI Financial Sector Use Cases
Will supplement other ML activities (not replace)
1) Dev Enablement
2) Business Intelligent (ask questions about your data)
3) Summarize Feedback
4) Know your customer
5) Chat Bots
6) Hyper personalized marketing material (auto gen)
