Over the past year, discussions about Artificial intelligence (AI) have captured the attention of everyone from everyday citizens to the Pope. Despite this surge in awareness, the public largely misunderstands how AI functions and how closely its development mirrors traditional manufacturing processes. This fundamental misconception is driving a fierce, increasingly bipartisan backlash against the physical infrastructure necessary to sustain AI’s growth.
As it has evolved, AI is, at its core, best understood as an extremely sophisticated form of automated pattern recognition. What we call computer neural networks are first trained on enormous amounts of data gathered from the internet using algorithms — really just mathematical equations — to discover patterns or relationships; once trained, they apply those patterns to new questions. Because modern computers can detect patterns sometimes invisible to humans and then use those patterns to solve problems at an almost unbelievably fast speed, their output seems magical. Since AI can answer questions and perform tasks much like very smart humans, it is tempting to anthropomorphize it.
Recommended Stories
But machine learning is not the same as human learning. People learn by grasping underlying concepts; a machine does not. Instead, AI “learns” the way a player learns in a game of “hot-and-cold.” As the neural network attempts to guess at answers, the patterns it has learned are equivalent to someone calling out “warmer” or “colder,” to a player. After multiple rounds of adjustment, guided only by those tiny corrections, AI, like the player, homes in on the desired object. But neither the blind folded player, nor the AI system, “knows” why any particular move worked. This is accuracy without true comprehension.
In fact, it’s this lack of comprehension, and by extension anything resembling compassion or empathy, that makes people nervous and fuels calls for regulation of AI. Of course, like many beneficial technologies, AI can be misused, and both AI and the infrastructure it runs on create externalities that need to be accounted for. But it’s important to recognize that AI is a real, valuable product, albeit a generated one. Given its use in everything from military logistics to traditional manufacturing, it is arguably the most important product a nation can lead in producing in the 21st century.
Because AI is fundamentally a system of processing and refinement, particularly when it comes to training what are called frontier models, AI can be understood as a kind of manufacturing process.
AI “manufacturing” takes massive datasets scraped from the internet as raw material and processes them with specialized computing hardware — clusters of advanced GPUs operating in unison in what we call data centers. The finished product is a trained neural network capable of generating text, images, or insights on demand.
As with most manufacturing processes — steel making, for example — the finished product enters the supply chain becoming the “raw material” for other factories. Although the “work” done in these data center “factories” is largely automated, processing units do not build, install, or maintain themselves. In addition to traditional IT engineers, data centers require, on an ongoing basis, pipefitters, electricians, and HVAC specialists, among other skilled trades. Rather than a job killer, AI will make many industries more productive while creating the sorts of blue-collar jobs the nation has been lacking.
For decades, American manufacturing has faced the challenges of offshoring and the resulting hollowed-out mill towns. The tech sector was celebrated, but with talk of places like “clouds” it was difficult for people to wrap their heads around exactly what they produced. The answer, “information,” didn’t seem tangible. But the cloud is a physical place, and the AI boom is pulling tech firmly back into the material world.
ANTHROPIC ARMAGEDDON: DARIO AMODEI’S DOOMSDAY FEARS CONVENIENTLY CRUSH HIS COMPETITION
Of course, opposition to data centers is not driven solely by confusion about AI; communities and policymakers also raise concerns about their energy consumption, water use, and environmental impacts. We have to be honest: As with earlier mills and factories, there is a trade-off between economic progress and environmental impact. When the steel mills were operating in places like Pittsburgh, Pennsylvania, and Youngstown, OH, the soot often darkened the sky, and the rivers suffered from industrial run-off. Through regulation and advancements in technology, those problems have been mitigated, although not completely eliminated. As with steel, however, if we don’t operate the mills, another country will, and probably in much less environmentally sound ways, and to the detriment of both our economic and national security.
The AI revolution will be coded in Silicon Valley, but it will be physically forged in steel mills, powered by upgraded grids, and built and maintained by blue-collar labor. By recognizing AI not as a magic box but as one the world’s newest manufacturing process, the U.S. can make sure that this is one industry that is not outsourced to our competitors.
Paul Sracic is a Senior Fellow at the Hudson Institute.
