Welcome to Snapshots of Social Science - a quarterly newsletter where we bring you recent developments in the diverse fields of social science. Learn more about us here.
This month will be the first part of a series looking at the evolution of artificial intelligence, from Cybernetics to ChatGPT. Inspired by ChatGPT making waves (I promise this newsletter is not written by ChatGPT), this newsletter will dive into the key moments in the history of how humans perceived intelligence in computers and the progress of artificial intelligence technologies. We present some early landmark research papers that shaped artificial intelligence as we know it today—perhaps advanced enough to do the computer equivalent of human metacognition: able to reason about its own shortcomings and potential to improve. This newsletter edition is in no way exhaustive. The field is so vast that we have picked just some early moments in a big trajectory of the evolution of modern artificial intelligence.
Before we begin, please consider subscribing to ICJS Social Science by clicking the button below! This will ensure that each edition of our newsletter makes its way to you every time it is released. We would truly appreciate it if you spread the word about us as well - our impact grows with each additional reader.
Research Map
Macy Conferences
In the late 1940s, a new field called cybernetics, was taking form. Perhaps one of the first and most important instances of interdisciplinary collaboration, this field brought together mathematicians, biologists, physicists, neuroscientists, philosophers, linguists, and sociologists, (just to name a few kinds of researchers), in a set of conferences called the Macy Conferences between 1946 and 1953.
But before that, let us look at the context surrounding these Macy Conferences.
Around the same time, related scholars were exploring the concepts of information and computation, drawing strong parallels between computers and biological processes.
The foundations of Neural Networks
In 1943, Warren McCulloch and Walter Pitts demonstrated how the behavior of neural circuits can be described in terms of mathematical logic. It was known that neurons get excited or inhibited to fire if the input they receive from other neurons reaches a certain threshold. It is a simple all-or-none mechanism: neurons do not partially fire - they either fire or they do not. Given that neurons work under these simple principles, when combined in terms of nets, the behavior of the network can be represented in terms of simple logical operations. McCulloch and Pitts described the different types of networks—with and without circular circuits—and wrote a set of theorems and proofs to prove their argument.
The foundations of Information Theory (1949)
Messages are not just random collections of characters. They contain information—there is value in the fact that these characters have been chosen among a bunch of other characters. A machine will not be able to recreate English-sounding messages if all it knows is that there are 26 letters. It needs to know other things too, like the frequency of letters. In English, certain letters appear together more often than others ( ‘q’ is almost always followed by ‘u’). The example mentioned here just showed a pair of letters, but what if the machine knew the probabilities of three-letter occurrences? Or four-letter occurrences? Claude Shannon showed, in his groundbreaking paper (written in a more accessible way by Warren Weaver) that the machine was able to generate messages much closer to English even with small increases in this number. So somehow, the design of the machine (how much the machine knew in order to be able to generate a close-enough message) was connected to the amount of information in the message. What, then, could be a measure of this information? In precisely this paper, the bit was born.
Alan Turing and claims about computers and intelligence (1950)
Can machines think? Alan Turing, the genius mathematician, biologist, cryptanalyst, and philosopher, attempted to answer this question with ambitious claims, and a novel test that he called the imitation game (and yes, it’s also the name of the movie on Alan Turing). A woman and a man in different rooms are questioned by an interrogator who does not know which is the man and which is the woman. The interrogator has to find out which is which, and each’s (man and woman) job is to mislead the interrogator. Turing then posed the question: ‘what if a computer were to take the place of either the man or the woman? Would the interrogator still be as wrong in their decisions as before, when deciding between man and woman?’ And this question, he argued, replaces the question of ‘can machines think?’ A philosophical question indeed, but the adapted version of this test is called the Turing test in modern artificial intelligence.
Self-reproducing machines—and DNA (1950s)
What is life? One could say that the ability to reproduce is distinct to living beings. But what if a computer could reproduce? (See Complexity: A Guided Tour for a detailed explanation, but here is the general idea) Imagine a computer program that could replicate itself. That is, executing the program would basically write the entire program itself. In this case, the execution of the program involves an interpreter and memory that is exterior to the program itself. A cursor needs to point to the right code line and another has to print the code. On the other hand, DNA does not—its interpreter is encoded in the machinery to self-replicate. And John von Neumann, mathematician, computer scientist, and physicist is credited to have described the first truly self-replicating machine, just like the DNA, before self-replication in DNA was discovered. Unfortunately, he died before he could publish his work, so the paper was edited and published by his colleague in 1966.
All of the aforementioned scholars, most of whom were strongly tied to the Macy conferences, were looking at technological and biological systems in a new light. And all of these insights came together as cybernetics.
So what is cybernetics—the theme of the Macy conferences?
Norbert Wiener, the father of cybernetics, called it the “control and communication in the animal and the machine,” in his 1948 book. In this book, he laid out all the foundations for the field of cybernetics. Chapters in this book explain a wide range of concepts in physics, information theory (in collaboration with Claude Shannon), meteorology and prediction, feedback in systems, psychology and neuroscience, sociology, replication in machines (featuring John Von Neumann’s work), and evolution theory. Now that all of these fields are coming together, we can only marvel at how ahead of its time this book was. After the success of his previous work, Wiener wrote another book in 1954 called the Human Use of Human Beings, in which he describes how humans and machines can communicate and the impact of that on society.
Cybernetics is such an all-encompassing field that it cannot be summarized in a single definition. Ross Ashby, another pioneer in this field and the author of what is considered the first textbook of cybernetics, called it the “art of steermanship.” He explains the main themes of his book as ‘coordination, regulation, and control.’ Cybernetics, according to Ashby, is very closely tied to physics, but is different from it in the sense that it focuses not on the physical material, but instead its behavior.
For next time
We will pick up this story starting from the birth of the discipline called Artificial Intelligence at the Dartmouth workshop (1956).
Special mention for this month
See Prof. Ronald Kline’s work in reviving cybernetics today and exploring its connections with modern AI.
Spread the Word!
Thank you so much for reading! Please share this newsletter with anyone who would be interested. Press the button below to share!