When the path isn't clear, the advisor you turn to should be.

10-Year Archive

Humanizing Tech

Notes from Gigaom’s AI 2017 Conference

Insights from Enterprise AI startups, CIOs, and Data Scientists

I. Setting the Stage

Yesterday we attended the Gigaom’s Artificial Intelligence conference in San Francisco. The focus was on the enterprise, from the perspective of what’s actually being implemented rather than the possibilities of what could come in the future.

All of the notes below are typed using the exact words said by the participants. We did not take any leeway with the language, interpretation, nor have we provided any opinion on this.

What you read below is factually accurate as best as we could copy from the participants themselves.

II. Blueprint for Building & Deploying Enterprise AI Solutions Across Industries

Panelists

  • Tim Crawford (moderator)
  • Somya Kapoor: ServiceNow
  • Josh Sutton: Publicis Sapient
  • Kumar Srivastava: BNY Mellon

Biggest industries from a survey (most participants work in a company with 1K to 25K employees)

  • Financial services
  • Healthcare
  • Manufacturing
  • Construction

Where do people get started with AI?

  • 33% are in evaluation stage
  • 25% in planning and getting to production

Why still early stages?

  • Most organizations are used to the past rate of change
  • But now time period is a much shorter cycle
  • Many people are hurrying to react and do it smartly
  • Now people are getting in deep

Are enterprise projects net new?

  • 90% of AI work is net new activity
  • Existing things in production means you have to refactor what’s already there and takes time
  • This it’s easier to start fresh
  • Traditional devops is not used to dealing with new models
  • Easier to try something new business production container is new. Deploy that up front and no legacy software to deal with
  • Smarter for enterprises to start with net new

Learnings from starting to implement AI (commands)

  • Modify how a user or employee operates
  • Based on patterns this might be the right thing to do, adds an extra step. Safer, harder to inspect a black box AI
  • How to retrain employees for working with AI

People seem to start with Bots as their first experience with AI

  • Vast majority of respondents sit to 5 to 100 bots, not 1–5 buckets
  • Intent and entity abstraction
  • Next up consolidation of bots

Entire customer journey experience (look at the entire thing, not just one small piece) is the holy grail

  • accounting, supply chain, customer service

Are people looking to solve a broad set of problems or specific?

  • Vast majority of respondents are split evenly between a broad approach and holistic approach
  • One problem that gets solved first then use that to build off of for the holistic strategy
  • Agile and BI opened the door
  • Initiatives can start in IT (becoming a service later, reduce headcount)
  • Password reset is one of the biggest requests for Bots

Problems

  • Training the people is one of the hardest things to do
  • Strategy situation from Board to CEO, lots of competition within the firm: different solutions for different problems that don’t work together

What part of the organization started the AI conversation?

  1. IT
  2. Finance (cost cutting as goal could result in problems in the future)
  3. Huge drop off after that

But there’s an expectation that other parts of the organization will be able to consume what comes out of IT. Needs to be a company driven strategy.

  • IT is becoming a provisioning layer so others can consume it
  • Doesn’t matter where it starts just that it does
  • But needs to become consistent with a coherent strategy

In House vs External?

  • 45% are building their AI services in-house rather than getting outsiders
  • Market is flooded with too many vendors
  • ServiceNow taking a two-prong approach (google, Watson) don’t feel like they can solve it all by themselves

What role does open source play? And how?

  • Have the best minds in the world working on this stuff, competing to give away the best services
  • Leverage but know you still have to build on top
  • Majority of AI effort starts with open source
  • Platform company so have to work with open source but have to add more
  • Two types: academic or large tech companies (incentives) both are starting points
  • Use them to go from zero to 60 very quickly
  • Many product companies are based on open source underpinnings
  • 75% of survey respondents say open source plays a very significant role

Explainability is the big problem. How do you rewind the decision-making for transparency for how it works?

Use case

  • Show how a customer incident is correlated to a knowledge article based on machine learning
  • Confidence interval. Can’t say why it’s the answer but can give you 99% confidence that it is right

III. How To Consider Proof of Concept Use Cases, Building & Scaling Your POC

Panelists

  • Soma: Madrona Group (moderator)
  • Jim McHugh: NVIDIA
  • Peter Marx: GE Digital (predix platform, former CTO city of LA)

Jim

  • NVIDIA DGX-1 is a supercomputer in a box
  • Most of the people working on AI don’t have the resources to set up a farm
  • NVIDIA wanted to come out with something turnkey, democratize it and make the compute power accessible
  • Only been about 4 years from AlexNET

Peter

  • GE has 380,000 employees
  • Turbines produce about 12% of the world’s power (scale is crazy)
  • 60% of airplanes run on a GE turbine
  • Been in AI world for a long time, came out of medical imaging, then Apple to work on videos, then work in video games (football AIs, race car AIs)
  • Techniques has been around a long time but what changed is access to data (and processing power

How do you get started in a POC?

Peter Marx:

  • GE looks at asset performance management across all their machines, when will they break?
  • LA has several million trees. Lifetime of a tree is 80 years, 3 years to establish a tree, fancy neighborhoods have more trees. Trees represent equity, how thriving the city is, and different than urban street islands with no trees.
  • With the drought, LA had to figure out what to do with the trees along with 20-year lawsuit on sidewalk damage where replacing a couple hundred thousand trees
  • Need a catalogue of trees, hiring a bunch of people to drive around the streets doesn’t work in today’s world (need more fidelity and automation)
  • Tensorflow and computer vision techniques used to identify species: did it on google street view, the computer is driving down the street and cataloguing getting down to sub-meter resolution
  • How do you do that in a mayor’s office
  • He did a small POC in Caltech then pulled up satellite image from 1996 data (pretty scratchy). Over 20 years many things changed. Get started with a small area, make it highly visible. Then we said we need money to do it across entire city of LA. By the time they got the money, had already done it across the entire city of LA and looked like Gods.
  • Reason GE is moving into Predix is because we’ve reached the limits of performance improvements in our engines, materials science, process engineering. But now we can start to optimize thrust across an entire fleet of planes and power performance. Level of complexity not taken by man before.

Jim McHugh

  • Bring people through their demo center
  • Put cameras in car and let them learn intuitively, after 100 hours they were hitting all the orange cones, after 3000 hours it’s driving through the streets of New Jersey in rain and showing the power of image recognition
  • Identifying objects: “do you know how many warehouse issues that solves for me?”
  • Showed it to consumer companies: P&G, Oil of Olay (do a study of your face and recommend right lotion, 60% to 70% didn’t understand product line, after app it’s a 90% product satisfaction and 88% repurchase rate)
  • Pattern recognition: studying network traffic patterns then detect an anomaly, understand where connections belong; same thing with Voice
  • Always start with Gaming in demo center, incredible amount of computational math, showing raw compute power and simulation (don’t need experience in real life)
  • PRISMA art overlays on top of your photo, generate new scenes for Blade Runner
  • Oxford getting 93% accuracy at lip reading, most humans are just above 50% (what can you do with cars and trucks)

IV. Fireside Chat: Auren Hoffman & Byron Reese

Auren Hoffman Bio

  • Started LiveRamp sold to Axiom
  • CEO of a startup that makes data sets for AI
  • Angel investor

Why was now the time AI is coming out?

  • We have enough data (stock market)
  • Over-predicting self-driving cars, most still driven by humans in the next 10 years
  • Under-predicting other interesting types of AI

What can we do to speed up the development of AI?

  • Really focus on getting the historical data on the truth of what happened
  • Truth is really important (chess moves, stock market mostly true)
  • Building models based on bad data can compound the mistakes very quickly

How much time are they spending on data?

  • 80% to 90% of their time is cleaning, munging, dealing with privacy
  • Spending only 5% of their time on the actual AI
  • One of the reasons Google gets all the researchers is they don’t have to do any of the other stuff around the data, just work your magic (compelling recruiting pitch)
  • Biggest complaint is they have to do all this other stuff, the problem is labeled training data

Enterprise problems

  • 500 vendors yesterday, 5,000 vendors today, another order of magnitude tomorrow
  • Wal-Mart has 1,000 vendors for just marketing technology alone
  • Number of vendors is astronomical and it’s growing
  • How do you manage vendors? Law firm, for instance
  • Instead of using a spreadsheet, now using an API
  • Almost a vendor assembly line, sometimes vendors work together to move data around
  • A company’s DNA is defined by their vendor: it’s like a fingerprint

Thesis for investing in AI companies

  • Have to tease out if someone wants to be cool, or do they have a passion about this particular thing
  • Hard thing to test for
  • Invests in B2B kind of companies, but seems a bad idea
  • Just investing in a good time to be a good time to be an investor

Where do you get data?

  • First, make sure the data is true (bad algorithms from bad data)
  • Scale is important, what is scale?
  • Watson Oncology, data from Sloan Kettering (procedures and outcomes); going to need more data from more sources, guessing the data is labeled, true, and clean
  • Probably want data from all hospitals that don’t have the best people and practitioners, and from all over the world. Variation.

Most interesting data sets that he has at his startup?

  • It’s not about one data set, but many together
  • Quant hedge funds
  • Graphing data cross data sets
  • Originally look at each data set by itself, train and that works relatively well, each data set had some sort of value to get better performance
  • But the real value they found was graphing all these data sets together, then asking questions
  • But by putting these things together, you start to learn a lot of interesting things
  • Nutrition is hard: everything you ate, wealth, DNA of people all together to graph for understanding on what’s going on
  • But if we could we would unlock the benefits to humanity
  • This is basically general intelligence

V. Line of Business AI: How Marketing, Sales, Customer Support, HR, Finance & Product Can Use AI Without Being a Data Scientist

Panelists:

  • Sam Charrington: CloudPulse strategies (moderator)
  • Simon Chan: Salesforce
  • Matt Gandolfo: Charlotte Russe
  • Terry Cordeiro: Lloyds

Charlotte Russe

  • Fast fashion, don’t rebuy products, have to make decisions quickly and repeatably
  • Issue is with a lot of people making decisions manually
  • Instead of just prescribing an action, actually take that action
  • Identify that the fleet is in the right location
  • If you shut a store down can ecommerce pick up the slack
  • They’re looking for commoditized solutions so they don’t have to hire data scientists or machine learning
  • Aren’t going to hire a bunch of data scientists and PhDs because they’re too expensive
  • Components of change management and retraining people when you get rid of their job of updating excel spreadsheets
  • AI is nobody’s full-time job, only have 50-person IT team total
  • So the question is how do they leverage solutions
  • Partnered with a vendor to take data around physical stores and eCom, other attributes to run a bunch of what-if scenarios
  • Problem is it’s built on an external vendor’s platform
  • cost was cheaper, shifting more data to them is more risk, what if data leaks out

Salesforce

  • Einstein platform, AI that runs on top of Salesforce as a CRM tool (came from Prediction.io startup he built)
  • Automate tasks they don’t want to do
  • In addition to metamind acquisition
  • Focus on improving the customer experience, customer improvement
  • Have an AI system that can help them
  • Make their business more efficient, and make AI customizable for any business to use
  • They have 100 machine learning researchers / PhDs
  • Data is already in the cloud and ready, so much easier to get benefit from it
  • Cloud is the infrastructure that can empower AI (get data ready, get model ready, get production ready)
  • When compare in-house development vs external: moving from POC to mass production they start to see the cost difference (maintenance in own IT infrastructure)
  • Pivotal point where people look to external solutions is when the scale and productionization happens
  • Keeping business users in the team with data scientists, that’s when real value is created
  • Privacy and data protection seems to be a big problem (at least with consumer use cases)
  • Trust is #1 value of salesforce

Lloyds

  • One of the products they want to build is a cognitive platform; started thinking about virtual assistants and then extended into the rest of the business (“empower colleagues as well as customers”)
  • Definitely a skills shortage in this area
  • Build a centralized AI team that offers AI as a Service to the rest of the organization
  • Good relationship with IBM Watson so chose them, use Bluemix as their dev environment, use data already had
  • Learning: spent too much time training the model, should have done it quicker and cheaper
  • Have 5 people: data scientist, product manager, scrum master, engineer, architect sit around a table and get something done quickly (paid for by different parts of the bank)
  • Can’t do this on-premise, would be impossible
  • Build components out as services, then put a wrapper around it
  • Use some open-source
  • Their differentiator will always be their data, and the bank should use that intelligently
  • IP will be in AI as a Service platform approach
  • Very worried about sensitivity of data in the cloud

VI. Super Powers of Innovation

Sandy Carter

CEO & Founder, Silicon Blitz (spent 20 years at IBM)

sandy@silicon-blitz.com

Stats

  • Narrative Science has been doing research on AI
  • 80% of business executives in 2017 believe AI will help improve worker performance and create new jobs
  • Gartner predicts 85% of interactions will be done without a human in 2020
  • Bloomberg: $300M as first investment in new AI startups
  • PWC: 1757 execs: 93% of execs depend on innovation to drive growth

Forrester: why companies DON’T use AI

  • 42% there is no defined business case
  • 39% not clear what AI can be used for
  • 33% don’t have the required skills
  • 29% need first to invest in modernizing data management platform

Jeremiah Olang (did a study on innovation)

  • 114 companies, asked them “What is innovation?”
  • Answer was across 4 areas: AI innovation in product, operational, client experience, business model

Startup: 360 Fashion collaborating with Intel, Chinese government, and CCTV (1.2 billion people watched new year’s celebration)

  • Smart glove fashion
  • 162 dancer synchronization, sensor based, learned gestures
  • Use of AI in media and entertainment using these smart gloves

Mastercard

  • #36 on Fortune’s most innovative companies list
  • Who are you, are you you, what can you do
  • 85% of payments are still done with cash

Disruptive business model

  • Atipica
  • Look at a profile and look at unconscious bias in hiring company (removes it from resume)
  • Diversity and inclusion, automatically review matches

Tesla

  • AI in cars, obviously
  • Ecosystem around their technology (if you leave your car at a charging station, keeps charging you)

Marketing AI Use Cases

  • Personalization
  • Content Creation & Curation
  • Recommendation Engine
  • Search Optimization
  • Product Pricing
  • Customer Service
  • Ad targeting
  • Voice recognition
  • Segmentation (her favorite as a past CMO)

Superpowers

  • Super Intelligence: understand the technology, but also the business and use case
  • Super Speed: ability to experiment rapidly
  • Super Synergy: build ecosystem of partners (not just a standalone system)

VII. CIO Panel: Do It Yourself, Partner, Or Buy?

Panelists

  • Tim Crawford (moderator)
  • Paul Chapman Box, prior at HP
  • Tom Keiser: Zendesk, prior Gap, then Victoria’s Secret
  • Elinor Mackinnon: Devcool, prior esurance, prior Blueshield California, prior biopharma company, prior Charles Schwab

Where is AI fitting into CIO agenda, or is it not?

  • Zendesk
  • consumer of AI right now, it is definitely one of the tools they use
  • Don’t really have an AI initiative, have a series of initiatives to drive productivity and scale (which have some AI components)
  • Growing rapidly globally requires customer service touchpoints
  • Global products that need to be secured
  • 13 data centers now, each year a few more, whole series of AI around capacity planning, knowing when bad things are going to happen
  • Box
  • Heavily millennial workforce, taking friction out of product
  • Deliberate set of investments to change how employees interact with technology
  • Dialogue-based user experience, Bots, voice controlled conference rooms, Alexa
  • Expectation that they’re a forward thinking tech company
  • Deliberate that there’s enough meat to move the company forward
  • Insurance (P&C)
  • See a pretty big spectrum
  • Use an internal team for service and support like Lloyds
  • Struggle going from low impact, low risk and moving to something that would have a big impact is very risky and hard for them
  • Attendees at InsurTech: many were from outside the US (lots of innovation globally), many called themselves purely AI
  • Small startups taking a narrow slice of AI combined with something else (dongles and data for car insurance)

Is AI like mobile apps where you plug into the cloud, or does it require something net new?

  • 400 service reps closing out a ticket from 100,000 tickets
  • A reply of “thank you” from the customer re-opens a ticket so machine learning can learn not to open that instead of “thank you, but”

Role of CIO, where does CIO fit in conversation around AI?

  • Depends on how role is defined, sometimes the expertise of other business unit leaders can bring AI to the table
  • Some CIOs are likely spending a lot of their time working on infrastructure-related issues
  • AI slash deep learning, AI slash IoT, AI slash robotics
  • AI is not just in a vacuum or in a silo, it’s always coupled to something else
  • AI for support functions is a big deal, take out the repetitive processes
  • CIO previously was a foreign relationship in the C Suite, now it’s a much closer relationship

The language of IT

  • “AI is the convergence of the physical, digital, and biological set to disrupt everything.”
  • In the world of insurance, who do we insure if nobody is driving a car, do we stay in the business, get out of it, what do we do with the actuaries

Build versus Buy?

  • Most of the survey respondents are using a best of breed approach solution rather than using a platform
  • Outside providers, all of them want to be a platform, all my data is in there, second platform
  • No one is selling you just the component you need, instead they want them to live in their platform
  • Insurance space: move to micro-products, go small instead of selling you an entire platform
  • Don’t know if there ever will be a single platform?
  • Before we pick a platform, lets have a strategic conversation, architecture could get really complicated
  • “Show me the org chart of Ralph Lauren” instead of having to go through single sign on, two-factor authentication, navigate to the person, then open it up

Big use case for email:

  • 3 suggested responses to any email based on your tone and writing style, inherent to the subject of the message

Most innovation in insurance tech

  • UK: less regulation
  • Micro products, people don’t want to buy an entire portfolio of insurance products, they just want to turn some product on or off
  • AI: use service side of it to give insurance customer an entirely different experience in the moment when the problem happens (it how they remain loyal to the companies)
  • Insurance: such a huge corpus of data, all available in the public domain
  • I want to take my bike out of the house, turn my insurance on, bring my bike home, turn the insurance off

— Sean

Recommended Reading For You

  1. Do We Want Artificial Intelligence or Augmented Intelligence
  2. Biologic Intelligence is NOT Artificial Intelligence
  3. A Novel Framework for Creating Self-Learning Artificial Intelligence
  4. The Next Motor of the World
  5. Early Stage Robotics & AI Funding Versus Market Size

Notes from Gigaom’s AI 2017 Conference was originally published in Humanizing Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.



from Stories by Sean Everett on Medium http://ift.tt/2lWt4Yq