Whatever a human may do, it can be described as a workflow of activities and resources. And with few exceptions, anyone with a task before them, once they have an objective in mind, seeks to find the fastest and cheapest, you might say most efficient, way of getting the job done.
Many of the leading narrative-makers in the world of technology have likened the arrival of retail artificial intelligence, whether ChatGPT or another, in terms of its impact on economic freedom and individual productivity, to either the arrival of the farm tractor or utility-scale electricity.
References to these examples are meant to imply that AI is something like an order of magnitude improvement in efficiency in the fundamental infrastructure of the industrial economy. Like the cheeseburger that proliferated from the tractor, or the automobile from electricity, everything downstream from AI will become cheaper and easier. And everything is downstream of AI.
The Economist writes:
Speculation about the consequences of AI—for jobs, productivity and quality of life—is at fever pitch. The technology is awe-inspiring. And yet AI’s economic impact will be muted unless millions of firms beyond tech centres like Silicon Valley adopt it.
The point about “beyond” Silicon Valley rightly highlights the problem of technology diffusion:
Horror stories abound. In 2017 a third of Japanese regional banks still used COBOL, a programming language invented a decade before man landed on the moon. Last year Britain imported more than £20m-($24m-) worth of floppy disks, MiniDiscs and cassettes. A fifth of rich-world firms do not even have a website.
In other words, any technology will not fulfill its full-potential promise to the economy if it’s used merely as a superficial appendage to existing workflows, or properly taken advantage of by only the top firms. In that case it is largely a driver of inequality rather than an engine of growth. Weak technological diffusion is identified as the culprit:
Three possibilities explain lower diffusion: the nature of new technology, sluggish competition, and growing regulation. Robert Gordon of Northwestern University has argued that the “great inventions” of the 19th and 20th centuries had a far bigger impact on productivity than more recent ones.
These comments are playing semantics. It is easy enough to say that the most significant benefits of AI and other such “electricity-like” technological breakthroughs will be realized when workers and their firms entirely redesign their workflows, with the new technology as a first principle rather than as a seasonal salad dressing. Blaming weak competition or egregious regulation are boring and lazy misinterpretations of the data.
The question not being asked: why is diffusion so weak, and why has it become more weaker over time? Beyond the obvious general issues with technology adoptions as implied by path dependence and the knowledge problem, there are two overlooked reasons why electricity and tractors are different category of economic prospect relative to quantum computing, AI, cloud computing, and websites.
The first is passive versus active participation or one-way versus two-way communication. This also relates to the concept of centralized versus decentralized systems. To cut to the point, electricity diffused because it could be forced upon the worker. If your factory did not adopt electricity, you would simply be inviting a competitor to create a replica across the street with only one difference: electricity. The worker did not have to “learn electricity”. If someone decides against using AI, the downside is unlikely to change.
Leverage is the second overlooked factor slowing the diffusion of AI. Let’s recall the basic structure of a unit of work products. Each work product is comprised of time (humans and machines) and resources (physical and digital). Each specific input can be described in terms of its cost contribution (percent of the retail price) and its variance (the degree to which the cost can be reduced). The tractor played a starring role in the transformation of the economy: agriculture went from almost 10% to less than 1% of GDP.
If AI becomes embedded into our systems, optimizing outcomes without front-line worker consent, and promises (like the tractor) to make chunks of the economy ten-times more efficient, you can bet that it will diffuse fast. Otherwise, its effects, however locally lucrative for a niche monopolist, will largely be reserved to the AI frontier of “special problems needing special data”.