Economic and Financial Effects of Advanced AI
By advanced AI we mean human-level AI and smarter-than-human AI that may be developed sometime in the next 30 years or so.
Human physical labor
We will consider two possibilities. First imagine a world in which there are no advances in robotics. The demand for physical labor will likely be little effected by advanced AI. Advanced AI could increase the demand for human physical labor; think third world factories making electronics. Or it could reduce it, by designing less labor intensify ways to manufacture the same products. In either case, physical labor will likely make up an increasingly small slice of the overall economic pie. Now imagine the more likely scenario in which advances in robotics go hand in hand with advances in AI. In this case not only would the demand for human physical labor relative to the overall economy be reduced, but its absolute size would also decline.
Human intellectual labor
If human-level AI is obtainable at a reasonable price, the demand for human intellectual labor is likely to all but disappear. Even if we knew how to design human-level AI, using today's technology the cost to do so would be prohibitive. It is projected that if computing costs continue to fall, assuming that by then we know how to design human-level AI, human level intellectual performance might be available for $0.70/hr in 2040 (see Prediction). This is based on looking at the human brain, and estimating how expensive it would be to perform an equivalent number of computations to that of the neurons in the human brain.
Gross world product
The reduced need for human labor will cause a relative drop in consumer consumption. I don't have a good argument for why this won't lead to a permanent recession, but assuming it doesn't a greater portion of economic returns will flow to capital, leading to an increase in business capital investment, which will in turn increase the rate of growth of gross world product.
Gross domestic product
Just because the gross world product grows significantly, doesn't mean individual country gross domestic products will all grow. Just as there are winners and losers in the current world economy, there will be winners and losers in the future, and the past need not determine the future. In particular it will depend on the nature of the capital resources possessed by each country.
Human-level AI and smarter-than-human AI might change the world, but still might not be a great investment. To understand this, a little history lesson regarding past investments in technology is in order. Note, AI is part of information technology.
Technology and the past
The NASDAQ Composite index was created in mid-1971 and has historically been made up of technology companies, and especially information technology companies. In 1971 Intel had just released the first microprocessor, and a few years later personal computers would be developed. We consider the NASDAQ Composite on account of its long history. From 1972 to 2015 the NASDAQ Composite returned an geometric mean of 4.7% per annum after inflation. The NASDAQ Composite does not include dividends, but by comparing the returns of the NASDAQ Composite Index (IXIC) to the returns of the NASDAQ Composite Total Return Index (XCMP), which has existed since 2003-09-24, dividends though 2016-09-09 averaged 1.0% per annum. So the total return from investing in technology from 1972 through 2015 is about 5.7%. For comparison the much broader S&P 500 stock index has returned 6.0% after inflation over the same period. Making this significantly worse, markets price volatility, and the annual volatility of the NASDAQ Composite has been significantly worse (25.6%) than that of the S&P 500 (17.2%). Thus technology has been a poor investment.
With only 44 years of historical data, and a variance of 25.6%, the standard error of the mean return for technology is 3.9%. Thus we are a fair way from being statistically confident that technology is a worse investment than the broader market, but so far it doesn't look good. The mean arithmetic return for the NASDAQ Composite would need to be 2.1% higher than it was for it to obtain the same Sharpe ratio as the S&P 500. Conventional financial theory holds that there are no investments that can be expected to have ahead of time superior expected returns on a risk adjusted basis, and the case of investing in technology tends to bear this out.
Technology clearly plays a much greater role today than it did 44 years ago. The market segment weighting of the information technology sector in the S&P 500 at the end of 2015 is a little more than double what it was at the end of 1979, which is the earliest date for which segment weighting data could be found. So clearly information technology has been booming (along with financials and healthcare). Most of the gains beyond those of the overall market however appear to have been picked up by new entrants. The considerable wealth of the founders of Microsoft, Apple, Google, and Facebook all come at the expense of regular investors.
A quick review of the global information technology sector shows the 11 largest companies account for just over half the value of the 876 public companies making up the sector. 9 of the 11 largest information technology companies didn't exist in 1972 (IBM and Intel are the exceptions). Without these 11 companies the information technology sector would be smaller today than it was in 1972. Thus to reliably capture the value of information technology it was necessary to invest in hundreds of companies. This can not reasonably be done through individual stock picking, and instead requires the use of mutual funds.
AI and the future
With advanced AI, what happened to investments in technology over the last 44 years seems ripe to repeat. Advanced AI is unlikely to require large amounts of capital equipment to develop. Instead what is needed is a few new ideas. These new ideas will probably tend to be either broadly shared, rather than the property of any single company. And even when they are the property of a single company history suggests they will be the property of a newly formed company rather than an existing one.
What happens to capital depends heavily on how advanced AI develops. If advanced AI is developed by a single entity, then that entity will reap the gains associated with the more efficient deployment of capital resources that it enables. If however there are multiple competing AI players, existing capital will be in the bargaining seat and be able to reap most of the rewards. On balance the latter scenario seems more likely. Self driving cars today are a long way from the level of the advanced AI with which we are most concerned, but they still serve as an instructive example. It seems there will be multiple players in the self driving car market, and so the primary benefits of self driving cars will mainly be reaped by the consumers not the manufacturers of the car AI.
We are in a Catch-22. We might prefer multiple competing AI players, so that the gains from advanced AI are spread broadly to capital, and power doesn't become too concentrated. But multiple competing players means the pressure of competition will cause corners to likely be cut, and far advanced AI of the Bostrom Superintelligence type may not be developed safely. To develop Superintelligence safely, we would like there to be a single entity methodically and rigorously developing it.
The demand for advanced AI technology might be almost limitless. Much as human evolution is largely driven by competition with other humans, advanced AI might be in a constant struggle to obtain an edge first in predicting and competing against humans and later even itself.
It is worth looking at a few individual sectors. Over the short term, 5-10 years, there will likely be booms and busts in various sectors. Attempting to capture them amounts to speculation. Over the longer term, say 30 years or more, these booms and busts should even out, and some more predictable trends might emerge.
- Information technology. Information technology, including AI, will clearly grow, but it seems likely the bulk of the growth will be captured by uninvestable new entrants.
Energy. Energy probably won't stand out, although there is some uncertainty regarding this. Power is a significant cost for deep learning today. However, IBM's TrueNorth neuromorphic chip showed that by using an alternative architecture it is possible to drastically scale back energy usage to around maybe 3% of the overall system cost. Part of the reason for this is there is a difference between the energy demands during the learning phase (high) and the energy demands during the execution phase (low). Unfortunately it is difficult today to predict the relative amounts of time spent in each phase.
- Real estate. Real estate in cities might slump as prices are currently driven by network effects of humans needing to be in close proximity to work effectively. With advanced AI this will no longer be the case. Moreover, in the absence of a basic income people won't be able to afford high real estate or rental prices.
Most humans have as their primary asset their human capital which will be worth little. In the absence of a basic income there will be an increasing divergence in the national and global income distribution. This applies even in the countries that are the economic winners.