Twenty years ago, Hal Varian and Carl Shapiro published what has become the classic introduction to network economics. Called Information Rules it described and illustrated key economic concepts like network effects, positive feedback loops, standards wars, market tipping points and switching costs, using examples that are now so dated that would not be recognizable to today’s digital natives. But the text drilled into a generation of entrepreneurs and policymakers the importance of understanding the basic economics of network industries before starting a network business or trying to throw a regulatory net around a network industry.
Today Hal Varian works as Google’s chief economist. In his personal capacity he delivered a crash course on AI and data to the U.S. Chamber of Commerce’s TecGlobal 2018 meeting on April 4.
He illustrated the familiar advances in machine learning through pattern recognition in voice and images, noting that it was the greater availability of data, improvements in hardware and software and the focus of thousands of human experts that made the machine algorithms so effective.
But data is a different sort of economic asset. It is not the new oil. Yes, it has to be processed before it becomes something useful, but it is not a rivalrous good. One company can use the same information that another company uses. Various measures like contracts or trade secrets or privacy rules can partly exclude others from data that one company has collected, and this makes data a little like a tennis or swimming club. For typical private goods like oil, it makes sense to ask who owns it. For club goods like data, it is more sensible to think in terms of access.
He illustrated this idea of access by asking the question who has access to autonomous vehicle data. Regardless of ownership, various actors have access to the data to do their job, including the manufacturer and the system designer and in some cases investigators for traffic safety or municipalities for planning transit systems. As in the case of airlines, a plurality of actors need and get access to data to make the system run effectively and to trace the sources of errors.
But because data is now so plentiful, it isn’t really the crucial thing anymore. Take image recognition. The data sets involved in training algorithms to recognize images have been more or less the same size for almost a decade. Yet the error rate has steadily declined from over 25% in 2011 to under 5% today, surpassing human performance in object recognition in 2016. The improvement comes from technical advances in the hardware, software and human expertise involved in machine learning.
He illustrated the image recognition software with the example of his cat, which Google’s image recognition software correctly identified as a cat with 99% probability and as a ragdoll cat with 78% probability.
His second big topic was competition. He asked who is providing this kind of image recognition software? Is it just one company with a proprietary hoard of data? He noted that image recognition software is made available for commercial use by Amazon, Microsoft, Google, Adobe, VMware, IBM, Rackspace, Red Hat, Oracle, and SAP. These companies also compete in providing related services including text detection, safe search, and logo detection.
The five biggest tech companies – Apple, Google, Facebook, Amazon and Microsoft – compete not just in a single market but in the provision of a wide range of products and services including advertising, browsers, digital assistants, email, search engines, maps, office tools, operating systems, productivity tools, smart phones, social networks, streaming music and video and video conferencing.
In an especially telling point, Varian highlighted the extent to which technology companies are leading in research and development investments. Amazon is far and away the biggest investor, spending a total $22.6 billion in 2017, followed by Google ($16.6 billion), Samsung ($14.9 billion), Volkswagen ($14.8 billion) and Microsoft ($13.9 billion). The idea that we should target these companies for dismantling because they are too big would be a big mistake. They are our biggest source of innovation.
To further illustrate the extent of competition in tech markets, Varian used the example of virtual assistants. Apple, Google, Amazon, Microsoft, Facebook and Samsung all vie for consumer loyalty in a vibrantly competitive marketplace.
As measured by investment by venture capital firms, innovation in start-ups is at the highest level since the late 1990s in both the U.S. and in Europe. And it has never been easier to start a new tech company. The basic tools - from computing services to accounting software to machine learning software - can all be rented rather than purchased, allowing the entrepreneurs to focus on their core competencies, not the construction and maintenance of basic business infrastructure.
But this is not a reason for complacency. A point that Varian did not make is that China’s AI startups scored more funding than America’s last year. Of $15.2 billion invested in AI startups globally in 2017, 48 percent went to China and just 38 percent to America. We cannot afford to let our leadership in this important new area slip away.
As we think through the need to address problems in the digital economy, it is useful to have a guide to the basic economic facts underlying the tech marketplace. Hal Varian’s talk provided this needed guidance.