AI Won’t Save Your Business – You Will!

What follows is a slightly serious look at AI with a focus on realistic, achievable goals and reducing your risk of painful, pointless burn. Target Audience: You, my friend!

TL;DR

Artificial Intelligence (AI) = Computer programs that make “intelligent” decisions — often developed using Machine Learning.  NOTE: A very broad *BUZZY* term.

Machine Learning (ML) = Programming technique which learns and improves in service of making “good” decisions. ML algorithms typically need a lot of data to train them.

AI can be an extremely effective tool in helping you serve your customer’s needs, but it bears repeating that it is still just a tool. You should be thinking how to use it to improve your product or business, not because “everyone else is doing it”. AI strategy entails unique challenges and thoughtful application that benefits from a buzz free approach. Read on for more of the soothing balm of a reality driven discussion!

STATE OF THE BUZZ

It’s well understood that the tech industry will lose their minds about some shiny “new” object every year.  This leads to an unhealthy FOMO (fear of missing out) and bad decision making. In SXSW (the freakishly huge annual music/film/interactive conference hosted every year in Austin, TX) that buzz is made manifest by freakishly long lines and completely full sessions.

When I attended the conference this year, I learned exactly two things:

  • Austin pedicabs are the least economical way to travel in the world
  • AI WILL CHANGE EVERYTHING (RRRRAAARRGGH!!!)

As a short technology refresher, the AI of today does some things very well:

  • pattern finding (find something in this gob of data that might be informative)
  • supervised learning (here is what a big bunch of data means, now tell me what this fresh data means in context)
  • collective intelligence (here’s some rules for how to live your life as part of a swarm, now go accomplish a shared goal and learn and adapt from each other)

There’s plenty more, but I mention these since they’re pretty well established and/or directly related to the sessions I’m about to listicle. There’s a good primer on kinds of AI here if you’re interested.

NOTE: A big part of being a boss-level AI person is assembling together those disparate pieces of AI (doing smaller things well) into a coherent SYSTEM that serves your business goals well.

STATE OF THE BIZ – MINI SXSW Report

Buzz aside, sharing is caring and conferences like this are great for getting ideas, learning and taking the pulse of our business. The lines were freakishly long for AI sessions and that says something as well. Here are some tidbits from interesting sessions I was actually able to attend.

Machine Learning: The Art of Explore vs. Exploit – Oscar Celma, Director of Research – Pandora

Absolutely the best session of the season. Oscar’s talk was grounded in real, practical application of what’s possible and beautifully tempered with Pandora’s business goals and realities.

Summary: A discussion of how Pandora uses ML to help their users EXPLORE new music or EXPLOIT what the system has already learned to give them music they know they like. Also how these two goals are frequently in opposition.

Key points:

  • Pandora’s business is great blend of human intelligence (using musicologists to author every song’s DNA characteristics AND using those characteristics as input into ML algos
    Pandora’s approach is also extremely data driven — i.e. one of the only ways to discover if a recommender is working is to run an A/B test on a subset of users.  These tests take at least 3 months in order to gather enough data.
  • Pandora uses lots of secret sauce — They assemble a subset of 75 different recommender (little ML/AI algos that help pick your music) into an ENSEMBLE RECOMMENDER that works for you. The ensemble changes and learns based on what works for you and usually ends up with 3-4 of the 75 that are your muscial BFFs.
  • THUMBS ARE WHY IT ALL WORKS! — A binary yes/no is how the system learns what works for you. (see Netflix’s ditching of stars for an argument of why it’s so! http://www.theverge.com/2017/3/16/14952434/netflix-five-star-ratings-going-away-thumbs-up-down)

Additional Reading: At Pandora, Every Listener Is A Test Subject

The Future of Warfare — Will Roper, Director — US Dept. of Defense’s Strategic Capabilities Office

Summary: Drone swarms. Drone swarms are coming. Don’t be afraid, but swarms of drones are coming soon. This was a general talk about the recently declassified SCO which is focusing heavily on collective intelligence, i.e. swarms!
Key Points:

  • Providing lots of scenarios where ML and AI can be applied is a force multiplier
  • SCO’s focus is on using “Teams of Machines” to achieve tactical and strategic objectives.
  • A human being should fall into a “Quarterback” role.  Give broad mission directives and make key decisions (i.e. ones that might take a life) — Translation: Leave the general intelligence to the humans.
    The future of warfare is software — i.e. those who have the best algo’s win.
  • Although the subject matter is mildly terrifying, it is still a very modest application of machine learning with a focus on serving the customer’s needs, the customer here being the DOD and their needs being more effective prosecution of war stuff.

Additional reading: The Autonomous Future of Warfare Looks a Lot Like Pokémon Go

Applying Science to Conversational UX Design — Bob Moore, Research Scientist & Raphael Arar, Designer and Researcher/Artist — IBM Research

Summary: If you’re building a chatbot, it pays to understand the science of conversation and apply your ML/AI techniques judiciously.

I include this session because it’s a great example of using basic science and technology — here the study of Conversational Analysis (a subset of Sociology) — to increase the return on your technology investment by having chat UX that’s maximizing your AI tech investment.

Key Points:

  • Social Science is cool again! I hadn’t heard of CA before this talk, but essentially it’s quantifying rules and systems around how people have an actual conversation. These generalizations can be used to create a virtual agent that feels much more natural.
  • Sometimes less is more.  In the case of a virtual agent (i.e. chatbot), using as little AI as possible is great. Less training means quicker turnaround. Less data, and less likelihood you’re going to alienate your user base by klunky UX.  You’re not trying to create generalized artificial intelligence, you’re just trying to achieve a short-term goal.

Additional Reading: DEEP LEARNING FOR CHATBOTS, PART 1 – INTRODUCTION

Epilogue

No matter how good your data, no matter how smart your algos are, wanton application of technology will not fix any fundamental flaws with your product or business. The most amazing recommendation of a house to buy for your clients doesn’t matter if it turns out all they want to do is rent and you don’t process transactions less than $100,000.

  • Never forget your core mission and that technology is there to serve (even tech that everyone says you ABSOLUTELY MUST HAVE in order to compete)
  • Chat calmly with experienced folks who know what’s possible with today’s AI technology
  • Make sure those folks understand YOUR customers and business
  • Use those learnings to experiment with new ideas and technology to minimize risk/cost and maximize learning — it’s still software development!

Be bad-ass, make informed decisions, ask for help where it makes sense, and get on getting on with the business of what you do that makes you awesome.

Recommended posts