Mark Ritson: Machine learning may be the future but it’s still pretty dumb
If machine learning is so clever, why does Spotify recommend such rubbish songs and Amazon suggest the stuff I already own in different colours?
Scott Galloway, the entirely fantastic NYU marketing professor who presents as if he has just imbibed a combination of uppers and downers all in one gulp, gives a great talk where he compares the winners and losers of the last 12 months. He then looks for patterns and themes in the two lists as a vehicle to pull out insights and develop predictions for the year ahead.
One of his many observations is the growing power of ‘algorithm businesses’. By his definition Galloway sees these companies as being able to “implicitly gather data from their consumers and then automatically update the consumer experience”. He sees the most powerful algorithm-driven businesses as those that have the most receptors in the market and move fastest to update and improve their offer based on what these receptors relay back to HQ.
READ MORE: Mark Ritson – In the AI era, it’s digital ads that face disruption
In traditional 20th-century case studies, big businesses usually grew too large to listen to consumers or pivot to meet their changing needs. Algorithm businesses don’t face this growth threat; rather, as their scale increases, so too does the number of receptors in the market and thus their ability to evolve and improve even further.
Every time you use Google, for example, it apparently gets 0.000000003% better. No surprise then that Google joins Facebook, Salesforce, Amazon, TripAdvisor, Priceline, Spotify and Uber among Galloway’s list of the top algorithm-driven businesses. Or that, of the 13 companies to beat the average performance of the S&P 500 consistently, each and every year for the last five years, seven of them were algorithm-driven businesses.
Although Galloway’s postulations are relatively new, his fundamental point is as old as marketing itself. Boil down the elemental advantage of marketing and it’s not communications or positioning or even pricing. It’s way simpler than that. The biggest gifts that good marketing bestows upon an organisation are the abilities to first see the world from the consumer’s point of view and then make appropriate changes to accommodate this perspective in future products or services.
AI versus insight
It does not sound much. You listen to the customer. Then you change what you do in response to the lessons learned from listening. There, I just summarised about 85% of the marketing concept. Simple.
But just because something is simple, does not make it easy. Try actually getting a big company to put down their products and technology and financial projections for even a minute to recognise that all the money comes from the customer. In my career – if I can sound like a wanker for just a second – nothing has stunned me more than my ability to disarm senior executives with the message that we should listen more to customers.
Algorithms should cut out the weak link in the whole understanding-and-responding-to-the-consumer bit.
After many years working for a company it’s too easy for a manager to go native and start thinking customers see the business the way they do. Even marketers, who are meant to always use the wet rag of research to douse the flames of corporate bullshit, have been known to lose the plot and start believing that their consumers want more than a toothpaste/beer/coffee and that they want peace/liberation/community instead from their brands.
And even when you stay market-oriented, the challenge of turning a big corporate ship around to deliver on what customers actually want is equally confounding. Many a marketing director drives home each evening knowing what the target customer wants, certain that his company can deliver on this but also convinced that the bureaucracy and politics and finance that run his company mean that there is no chance that this delivery will ever actually occur.
It usually takes a slightly deranged CEO like Elon Musk or the late, great Sir Simon Marks from M&S to be able to take an insight and, with a dismissive wave of a hand, make the insight a directive that becomes a reality. Usually consumer insight remains with the front line staff and organisational ability to drive change stays with the C-suite – and never the twain shall meet.
Seen in this light, it’s clear why Galloway’s algorithms make so much sense and, apparently, so much money for the companies behind them. Algorithms cut out the weak link in the whole understanding-and-responding-to-the-consumer bit – they remove us, the humans, from the process. Amazon does not have to know why the new Tom Cruise movie sells better with a picture of his upper torso, versus the one of just his face. It just knows that this is the case from the thousands of sales it has already made and not made this week and instantly changes how the product appears to the 98,000 shoppers who will view it in the next hour.
READ MORE: Rise of the machines: Are robots after your job?
There’s no moron from finance questioning whether the research is ‘accurate’. No merchandiser with a long-running hatred for marketing who won’t change the Cruise packaging because his movies ‘usually do better with just the face’. They are washed away in a sea of terabytes and predictive modelling that takes barely a flash of time to understand, implement and move on.
Limits of machine learning
But is there a bit of hype here too? I get the power of algorithm learning. I see it as the modern epitome of market-orientation. I get the correlation with oversized business performance too. I just don’t see any of this first-hand. I am a consumer of many of the companies that Galloway lionises as ultimate algorithm businesses and yet they just don’t seem that smart to me.
Spotify is meant to be pretty clever. It knows what songs I have saved, what I play most often, what my mates play most often and what other people who have the same vaguely 1980s, rubbish taste in music as me play most often. This should make for a pretty convincing ability to recommend some excellent musical additions to my playlists that I was unaware of.
Even after I click on Amazon’s ‘recommended for you’ tab I just get a list of stuff that I have already bought.
And Spotify has a go. It creates ‘daily mixes’, which combine a set of my own chosen tracks with some ‘new discoveries’ based on what the algorithm says I will enjoy. So along with my own tracks (Van Morrison, Billy Joel, Robert Palmer and a bit of Jackson Browne) today I got late Paul Simon (shit), Boz Scaggs (shit) and Toto – you know the one about Africa and Kilimanjaro and the Serengenti-i-i (very shit).
It did serve me Warren Zevon and ‘Werewolves of London’, which I had not heard in a decade, and this made me happy. But it’s hardly the stuff of dystopian nightmares or the business engine of the future that Galloway suggests we might now expect.
A visitor to my house, armed with only with the recent experience of three very good bottles of red wine while I blasted him with cigarette smoke and offensively loud tunes from my outdoor speaker, could have – would have – done much better in the suggestions department.
The improvement is not apparent
Similarly, signing into Amazon should, after about 60 grand worth of purchases over the years, make it know everything about me and then some. But even after I click on the ‘recommended for you’ tab I just get a list of stuff that I have already bought. Basically, baby clothes, a carafe for my coffee maker, and then more baby clothes. I don’t need any of this because I just bought a barn load of clothes for the young Ritson mutant last week and the coffee maker is already gone because it was crap.
Now I appreciate all this is very hard to discern if you’re Amazon just selling me stuff online, but I expected something better or a bit more left-field from the Dark King of Algorithms than a reproduction of what I’ve already ordered. Nestled in my Amazon top 10 list of baby leggings and coffee mugs I expect one completely unexpected, immediately desirable thing that I am strangely drawn to. An exotic musical instrument that only I can play or perhaps a piece of clothing I would never have considered but now realise will make me look fantastic. But there is nothing. Just stuff I already own in different colours.
I should pause here and note that I may not actually be aware of the algorithms around me. I say that because I famously got pissed once and accosted the CEO of a very large airline that I was a patron of. He was very patient with me as I explained to him how bad his airline once was and how, when I had first flown them many years before he took over, the cabin crew never had my first-choice meal in business class.
But, I blustered on, they had improved in recent years. So much so that I could not recall any incident where my preferred food was not available. I tried to ascribe this improvement to his leadership but he stopped me short, explained that I had simply risen up the ranks of his airline’s loyalty program and was now visited first by the air crew to ensure I got what I wanted, and the bloke behind me got the dodgy chicken.
My point is that consumers – even those of us who masquerade as marketers in the daylight hours – rarely spot the wheels and cogs that surround us when we buy from a well-run company. Perhaps the algorithms really are as good as Galloway says. Maybe the likes of Spotify and Amazon are getting ever better with each click of my mouse. I’m just waiting for that improvement to become apparent and for that Toto song to get out of my head.
Compelling stuff as always. Mark, in your career I suspect you’ve sounded like a wanker for many seconds. It only occurred in this piece though, when you ventured into music criticism. Paul Simon, Boz Scaggs and Toto have all released some great albums. Werewolves of London, however, is shit. Zevon’s self titled 1976 LP is a masterpiece though. The best music algorithm out there is Pandora’s. It’s not available in the UK but it is in Australia, so fill yer boots. In order for Spotify and Amazon’s algorithms to really work for you, one has to do some fine tuning. Rate your purchases and you’ll see the difference. You’re welcome.
Damn blast to buggery- I’ve got Africa in my noodle now too.
Hurry, boy, it’s waiting there for you.
Toto’s Africa is an absolute monster of a tune.
Sorry Ed, I am having none of that! Any song featuring the following lyrics has to be relegated to the dumpster:
“I know that I must do what’s right
As sure as Kilimanjaro rises like Olympus above the Serengeti-i-i-i.”
The prosecution rests.
As you point out, when it comes to music, it’s often not enough to know what you’ve listened to before. The “secret sauce” to music recommendation arguably involves both learning about your preference for different artists and genres and why you’re motivated to listen in the first place.
That’s why our team at Sync Project made a personalized music bot for Slack, that learns from you to recommend the best songs to work, relax and exercise to.
You can learn more about how we are integrating AI techniques with real preference information to find the best songs for a given purpose (https://medium.com/sync-project/work-is-where-music-happens-to-slack-with-love-f173d38372a4) Let us know what you think!
This is seriously cool! will try it tomorrow.
I haven’t seen Galloway’s talk, but from the examples you give it seems like all the winners are web endemic businesses. Aren’t all web endemic businesses algorithm businesses? Since online business is growing at a much faster pace than off-line business is it really logical to say that these guys are all winners because they are algorithm businesses? Seems like a tautology to me. You can put any web-based company in the winners circle and call it an algorithm business. Are the winners winners because of algorithms or because they’re just the most successful web-based companies?
Hi Mark, good article. Building on your comments about the duh! shortcomings of firms who have the data but are using it poorly, here are two pet peeves. Travel sites which offer deals on places we have visited in the recent past (e.g., “Great prices on hotels in towns on the east coast of Iceland !”). And companies who have my loyalty data and don’t react when my purchases decrease significantly over time.
Couple those genius lyrics with some of the most memorable synth work of all time and what do you get?
An absolute smash!
I’m warning you Smallman, any more of this Toto madness and I will have you banned from the MW site forever. You can be as sure as the mighty Kilimanjaro towers over the serengenti-i-i that I am not bluffing.
There is no algorithm for Serendipity.
not yet! Serendipity is most definitely the secret sauce.
£60K on coffee paraphernalia and baby clothes, Mark? No wonder you have uprooted the Amazonian decision trees…
The learning of the AI machines is predicated on patterns
Feeding on our pure news or on our acting more like slatterns?
They learn what they think we like and why they think we choose to pause
They learn how to spot our symptoms but fail to grasp the underlying cause?
Understanding our random quests or our rationale is not within their gift
Cos the answers, my friend, are blowin’ in the wind – tempestuously – if you get my drift?
Hi Mark – the best examples of how algorithms go wrong is to check out Private Eye’s Malgorithms section every issue. It totally exposes the flaws in algorithms at the moment. And they’re funny. See these two examples
there’s a reason for your experience. Let me try to explain it:
First off, algorithms are not a 100% science and sometimes really wrong for two reasons:
#1 – They don’t have all/enough data.
#2 – We, the people, are not rational. By our very nature, we are irrational beings.
The Amazon example is the easiest to explain. Think of products you’d recommend for a letter scale. Then go on Amazon, search for “American Weigh Scales AWS-600-BLK” and check out the recommendations. If you guessed it, congratulations. If not, no problem, the algorithm that is optimising for cross sales has got you covered. Since it sees all data, and Amazon carries all these products, it’s able to make the connection.
Regarding products you’ve bought before, the topic is called “replenishable items”. There are some items you’ll need over and over again because you run out of them, like coffee, pasta or even printer cartridges. Replenishment algorithms try to figure out your personal replenishment cycle and offer you these products at the right time.
But this gets a bit more complex very soon. Example: is paint a replenishable item? Depends. If you’re a dad painting the children’s room in pink, no. If you’re a professional painter, yes. Having more data, for instance about your profession, would allow such algorithms to be more precise. That’s reason #1.
Then there are some items you’d certainly not categorize as replenishable, like a big flat screen TV. Surely people will not buy the same TV twice. Guess what: they do. We’ve carried out several tests showing recommendations of products people have bought and would, in our opinion, never buy a second time, yet they did. Like the TV. The reasoning behind this is: the customer liked the TV so much, he bought one for his uncle, friend, etc., too. That’s reason #2.
When it comes to music, it’s a bit more complicated but essentially works the same way. There are people who actually like Toto. You might not understand that, just like you might not have guessed what the top cross sells are for the letter scale (weed stuff), but they do exist. The AI sees patterns in the data we cannot see. And there seems to be an astonishing number of people that do like Toto.
Music is a bit more difficult to figure out than products, because here we act even more irrationaly than when we go shopping. Example: Tanita Tikaram’s ‘Twist in my Sobriety’ is a great song I’ll always give five stars. But it’s a song I never want to hear except for at my funeral. The signals I send out contradict themselves. Multiply that by a billion potential combinations of signals and listeners and it soon becomes really hard to guess the next song you’d like to hear.
Given all data and complete rationality, algorithms will always be right. They’re pure math. But our world is not. And the least rational of all beings is us. That’s the essence of why AI is not a 100% science.
Disclosure: I do work in the field of AI and deal with personalization algorithms daily.