AI will be good for jobs (and humans)
Everyone is saying that artificial intelligence will destroy jobs. I think they’ve got it wrong.
We invest in AI technology now, to create value in the future. It’s easy to think that this value is created by making something faster or cheaper — automation they call it. Take a job that a person does today, and have a machine do it tomorrow.
That’s hardly ever been the real story.
The massive value comes when a technology makes something new possible, which was effectively not possible before. Something that wasn’t being done at all, or available only to the very few.
An easy example is Google. In the mid-90s we had web search (and Yahoo! directories) for finding websites. But not until Google made search great did we realize that we didn’t need to remember things. As cell phones and wifi became more common and Google search kept getting better, Google replaced our need to remember facts and formulas, to look up stock prices and movie listings in a newspaper, or to program a scientific calculator. OK, maybe Google replaced different use cases for you.
AI is just math (and engineering)
Artificial intelligence is just a technology. As Shivon Zilis writes in her annual Current State of Machine Intelligence, the companies that “get” machine learning, tend to understand that it’s just hard math — these companies tend to include a math Ph.D in their executive suite.
People don’t refer to Google search or to the Facebook news feed as artificial intelligence — because these systems work and so they cease to be viewed as AI. Once you know how it works, it’s just engineering. Deep learning is fundamentally no different.
The AI that will power self-driving cars will not be any different either. You can already see self-driving cars on the roads, so that’s just engineering.
We’ll get back to self-driving cars in a moment. I focus on driver automation since this is the problem on the top of most people’s minds when they talk about AI replacing American jobs. They point to this graph of “truck driver” being the most common profession in many US states:
And then you see something like the night driving demo from drive.ai, a startup which if I’m not mistaking is also working on a “truck driver auto pilot,” as is Otto, acquired by Uber in 2016.
But what actually happened when a new technology replaced work done by humans?
Entry level jobs
Bank teller used to be the entry level white collar job in America. You can still read stories about people rising up from bank teller to bank president, as well as Wall Street traders starting from the mail room. Dilbert creator Scott Adams started out as a bank teller when he moved to California after graduating from college with no plan.
When the ATMs started popping up in Long Island staring in the early 1970s, many lamented that the bank teller job would go away. In fact — the number of bank tellers increased, as ATMs became commonplace. I wasn’t around for that transition. We had no ATMs (and not too many bank tellers either) in communist Russia. But I’d imagine that people did a whole lot more banking, more branches opened up, and the tellers employed at these branches ended up doing more valuable work than counting banknotes before handing them over to customers. Customers in turn could get those banknotes at all kinds of odd hours, until the ATM ran out and had to be re-stocked from the vault or the Brink’s truck.
I’d assumed that the total number of bank tellers would eventually start to go down over the decades— as ATMs began to accept checks, to use OCR to verify the deposit amounts written on those checks, and as most of us moved to direct deposit and internet banking altogether.
Surprising, the number of U.S. bank teller jobs was still rising through 2007, until it finally dropped off after the recession in 2008. It might even have come up in the years since.
Of course one could argue that even more tellers would be employed in the US if the ATM was never invented, or banned, or somehow taxed and required to work 40-hour weeks like a human. We know that this is impossible. The ATM created value. It replaced some human tasks, yes, but it opened up others. No modern bank could compete by rejecting the ATM, and that was clear by the 1980s.
Some jobs do go away, of course. There are no more telephone operators, no more typing pools, and no human-powered spreadsheets.
However, are more people or fewer total people employed by accounting firms and telecoms? Automation in telephone routing and spreadsheets made the basic operations a lot more accessible. The volume went up, it generated tremendous surplus, and some of that surplus went into new jobs in those sectors.
Back to the 🚘
I recently rode in an Uber self-driving cab while visiting Pittsburgh — fittingly enough for the #BrainsVsAI poker challenge, which the 🤖 won over 120,000 poker hands of heads-up No Limit Hold’em.
This custom Volvo will eventually drive itself, with no human other than me in the car. But for now… there were two Uber employees in the front seat: the driver with his hands on the wheel, read to take over, and an engineer running diagnostics as the car made turns, obeyed traffic lights, and stayed in the lane.
Soon, Uber will remove the engineer. Eventually, the driver. But for now the self-driving cab requires more jobs than a conventional cab. Given the story of the bank tellers, I’m not ready to accept that dramatically fewer people will be employed in the professional driving business overnight.
Where to, Brute?
What will people do in those new jobs in support of self driving car fleets? Some of them will monitor cars remotely, getting them out of situations where the AI gets confused. The car will likely drive very conservatively, stopping and pulling over as soon as something looks amiss. Expect this to happen often in the early stages, and it will always happen with some frequency for as long as I can imagine. Other people will service the cars, as it will become possible and hopefully expected for the cars will be cleaned between every ride. Unlike sadly my seat pocket on JetBlue.
Others will drive emergency vehicles, picking up customers whose cars break down or get confused. Don’t expect the self driving cab to be able to figure out every cobblestone street or confusing local neighborhood.
This almost begs the question: why have self-driving cars then?
In a sense, it doesn’t matter. Google and Uber (and Ford and Volvo and Tesla) will spend billions on this technology, with or without much benefit. Or it may end up like the optical fiber rush of the 1990s — leading to the fast internet connections that we enjoy today, while bankrupting the companies that built them.
Self-driving cabs will increase peak carrying capacity. Right now it’s pretty easy to get an Uber at most times and in most places, but very hard to get one at some times and not at all is some places. Meanwhile the cost of an unscheduled Uber car is not that different from one that’s in service (just add gas).
If a central operator can oversee up to five cabs, and even if those cabs drive slowly, stop often, and can’t drive well at night — you’re still able to deploy a lot more cabs at peak hours, as well as to pick up folks en masse from the ⚾ game. At least from a day game when it’s not raining.
With the self driving car on its way, it will finally make sense to invest into automatic pickup — do you really need to order and wait for a specific car at the airport? We’ll have better location services, you’ll use your phone to log in to the next car that’s ready, and we’ll have excellent communication features for when the car can’t find your pickup location. All of these improvements would benefit the existing Uber/Lyft human-based model as well, but self-driving cars will make them more necessary and more valuable, therefore making the upgrades more likely to happen.
All of this work will require investment, and that will create jobs.
Meanwhile despite the companies’ massive investment into self driving technology, the decreased unit cost will make car sharing more accessible. Just as technology made telephony and banking more accessible, mass use leads to employment growth in those sectors along the way.
Known unknowns
I don’t know exactly which jobs AI will make possible. That’s why the “AI will be bad for workers” crowd appears to be winning. They have a simple message. Then again we’ve never been particularly good at predicting the future, especially the future of the economy.
After WWII, many economists expected mass unemployment, as millions of Americans would return to the work force:
Paul Samuelson, a future Nobel Prize winner, wrote in 1943 that upon cessation of hostilities and demobilization “some ten million men will be thrown on the labor market.” He warned that unless wartime controls were extended there would be “the greatest period of unemployment and industrial dislocation which any economy has ever faced.”
Instead, the economy adjusted, and most of the GIs found good jobs. The 1950s are now remembered as a time of great prosperity.
In the 1990s, Israel absorbed 1M+ ex-Soviet immigrants (into a country of just 6 million people). There were some adjustments, but no ultimate collapse.
Human computer cooperation
Just as there will be a pot of gold for the quants who automate ways of combining stock news and rumors with technical analysis, there will be value created by those who invent ways to combine human work with what intelligence machines do best.
Yes, for now a lot of that non-coding collaboration has focused on building human teams for data labeling. Not necessarily the most fun job, but very important work that the AI engineers and researchers benefit from immediately. As your company is spending millions (or at least hundreds of thousands) on AI researchers and GPU machines, it becomes inexcusable not to spend tens of thousands on building good datasets.
One of our super-powers at Twitter Cortex was a great in-house data curation team — who are now curating data at Clarifai.
I’m not suggesting that every non-tech graduate’s future lies in labeling data to feed to the machines, or in writing articles to be fed into a giant LSTM for hedge funds to predict the news. Then again workers born in the late 19th century did not anticipate moving to Detroit and spending eight hours a day working on Henry Ford’s assembly lines. Those were considered great jobs.
New jobs will be enabled by automation and AI, probably jobs that effectively don’t exist yet. (Who was a social media specialist, or for that matter a blogger, a mere ten years ago?) These new jobs will be enabled by the surplus value that technology creates, much of it AI technology. We’ve seen this story before.
P.S.
So what should a young person study to prepare themselves for this robot-enhanced future?
Systems over goals, as Scott Adams says. Don’t aim for a goal of getting a specific job. Instead, learn skills that will increase your overall “talent stack.”
Human-computer interaction is a good skill. Not just Word, Excel and Photoshop, but it’s great to have someone on your team who can lay out a Wordpress page, hook up Disqus, and figure out how to embed Powerpoint presentations into your company’s Confluence.
Writing, and general communication skills are always in short supply, and helpful in any role. Again from Scott Adams, here’s a excellent post about becoming a better writer.
Measurement, Detection and Classification
New processes create friction. None more so than a new computer-aided process. People will always be needed to evaluate how the new process is doing, and to detect and fix the mistakes. The most valuable people will also classify those mistakes, allowing engineers (or the AI itself) to identify and work on the common errors that it’s making over and over.
How do you make a computer process better? Yes, some computer skills might be useful. Regular expression, for example. I’ve heard that the best tech recruiters have becomes wizards at searching resumes on LinkedIn with complicated search queries. Similarly, if you’re asked to label thousands of examples of text data, a bit of regular expression work and some sorting can save a lot of time.
Mostly though, if you can take notes and organize every case that came up for human review, this “data exhaust” can be used to make the AI more powerful in the future. According to leading machine learning VC Matt Turck, data exhaust will be a leading form of monetization for many startups. Buying and selling anonymized medical data is already a huge business.
Put another way, with a few exceptions derived from processes like physics simulation, today’s AI systems are trained on data that was at some point generated by humans. The more that the humans can record human reasoning, at some point the AI can take advantage of that.
The dream is unsupervised data at scale + supervised labels for what really matters:
Labeled data are the fuel for today’s machine learning. Collecting data is easy, but scalably labeling that data is hard. It’s only feasible to generate labels for important problems where the reward is worth the effort, like machine translation, speech recognition, or self-driving.
Training domain-specific AIs for every problem will take some time, and the machines will never be able to do everything that a human can learn to do, often pretty quickly.
We’ll always need machines to help enhance human performance, and vice versa.
The future of jobs is bright in an increasingly AI world, in my opinion, for this and other reasons. They won’t be the same jobs as we have now, and that concerns some people. But how many of us do what our grandparents used to do, or would want to?
