Power Relationships in AI and the Historical Evolution of Engineering
It's clear now that future data centers will require substantial power for operation.
Recently, there have been quick developments in safe and widespread U.S. nuclear power, such as Microsoft and Constellation reviving Three Mile Island, or advances in small-scale local nuclear plants.
Additionally, research on renewable energy sources is ongoing, and there are plans to generate power and perform AI operations on satellites orbiting Earth to harness power differently.
Let's delve into some of the underlying obstacles and our past successes in addressing similar issues.
Electrical Engineering as a Reference
In a recent presentation, Deborah Douglas emphasized that as we tackle the power issue in AI, we have historical instances to guide us. Douglas serves as the Senior Director of Collections and Curator of Science and Technology at the MIT Museum.
She discusses early electrical systems, pointing out how the grid transformed from independent networks.
You may also relate this to the development of the Internet and its decentralized structure. Douglas further highlights the instruments engineers used for engineering prior to our current modern electrical systems.
(It's essential to note that electricity existed even before it was present in the modern systems we see today.)
Douglas then takes us back to the 1940s, a time filled with significant changes – not just in power grids but in society as well.
One Woman's Journey
She introduces the story of Phyllis Fox, who completed her master's thesis at MIT in 1949. Her title: "The solution of power network problems on large scale digital computers."
This is a narrative of females breaking through the glass ceiling, a tale of determination, and the power to impact the world.
Douglas also mentions Vannevar Bush, an influential figure at MIT, who taught electrical engineering in 1919. Fox was intrigued by Bush's differential analyzer.
A Tiring Day
Douglas illustrates Fox's working conditions as an engineering assistant at the GE company.
“She was given an office with a Marchant calculator and provided with tasks to calculate. On average, she would perform one equation per minute, resulting in 60 equations per hour, 300 to 400 equations per day, if you will. And she was expected to keep up this pace. A supervisor, known as a ‘minder,’ would walk around to ensure they were on task and prohibited speaking. Fox, however, was one of those adventurous souls. She would venture around the campus during her lunch break and found an office with a calculator where they were solving equations by hand. She would spend her lunch break completing a full day’s work on that calculator and then return to her desk to appear as if she was still working. Her curiosity did not go unnoticed, and she was eventually assigned to work with the differential analyzer.”
This calculator brings to mind MIT professor Ethan Mollick’s concept of the ‘wait calculation’ – the notion that you don't need to do manual labor extensively if you can wait for modern tools to automate the same work.
Fox was eventually rewarded for her adventurous spirit and contributed to the evolution of this important field.
However, Douglas notes that Fox was fired after World War II, then pursued her master’s degree, and eventually found another employer.
His name, she adds, was Jay Forrester.
“He fostered an inclusive hiring policy, hiring not only men but women, black individuals, the blind, and even Japanese Americans during a controversial period in U.S. history. He hired Phyllis Fox, who he believed was exceptionally intelligent.”
So, what did these pioneers accomplish with the differential analyzers?
Large Machines
I visited the MIT library website and discovered some of these early machines, including the Rockefeller differential analyzer.
These are massive machines akin to washer-dryer systems, equipped with big gears and wheels to solve differential equations.
These are extraordinary examples of analog computing.
Eventually, Fox ended up teaching at MIT, revealing more about how calculation processes work using flowcharts, block diagrams, and so forth.
She moved on to work at the Atomic Energy Commission’s Computer Center and developed Dynamo, an early computer simulation language. Fox is also reported to have written the first LISP manual, taught at the Newark College of Engineering (now the Jersey Institute of Technology), and obtained tenure in 1972.
It's all quite commendable, and Douglas concluded her presentation by suggesting that there might be someone today who can play a similar role in resolving our current obstacles and challenges.
Certainly, I hope so. However, the history of differential analyzers and analog equipment reminds me that we have the capacity to create new tools to tackle our present challenges in pioneering AI power solutions.
It could appear daunting to ponder over the fact that data centers might require vast amounts of power, like X gigawatts and terawatts. Yet, who knows what our perspective might be in a decade or two? Or perhaps even sooner?
Draw inspiration from such examples and revisit some of the outdated machines from the past. Once upon a time, they were the epitome of advanced technology, having been created less than a century ago.
In light of the digital transformation, big tech companies like Microsoft are exploring alternative energy sources to power their data centers, recognizing the potential of nuclear energy to provide the necessary energy.
As we delve into the future of AI, the story of Phyllis Fox serves as a reminder that digital transformation often requires substantial investments, just as her pioneering work in differential analyzers required significant financial backing.
