This post also appears on Forbes.com.
There have been and will continue to be multiple big technology revolutions, but the most impactful on human society may be the one that finally builds systems with judgment and decision-making capability more sophisticated and nuanced than trained human judgment. Machine learning, sometimes called big data or artificial intelligence, is making rapid progress in complex decision-making (for instance: driving a car was thought to be too difficult for computers even five years ago). Without speculating on what is probable, it is at least possible that such systems may even be better at creativity, emotion and empathy than human beings (for instance: writing the best music, love story or creative fiction). At the very least these systems may be able to handle much more data to which we now have access and use it to make better judgments than humans with their supposed instinct, gut, holistic and integrative decision capability. Although any one software program may not do everything a human brain can do, specialized programs will likely make decisions and predictions in their domain better than most trained humans. Many, if not most, domains will be well covered by such programs. Many problems in our work environments aren’t ones the human brain evolved to solve for in the African savannah. To achieve these goals, a machine learning system does not need to exactly replicate the brain or even use brain like techniques.
While the future is promising and this technology revolution may result in dramatically increasing productivity and abundance, the process of getting there raises all sorts of questions about the changing nature of work and the likely increase in income disparity. With less need for human labor and judgment, labor will be devalued relative to capital and even more so relative to ideas and machine learning technology. In an era of abundance and increasing income disparity, we may need a version of capitalism that is focused on more than just efficient production and also places greater prioritization on the less desirable side effects of capitalism.
Let’s look at the scale of change that the new machine learning and data revolution may bring and why it potentially could be different than prior technology revolutions like mobile phones, accessible computing and automobiles. Just in the Khosla Ventures portfolio alone, entrepreneurs already are trying to use machine learning technologies to replace human judgment in many areas including farm workers, warehouse workers, hamburger flippers, legal researchers, financial investment intermediaries, some areas of a cardiologist’s functions, ear-nose-throat (ENT) specialists, psychiatrists and many others. Efficiency in the business world generally means reducing costs, which results in using fewer well-paid but highly skilled minds and the technology they develop or capital to replace lower paid and less skilled workers.
Our portfolio represents only a tiny fraction of the efforts around machine learning. Consider replacing taxi drivers (Google’s driverless cars), IT administrators (Grok on Amazon Web Services) and even hedge fund traders. Renaissance Capital, one of the top performing hedge funds that has consistently topped the Standard & Poor's 500-stock index, does not hire traditional Wall Street talent like analysts but instead uses machine intelligence. “The firm's scientists tap decades of diverse data in Renaissance's vast computer banks to assess statistical probabilities for the direction of securities prices in any given market.” Another machine learning system even performs difficult jobs like scheduling night workers for the Hong Kong subway system, the busiest and most efficient in the world. These are not just traditional low skilled jobs susceptible to replacement.
In past economic history, each technology revolution—while replacing some jobs—has created more new types of job opportunities and productivity improvements, but this time could be different. Economic theory is largely based on an extrapolation of the past rather than causality, but if basic drivers of job creation change then outcomes may be different. Historically, technology augmented and amplified human capability, which increased the productivity of human labor. Education was one method for humans to leverage technology as it evolved and improved. However, if machine learning technologies become superior in both intelligence and the knowledge relevant to a particular job, human employees may be rendered unnecessary or in the very least, they will be in far less demand and command lower pay.
Machines with unlimited and rapidly expanding human-like capabilities may mean there will no longer be as much need to leverage human capabilities. In fact, there may be little for humans to augment or amplify even as productivity per human hour of labor increases dramatically all while far fewer people are needed for most tasks. This is not to say all human functions will be replaced but rather that many, and maybe even a majority, may not be needed.
It is possible that machine learning technologies in the next 50 years will create a greater abundance of goods and services than we could imagine. Initially, machine intelligence will exceed human judgment in a few narrow areas and then, more broadly over time, will increase traditional measures of productivity and increase economic growth over where it might otherwise have been. In my view, capitalism is very good at promoting efficiency but now has moved to demand generation, making us want things we did not know we wanted. I suspect this trend will persist and the demand for goods and services will continue on an upward trajectory.
Many like Steve Rattner and Marc Andreessen have written on the subject of technology and proposed arguments ranging from Luddism and the “lump-of-labor” fallacy to economist Milton Friedman’s take that human wants and needs are infinite. I suspect they are right but that does not mean we will not see increasing income disparity with the next machine learning based technology revolution. Others like Erik Brynjolfsson are more contemplative but still miss the difference between past technology revolutions and machine learning technology.
The traditional view is that historically over time as jobs have been displaced, new ones have been created and to think otherwise is a Luddite fallacy. Steve Rattner argues that technology comes down to the concept of producing more with fewer workers or becoming more efficient (what economists call “productivity”). Without higher productivity, wages and standards of living cannot go up. He goes on to state that as technology has changed the nature of work—more specialized training is now required for many jobs—and consequently, it has contributed to a sharp rise in income inequality. We should be embracing technology not fearing it and that means educating and training Americans to perform more skilled jobs. He agrees that not every worker can be retrained, and so we must help those who aren’t suitable for new jobs with more robust social welfare programs, but he seems to treat it as a minor, not major problem.
What if machines, which may soon exceed the capability of human judgment, do most jobs better than humans even if people receive additional training? The magnitude of the problem of displaced workers and increasing income disparity especially in the face of abundance (increasing GDP) may become substantially larger. It is possible that this particular technology revolution does not allow for human augmentation and amplification by technology to a large enough degree and that education and retraining are not solutions at all, except for a very small percentage of the workforce. As Karl Marx said, “when the train of history hits a curve, the intellectuals fall off”. Extrapolation of our past experiences, a favorite technique of economists, may not be a valid predictor of the future—the historical correlation may be broken by a new causality. Efforts at estimating the number of jobs that are susceptible to computerization underestimate how technology may evolve and make assumptions that seem very likely to be false, similar to past “truths” (like the waning correlation between productivity and income growth for labor). Even with this underestimate, researchers concluded that of the 702 job functions studied, 47-percent are at risk of being automated.
I am not advocating for slowing technological change to preserve jobs but rather worry that the machine learning technology revolution will lead to increasing income disparity, and disparity beyond a certain point will lead to social unrest. I grew up a fan of some inequality (read “incentives to work harder”) but lots of social mobility. I suspect that if and when software systems exceed the capability of the median, and eventually the best humans in judgment and skill, an avenue of personal growth through education that previously has always been open for labor advancement may be closed (note that when I say “software systems”, I do not mean robots alone as I believe there is a large and non-linear difference between general productivity technologies and technology that surpasses human judgment capability, although other revolutions like robotics, mobile technology, computing power and others all enhance and leverage machine learning).
It is happening in every type and level of job from farm work to radiology, warehouse work to legal research and taxi drivers to medical specialties like robotic surgery and medical diagnosis. Some new jobs will be created but given the need to exceed the capability of intelligent, fast learning and ever evolving software systems, the typical human will not be able to keep up in most tasks even if the overall supply of cheaper goods and services is increasing and GDP growth is healthy. Even with access to better education and skills, not enough humans could adapt quickly enough to outperform intelligent software systems. It seems likely that humans will lose this “race against the machine” in many, if not most, work domains causing a large shift in employment much like the transition away from an agrarian economy in the early 20th century. First, we lost the physical labor battle to engines, and now, we may lose the mental labor battle. What else may humans offer? Creativity is an option, but machines may even get better than humans at that, too.
For instance, David Cope is stretching machinecreativity by inspiring questions as to whether or not great new music can emerge from the extraction and recombination of patterns in earlier music using computers. Will the deepest of human emotions be triggered by computer patterns of notes? According to Wired magazine, “The result is astounding to even the casual listener—rife with emotional complexity and deep textures that belie the music’s artificial origin.” Will these kinds of developments replace human capability or enhance human experience? Both the argument that there will be new kinds of jobs and new creativity unleashed rings true. Before the film camera, there was no job for a film producer. Entire industries have erupted. For instance, entertainment has become more popular, and extreme sports has turned into an income generating profession for many. Snowboarding, which before was not a profession, now is. Etsy and EBay have facilitated global artisans and entrepreneurs. New technology will most likely enable an entirely new world of professions. For instance, Wattpad has enabled a bevy of new creative writers. Others, like Pinterest and Tumblr, have provided people with an outlet for their creativity and allowed them to be more expressive about their tastes and individuality.
Will new jobs be based on human intelligence or creativity? It seems likely that the top 10 to 20-percent of any profession, be they computer programmers, civil engineers, musicians, athletes or artists, will continue to do well. What happens to the bottom 20-percent or even 80-percent, if that is the delineation? Will the bottom 80-percent be able to compete effectively against computer systems that are superior to human intelligence?
Productivity may increase average incomes while reducing both median incomes and the Gini coefficient (a measure of statistical dispersion intended to represent income distribution). Ironically, citizens in developed countries like the U.S. could conceivably have a higher standard of living even though income disparity may widen. Many economists argue for increasing access to education and skill development (which itself will be delivered via technology), but will that be sufficient to stem the growing income disparity? The traditional recipe for education and retraining may become far less relevant in the face of rapidly increasing capabilities of machines, even in areas that were previously considered human strengths like integrative judgment, knowledge, complexity and creativity.
Marc Andreessen argues that the subtext to the argument that “this time is different” is that “there won’t be new ideas, fields, industries, businesses and jobs”. He compels us to consider a thought experiment, “Imagine a world in which all material needs are provided for free, by robots and material synthesizers.” I believe the subtext of this assumption is wrong; if we wait long enough, the utopian world for which we strive is in fact, possible—that’s the technology optimist in me.
He goes on to hypothesize:
Housing, energy, health care, food, and transportation – they’re all delivered to everyone for free by machines. Zero jobs in those fields remain…. What would be the key characteristics of that world, and what would it be like to live in it? For starters, it’s a consumer utopia. Everyone enjoys a standard of living that kings and popes could have only dreamed of.
In fact, at two-percent per capita income growth for a hundred years, someone making $40,000 annually today will, in the future, make real income in today’s purchasing power of almost $300,000 (assuming that the cost of goods and services stays the same). I suspect cost of living at a certain standard in our future utopian society will further decline, thus buying substantially more for the individual who today earns $40,000 annually than someone making $300,000 annually today can buy. Happily, technology will be even more deflationary for goods and services than outsourcing to China has been over the last decade or two.
Going even further, Andreessen writes,
Since our basic needs are taken care of, all human time, labor, energy, ambition, and goals reorient to the intangibles: the big questions, the deep needs. Human nature expresses itself fully, for the first time in history. Without physical need constraints, we will be whoever we want to be.
I agree that the main fields of human endeavor may be culture, arts, sciences, creativity, philosophy, experimentation, exploration and adventure. The real question is whether or not everyone will be able to participate. Will everyone be able to keep up on the income curve and leverage new technologies equally if the rules stay as they currently are?
This new quantum jump in computer capabilities will likely lead to increasing income disparity and abundance at the same time. It is possible that this time the technology evolution really is different, because for the first time, it is not about productivity enhancement but rather exceeding human intelligence. If this scenario comes to fruition, we will need to make structural changes in our social and political systems to optimize for fairness or whatever we determine are our society’s goals (that is, if we can agree on goals at all).
So, how do we address the issue of income disparity? The easy answer seems to be what Thomas Piketty has advocated: some form of income redistribution. I suspect it will be a necessary component, but it should be the last resort. Our capitalist system is easily biased with some arbitrary policies in favor of labor or capital or new ideas as is common in the innovation and entrepreneurial economy. Giving a research and development (R&D) tax credit back to companies will favor R&D and innovation, whereas giving favorable depreciation is a bias towards capital instead of labor. Keep in mind that larger incumbents tend to shape more of the rules and regulation, at least in the US.
Requiring corporate healthcare makes labor more expensive and disadvantaged relative to countries, like many in Europe, that provide centralized healthcare. The cost of labor or the cost of capital can be effectively altered by simple changes in rules, regulations and laws; many of these biases have been engineered into today’s seemingly neutral capitalist economy. More and significant manipulation will be needed to achieve reasonable income disparity goals. Income or social mobility is an even harder goal to engineer into society’s “rules”. I suspect the situation will become even more complex as traditional economic arguments of labor versus capital are upended by a new factor many economists don't adequately credit—the economy of ideas driven by entrepreneurial energy and knowledge. This last factor may become a more important driver of the economy than either labor or capital.
These are mere speculations and the future is nearly impossible to predict. Hence a word of caution is necessary in recommending any specific solutions or premature action at a national scale that may be drastic or irreversible. Debate and discussion are definitely called for. Point solutions for those hurt by the increasing income disparity need to be found, and the well off should be willing to tilt the playing field somewhat towards those who are not as fortunate. Local disparities (like housing in San Francisco as technology firms grow rapidly and non-technology workers get crowded out) may need temporary solutions, but structural changes at the national level will probably be necessary over the long term in order to solve the larger side effects of technology exceeding human capability. Economic policy will need to include not just economic growth tuning as the U.S. Federal Reserve does today but also be driven by disparity and social mobility biases.
As an unapologetic capitalist and technology optimist, I will argue for the continued rapid support and deployment of machine learning systems. Let’s not slow down the hand of the market or technological progress but rather realize human labor may be devalued in many instances putting downward pressure on wages of lower-skilled workers and even many higher-skilled workers. These represent just some of the looming challenges as we continue on the path towards technological progress. We must be thoughtful about the society we live in and the future we create.