In August of 2021, Elon Musk revealed the Tesla Bot at the company’s “AI Day” event. At the time the announcement was so unexpected that some journalists, pundits, and analysts believed it to be a joke. But Musk subsequently clarified not only that the humanoid robot is a real product in development, but that the Tesla Bot, nicknamed Optimus, “[has] the potential to be more significant than the vehicle business.”
On a recent earnings call, Musk said, “I was surprised that people did not realize the magnitude of the Optimus robot program … The importance of Optimus will become apparent in the coming years. Those who are insightful or looking, listening carefully, will understand that Optimus will ultimately be worth more than the car business, worth more than [autonomous driving].”
I anticipated this move in an article I wrote five years ago in 2017, in which I explained why autonomous vehicles are the pebble that will trigger the automation avalanche. Given its relevance, RethinkX is republishing an edited version of this piece.
[The article below was originally published on www.adamdorr.com in 2017]
The prospect of self-driving autonomous vehicles (AVs), long regarded as a futurist fantasy, is now being taken seriously by industries, investors, legislators, and other decision-makers, as well by the media and public at large. The world is paying close attention as AVs near market readiness, and analysts anticipate a combination of benefits and detriments from their widespread adoption. On the one hand AVs will lower costs, improve safety, increase access, and reduce the overall environmental footprint of transportation. That is unequivocally good news. On the other hand, however, millions of people around the world who currently earn their living by driving are going lose their jobs to these machines over the next decade or so. Indeed, the looming prospect of large-scale technological unemployment resulting from the automation of one of the most popular jobs on the planet is perhaps the single greatest concern surrounding AVs.
But we aren’t nearly concerned enough.
AVs aren’t just coming for drivers. They’re coming for everyone. At least one billion jobs are in the crosshairs across thousands of industries. Whoever owns the dominant AV technology is going to own the future of those other occupations as well. AVs are the ultimate “killer app” – the job killer app. Yet as we will see, hidden within this prospect is an unprecedented opportunity. Removing labor costs as the main limiting factor in production will unleash the potential for new levels of prosperity that could completely transform the economy.
Concerns about the looming specter of mass technological unemployment from automation are now being voiced from all quarters. It isn’t just technology innovators like Elon Musk and Bill Gates who’ve expressed concerns. It is also economists like Erik Brynjolfsson, scientists like Stephen Hawking, philosophers like Nick Bostrom, public officials like Colorado Governor John Hickenlooper, and investors like Warren Buffet. Some observers are optimistic, others more pessimistic, but few if any have yet to recognize the direct connection between AVs and the tidal wave of technological unemployment to which they will give rise. As a result, there is essentially zero public awareness that large-scale disruption of the global labor market by automation will be well underway within a decade.
Technological unemployment is as old as the Industrial Revolution itself, but in the past jobs were lost to mechanization which is not the same as automation. Mechanization shifts monotonous physical drudgery from muscles to machinery, the results of which have been admittedly spectacular: one person with a bulldozer can move more earth in an hour than a hundred people with shovels can move in a month. But most labor requires more than just mindless toil. It requires decisions in response to a complex and changing physical environment, and those decisions in turn require recognition of and purposeful interaction with hundreds or even thousands of different objects. In other words, physical labor also requires intelligence – lots of it. And machines today are not particularly smart. But that is about to change in a very big way.
AVs are an exemplar of narrow artificial intelligence, which is distinct from the general intelligence that humans possess. Narrow AI, as the term suggests, refers to machines that possess competence in only one very specific domain. By contrast, truly general intelligence capable of adapting to any new task or situation probably requires self-awareness and perhaps even the kind of subjective conscious experience that humans and other mammals have, and we are still a long way from that in computers. The threat that narrow AI represents to employment, however, is very real – and it is far closer than most of us realize because of AVs.
Modeling and Navigating
To understand why AVs threaten to automate far more than just driving, we first need to appreciate that AV technology is software that performs a complex set of tasks, and that the same software can be adapted to other seemingly dissimilar sets of tasks with surprising ease.
The tasks that AV software performs can be divided into two broad functions: modeling and navigating. First, the software builds an accurate model of the dynamic three-dimensional environment around itself from sensory input, just as human drivers do. Cameras, lidar, radar, sonar, and microphones gather huge quantities of data, which are then analyzed by a powerful onboard computer. The result is a detailed picture of the world surrounding the vehicle that is updated in real time. Next, with this picture in hand, the vehicle navigates within this dynamic environment – meaning that it recognizes other objects around itself and makes decisions about how to interact with them.
So, AVs build a detailed picture of the world around themselves and then navigate through it by recognizing and deciding how to interact with other objects. Sound familiar? It certainly should. Whether it is washing dishes in a restaurant kitchen, stocking items on a supermarket shelf, harvesting potatoes in a field, or assembling a smart phone on a production line, performing physical labor entails modeling the physical environment from sensory data and then navigating through that environment by recognizing and interacting purposefully with the other objects within it.
These feats of modeling and navigating – so effortless for humans and other animals – have proven devilishly difficult for robots to master. But thanks to key breakthroughs in the mathematics and computer science of machine learning, together with astounding growth in the power of computing hardware, these challenges have now been met. Waymo, the AV company started by Google, already has a fleet of cars operating on public streets with no human behind the wheel, and nearly all major automotive manufacturers have plans to have AVs on the market in the 2020s. The technology is science fiction no more.
If we grant that the work of driving is fundamentally similar to other types of physical labor, the question that logically follows is: how quickly can the underlying AV technology be adapted to a new task? It is impossible to know for sure, but we are starting to see strong evidence that the adaptation process will happen extremely rapidly – in some instances, quite literally overnight.
Lessons from Machine Learning
Two recent cases offer an instructive window into the world of machine learning. The first is robots that are able to walk. Walking robots aren’t breaking news, but what is extraordinary is that they are now able to learn to walk in a natural way from scratch in a matter of hours. This feat is achieved through the techniques of reinforcement learning and imitation learning, which are specific applications of the more general deep learning process utilized by deep neural networks.
These systems set a goal for the machine and then allow it to learn in a human-like way from its own attempts in both the real world and in simulated reality – just as a person might learn by both trial-and-error and by running through different possibilities in his or her imagination. What is perhaps most instructive for our purposes here is that these deep learning systems have robots up and walking in hours, no matter whether they have two legs, or four, or more. Companies like Boston Robotics are leaders in this area of applied deep learning technology.
The second instructive case is the spectacular example of AlphaZero, a program created by DeepMind – another member of the Google family held by Alphabet Inc. In 2016, DeepMind’s program AlphaGo beat Lee Sedol, the world champion Go player, to much fanfare. Observers had predicted that defeating humans would not be possible for years, if ever, given the open-ended nature of the game. An updated version of the program called AlphaGo Zero was developed in 2017 that is able to crush the original AlphaGo. AlphaZero is a generalized version of that successor program which, to almost everyone’s astonishment, was able to teach itself chess from scratch without assistance in just four hours. Within nine hours it was able to defeat the world’s best chess program. Not the best human player, the best chess program.
To put this astounding accomplishment into context, keep in mind that thousands of brilliant human players and computer programmers have spent more than forty years competitively designing chess software. (The first machine to defeat the best human player of its day – Gary Kasparov – was IBM’s Deep Blue in 1997, and the program that AlphaZero defeated, called Stockfish 8, is a much stronger player than Deep Blue was). Now to be fair, the hardware that AlphaZero and Stockfish 8 were running on was not identical, so a true apples-to-apples comparison between the two is not yet available. But the relevant point for our discussion here is that DeepMind was able to adapt AlphaZero to a new purpose extremely quickly.
The case of AlphaZero may also hold an additional lesson for the future trajectory of automation technology: hard-won solutions to very difficult machine learning problems can sometimes be surprisingly easy to generalize to other similar but simpler problems. For DeepMind, solving the challenge of Go took years and tens of millions of dollars. With that solution in hand, solving the much simpler challenge of chess only took nine hours and a few dollars’ worth of electricity. For AVs, solving driving is Go, and every other physical job is chess.
A Light Bulb Moment
If the analogy of Go and chess doesn’t resonate well enough, consider another one instead: the light bulb. It took years of research effort among competing firms before Thomas Edison’s company found a commercially viable way to produce light with electricity. The success of the light bulb then threw open the floodgates to thousands of other applications of electricity, from electric fans and refrigerators to washing machines and computers. Our world now revolves around electronics, and Edison’s firm went on to become General Electric which more than a century later remains one of the largest companies in the world – and is still a leading manufacturer of light bulbs. In the same way, AVs are the ‘killer app’ that will open the door to thousands of other applications – not of electricity, but of narrow artificial intelligence.
Winner Take All?
Competition among firms to win the AV technology arms race is fierce, as it has been with countless other technologies since the dawn of the Industrial Revolution. But what very few observers realize is that the form of narrow artificial intelligence that AVs represent will be quickly generalizable to thousands of other applications beyond vehicles – in some cases, literally overnight. The nature of machine learning means that whatever company wins the AV arms race stands to capture (and undermine) a substantial portion of the world’s labor market by around 2030 – the value of which is at least $20 trillion, or around one quarter of the global economy.
If that didn’t properly shock you, let me state it again for clarity: whoever wins the AV technology race will be poised to take ownership of up to one quarter of the global economy by the 2030s. If you are an institutional investor or policymaker, now would be a very good time to take a close look at the AV leaderboard and dial all of your concerns about technological unemployment up to eleven.
Lessons from History
We can look to previous examples of software-driven disruption for insight. In the 1980s, Microsoft and Apple overtook incumbent giants like IBM and Xerox to capture the lion’s share of the operating system market for personal computers, and with it control over much of the surrounding technology ecosystem. In the 2000s, Apple and Google overtook incumbent giants like Nokia and Blackberry to capture over 90% of the operating system market for mobile phones, which they very quickly leveraged into control over most of the surrounding technology ecosystem. Today, one or two companies like Google or Tesla are similarly poised to overtake incumbent automakers like GM and Toyota to capture the operating system market for Avs. Therefore, they too may quickly come to control the majority of the surrounding technology ecosystem.
But lessons from these past cases can only take us so far, because the technology ecosystem surrounding AVs will encompass the bulk of physical labor as we know it. The stakes are therefore so titanic that the case of AVs is likely to break all the conventional rules in business, economics, and policymaking.
An Unprecedented Opportunity
AVs represent the cusp of an unemployment tsunami. Even if a monopoly or duopoly on AVs is prevented by regulation or market competition, the automation genie will still be out of the bottle. It took an IBM supercomputer to beat the best human chess player in 1997, but within ten years dozens of different chess programs could manage the same feat running on a home computer. Robots will still take all the same jobs, whether they are powered by the software of one, two, or a thousand different companies.
On the one hand, this will be one of the most extraordinary challenges human civilization has ever had to deal with, as a chorus of observers are now cautioning. But on the other hand, it also presents humanity with an unprecedented opportunity.
Labor has always been the limiting factor of production. In a world where labor is abundant and nearly costless, so is everything else. In principle, giving over all human toil to machines therefore fulfills one of the basic promises of technology itself: the work still gets done, the raw materials still get harvested and mined, the goods still get manufactured and packaged and shipped – but no person has to suffer any drudgery along the way.
There is much to love about this long-cherished sci-fi vision of the future. Radical automation represents a giant leap forward on the pathway toward providing material quality of life and an abundance of opportunity to all the world’s people. But actually realizing the promise of a more prosperous, equitable, and sustainable world with the aid of technological marvels will be no simple matter, and of course will not be achievable with technology alone. While technology will provide the means, the next question is how do we distribute the extraordinary benefits of a world in which labor is virtually free. Our current institutions are built to govern scarcity, not abundance. We must therefore get ready to rethink all of our social, economic, and political apparatus from first principles.
This is a thrilling moment at a key crossroads in human history. We must choose whether to use the marvels of technology to accomplish wonders and build a world that is truly equitable and sustainable, or whether to allow our current system to continue deepening disparities of income, wealth, and opportunity while running roughshod over the natural environment.