Erik Brynjolfsson, Daniel Rock, and Chad Syverson write,
Systems using artificial intelligence match or surpass human level performance in more and more domains, leveraging rapid advances in other technologies and driving soaring stock prices. Yet measured productivity growth has declined by half over the past decade, and real income has stagnated since the late 1990s for a majority of Americans. We describe four potential explanations for this clash of expectations and statistics: false hopes, mismeasurement, redistribution, and implementation lags. While a case can be made for each, we argue that lags have likely been the biggest contributor to the paradox. The most impressive capabilities of AI, particularly those based on machine learning, have not yet diffused widely. More importantly, like other general purpose technologies, their full effects won’t be realized until waves of complementary innovations are developed and implemented. The required adjustment costs, organizational changes, and new skills can be modeled as a kind of intangible capital. A portion of the value of this intangible capital is already reflected in the market value of firms. However, going forward, national statistics could fail to measure the full benefits of the new technologies and some may even have the wrong sign.
That is from the abstract. I cannot find a free ungated version of the full paper. Meanwhile, my thoughts:
1. I don’t have enough confidence in productivity data to believe a statement like “productivity growth has declined by half.” I’ve already explained why. I already think that the national statistics are misleading.
2. As to the diffusion explanation, maybe we’re in a situation today where AI and machine learning are like mainframe computers, with seemingly only a few giant firms able to take advantage. Maybe if somebody comes up with “AI and machine learning for the rest of us” the story will be different.
“…maybe we’re in a situation today where AI and machine learning are like mainframe computers, with seemingly only a few giant firms able to take advantage.”
No. The new AI with deep learning convolutional nets is extremely accessible to all firms, maybe even hobbyists. The tools are readily available and mostly open source. If you’re curious, check out:
https://developer.nvidia.com/embedded/twodaystoademo
(NVidia is a graphics company and their GPUs (Graphical Processing Units) happen to be nearly ideal for implementing the new AI).
More importantly, like other general purpose technologies, their full effects won’t be realized until waves of complementary innovations are developed and implemented.
Well, most technologies take decades to roll out if you analyze histories. All the basics of indoor plumbing will around in 1880s while it several decades to roll to the general populations in the 1910s – 1920s. We act like everybody had cars 5 years after the invention and the even the basics of TVs were around in the early 1930s. So AI will take years for the rest of the economy to grow into the growth.
Also why have incomes stagnated. First China and India have lots of wage increases the last generation and the growth of AI diminishes the value of human labor.
“Maybe if somebody comes up with “AI and machine learning for the rest of us” the story will be different.”
I don’t think we are waiting for anyone to come up with the idea. It seems like I use the techniques every time I yell, whisper or say in a normal voice to my phone “set an alarm clock for 6 AM” and even with random “ums” and “ahs” it processes the audio signal and sets my alarm clock. It is part of facial recognition, which is kind of boring but also used in predicting flat surfaces or walls and other visual processing tricks. No doubt it is used in web services and AB testing to improve some QOS metric. If it is working well it wont be calling the user’s attention to itself, as that would be inefficient, though I imagine the team running these services is constantly messing with it and think about it.
My guess is it will become more obvious as the price of hardware goes down and people can work on applications less intuitively valuable then sight and sound. Even with low hardware prices, the need for scored data that reflects the environment, together with an environment that changes, means you will constantly need new data. You will be running a service which is harder to do then selling a program.
You can rent AI of various flavors by the hour.
https://aws.amazon.com/amazon-ai/
“seemingly only a few giant firms able to take advantage”
So long as what you really mean is “only a few firms have enough data to train the AI and take advantage” then yes, I agree with you. My team and I did a morning training class on machine learning in the Spring. Between an open source dataset and a CloudFormation set of AWS servers, we had an AI recognizing cat pictures in about 45 minutes.
The key is the data.
Diffusion lags. Like we needed the browser to get wide diffusion on the web. Needed the cheap digital device for mobile integration.
Search engines are quite exclusive, not diffused. The search engine is used very productively by less than 10% of the population.
Listening to the Econtalk with Dennis Rasmussen today, Dennis cited Hume in the context of free trade:
“But, Hume is arguing for free trade decades before the Wealth of Nations appears, saying, ‘What’s the true source of a nation’s wealth? It’s not gold, it’s not silver; it’s not a positive balance of trade. It’s a productive citizenry.’ That, politicians’ attempts to guide or control people’s economic choices are going to be, um, either just futile or maybe even positively counterproductive. ”
The nation’s wealth being it’s productive citizenry set me to wondering about the changes in recent decades as more and more of workers have been shifted from industry, where increase efficiencies and cost reductions are the norm, to trade and “clientry” (long term relationship services, such as your dentist), where productivity growth is more difficult to achieve. This shift came from both the movement of manufacturing “off shore” and from automation.
While productivity may grow due to AI in some businesses, the overall productive citizenry seems to have stagnated in productivity growth mostly due to being pushed to low productivity growth sectors of personal service and trade. The Youtube creator can only produce so many videos a week, even if they may earn more by getting more views on what they are able to create. A doctor can only see so many patients in a day. Saving a nickel on every exam for 20 exams in a day doesn’t compare to saving a nickel on every unit in a 100,000 unit run.
And Satya Nadella is saying the same in his book “Hit reset”. Through Azure AI and powerful deep learning can be bought cheaply by everyone from cloud services. He mentions web-services for recognizing human emotions that can be tapped by everyone.