New advancements in artificial intelligence have a lot of us asking the question: “Can computers think?” I think the answer is currently no, though they might achieve this at some point in the distant future.
Meanwhile, humans have achieved a state that I will call the “beehive brain”, where the written thoughts of all of humanity can be quickly assimilated and analyzed, producing a result that looks very much like thinking.
The big difference between a tool such as ChatGPT and an actual beehive (or ant colony) is that in a beehive the collective thinking evolved organically over millions – if not billions – of years, whereas this new way of “thinking” among humans has developed in a matter of decades thanks to the coupling of mathematics and technology, both products of the actual human brain.
Let’s step back and look at the history of AI. Mathematically, it goes back to George Boole, a 19th-century English mathematician and philosopher who developed a branch of mathematics, now called Boolean algebra, which is based on postulates using zeroes and ones. Boole considered his algebra to be the basis of thought.
Another 19th-century genius, a neuroscientist called Santiago Ramon y Cajal, then set the stage for AI by creating beautifully hand-drawn pictures of the brain’s intricate neurons. (On a side note, I recently read a wonderful book on this almost forgotten Nobel laureate, called The Brain in Search of Itself by Benjamin Ehrlich).
Donald Hebb, a Canadian psychologist who taught at McGill University, then postulated that these neurons were activated by the electrical firing of their axons, the connections between cell bodies. Hebb’s non-mathematical theory inspired two mathematical approaches to modelling the brain: Frank Rosenblatt’s 1958 perceptron model, which modelled a simple neuron with its weighted and summed inputs, and John Hopfield’s 1982 model based on statistical physics. Although most modern AI programs trace their origin to the Rosenblatt model, Hopfield was a co-recipient of the 2024 Nobel Prize in Physics for his pioneering work and some machine learning methods are still based on his approach.
In the perceptron model, Rosenblatt came up with a simple algorithm that allowed a perceptron to “learn” by adjusting the weights of its inputs, which were like the inputs into a neuron. This gets us back to Boolean algebra, because the perceptron was able to learn simple linear Boolean concepts such as AND and OR, but failed to learn the most basic non-linear Boolean function, called the XOR. (As an aside, the “thinking” algorithm in Microsoft Word thinks I just made a mistake by repeating the word “and”, but it does not understand the context!).
This led to the first “AI winter” in which AI fell out of favour. But a ground-breaking paper by Rummelhart, Hinton, and Williams in 1986 showed how to overcome this limitation by putting multiple perceptrons together in layers and using a more sophisticated learning technique called error backpropagation.
Backpropagation and its variations are the mathematical tools that have been used to develop modern AI algorithms. The large language models (LLMs) that are the basis of ChatGPT and other “thinking” programs, use hundreds of layers of perceptrons and are trained by backpropagation. However, they have an added feature called a transformer that allows them to assimilate large amounts of training data from the web and predict words in advance, which makes them almost look like they are thinking.
Geoffrey Hinton, the second author of the paper I just mentioned, is an English/Canadian researcher at the University of Toronto. He has done an amazing amount of work on AI since the 1986 paper and co-shared the 2024 Nobel Prize in Physics. I have hit some of the highlights of the development of AI, but space does not allow me to mention all the researchers involved.
But it took more than mathematical algorithms to get us to today’s large language models. Two huge technological breakthroughs were also necessary. The first was the discovery of electromagnetic wave propagation, which traces its origins back to the English experimental physicist Michael Faraday and Scottish theoretical physicist James Clerk Maxwell, whose experiments and equations explained the coupling of the electrical and magnetic fields. Learning how to create these propagating electromagnetic waves, and to carry information over them, has led us all the way to the modern internet.
The second was the development of the digital computer, which first relied on vacuum tubes and then on digital transistors. Interestingly, Boolean algebra also played a role in the development of the digital computer, because it was the brilliant U.S. mathematician/electrical engineer Claude Shannon who recognized that the Boolean XOR function to which I referred earlier could also be built using transistors. This led to the development of the digital adder, the heart and soul of every computer. Shannon, along with Alan Turing of England, are credited with the development of what is called “information theory”, which is also at the heart of modern computing.
On the backs of the digital computer and information theory, programming languages such as Fortran, C++, and Python have allowed us to implement the mathematical algorithms I described earlier.
Thus, mathematics, physics, electrical engineering, and computer programming have got us to where we are today, with computers that almost think. Is this a good thing or a bad thing? I go back and forth on this question but currently my thinking is that it is a good thing if handled correctly. I described the state of AI as the creation of a “beehive” brain that connects all our knowledge in one giant web. Another way to think of this is as having access to the world’s most brilliant professor, who not only knows everything about everything, but has infinite patience to explain anything you want to know. To me, this is the most beneficial thing that has come out of machine learning.
If I take you back to my childhood 60 years ago, the only way to access information was through an individual teacher or a library. And you had to be lucky to find the right teacher or book to explain a subject to you, one that you understood. Each teacher and author brought their own knowledge and biases to the difficult task of teaching. We now can put every book and idea ever written into a big stew, thanks to AI, and decide for ourselves where the truth lies.
A common criticism I hear about AI tools is that they often “hallucinate” and tell you things that are not true. I have experienced this myself. But this is no different from the way it has always been. There have always been books and teachers who distort or misrepresent the truth. In fact, this is where the “hallucinations” come from, since the input to a large language model comes from every source, good and bad. So, it is up to you to decide when to trust what AI is telling you. I must confess to having had long “discussions” with ChatGPT. I often point out when I think it is feeding me incorrect information (or BS, if you want to put it more crudely). Surprisingly, the new versions of ChatGPT will admit to you when they are wrong and will work with you to correct the narrative. These “discussions” are almost Socratic in nature. I find that I have learned a lot on many subjects from ChatGPT.
Finally, should we be worried that the new AI is going to destroy our education system and our kids’ brains? Again, it depends on how it is handled. These tools will allow us to re-think the way kids are educated, and we need to show them how to use the tools in a positive way, like the way I used the library when I was in university 50 years ago. Large language models have the potential to vastly improve the way our kids learn if they are handled correctly, but also to make them lazy if they are not. In summary, like every other technological advance, AI tools have the potential for harm or for good, and we must decide how to go forward with them. Two things we count on are that they are here to stay and that they are only going to become more powerful.
