The most important thing to get right about AI is what you are trying to optimize.
For all the hype that exists around the smart home, self-driving vehicles, fear about losing jobs, I have shut off the “auto-scheduling” on my Nest, despite being a machine learning expert. I have an Audi that I find impossible to drive smoothly. And I grew up on a farm with a Cockerel that was relentless in his attacks on my legs.
What do all these things have in common? Systems optimizing the wrong metric!
Take my Nest – my energy utility gives me money back for having installed it, because of its energy efficiency, which is great. But how is this energy efficiency achieved?
Well, the Nest spends as much time as is possible turning the heat down. Nest uses its motion detection and some location cues in the app and interacts with my Xfinity@Home service to make decisions about when to turn the heat up, but I find we are continually fighting the Nest’s “auto-scheduling” to turn the heat up. Why the problem? Well, our primary metric is the comfort of our home, and our secondary metric is to not waste energy, rather than our primary metric being to cut energy cost.
And why can’t I drive my Audi smoothly? Well, some of my friends would say I’m a terrible driver. Others, our 8-year-old son included, would say “you drive everything like it’s your race car.” Setting those elements of potential “user at fault” aside let me tell you:
- I can’t smoothly slow it down and come to a halt – because the car changes phases of braking to allow it to engage the alternator to recharge the battery, and because it avoids engine braking.
- I can’t pull away smoothly because the engine stop-start takes a moment to restart the engine from a stop, and because the calibration of the gas pedal is such that a large amount of travel is needed to get the vehicle to pull away with ease.
Why are these things that way? Because Audi’s key metric in developing the car, once meeting the demands of the market segment, is fuel economy. My key metric, however, is ensuring a comfortable ride, particularly when it is all five of us in my family plus the dog.
So what about the Cockerel? He’s not tech, so why mention him at all?
Well, while perceived to be otherwise, chickens have been found to be quite intelligent. That said, the male of the species’ response to provocation is the very definition of singular focus on a primary metric. I remember, growing up, the Cockerel deciding that I was a threat and relentlessly attacking me – spurs and beak fully bared. Even after flinging him away from me 20-30 ft., he would come right back at me. His primary metric being to kill me.
Artificial intelligence and machine learning are all about optimization – maximizing or minimizing a metric (it’s what led me to thinking about unique approaches to optimization using quantum mechanics early in my career, thinking which now underpins so-called “noisy quantum computers” like the D-Wave 2000Q). What the Nest, Audi, and Cockerel all have in common is that, at least for my liking, they all optimize the wrong metric, using either biological or artificial intelligence.
The cockerel epitomizes what is scary about AI – that decision-making directed only at the singular motivation to dispassionately kill is quite terrifying. More benignly, the Nest and the Audi are chasing someone else’s definition of optimum: temperature management or driving dynamics, rather than my own.
I have had the privilege of working with many enterprises’ first explorations of applying machine learning technologies to their business, and seeing first-hand the intuition-shattering effect of what happens when you have a technology that enables you to chase what you believe is your primary metric:
“I want to engage customers and make them fans of my brand – by giving them access to some of my products for free for a period of time – and I’m sure it will drive my revenue.”
Well, in this customer’s case, machine learning engaged tens of thousands of customers, with their free products, with the impact that it distracted them from spending money on their paid products – engaging customers increased “fan-dom” but impacted revenue substantially to the negative.
As a result, the customer optimized their program to target customers in a low revenue, low engagement segment, with a new primary metric looking for incremental revenue.
“I need to retain customers at all costs – emphasize the richest possible retention incentives – we’ll make it up in lifetime value of the retained customers.”
This customer was wildly successful in converting their worst-in-industry churn into a close-second best, but the revenue cannibalization was killing the topline of their business.
The customer harnessed some of the best learnings about the marketing elements that went into successful retention, but focused marketing optimization on a primary metric of revenue-neutral customer retention.
It’s remarkable how frequently I have heard “we tried some machine learning at our company, but it didn’t work” – the scenarios above are examples of that. I can guarantee you ML/AI technologies work – indeed AI will relentlessly seek whatever primary metric you set for it. That just means that to get your intended result, never was the saying more true that you must “be careful what you wish for.”
When I am developing solutions for automotive, mobile, banking, gaming and retailers I ask myself and our customers – what is the primary metric you care about? Will your technology drive you toward it or away from it?