Is a calculator AI? What is AI?
- Ian Vicino
- Aug 2
- 6 min read
I was speaking to my colleague at work about artificial intelligence (AI). I don’t remember how the conversation started, but what I do remember is that he wanted to know whether a calculator can be considered AI. During the conversation, we spoke about how ChatGPT does not “learn” while it is being used, but during its creation. (I put learn in quotations because what we refer to as AI learning is just our way of anthropomorphizing the AI. You will learn more about how AI is trained in this article.) After our conversation, I was inspired to write this article, not only to teach my colleague and you how AI works and how it may be different than a simple calculator, but also as reference material for my AI class.
So, it turns out the definition of AI is nebulous, and I may have been wrong when saying a calculator was not AI. While researching to verify that what I told my colleague was correct, I asked Google whether a calculator was AI the Google’s AI model initially responded in the negative, but when I specified the question, asking, “is a calculator and ai, are they not a reactive machine,” it responded that a calculator could be a simple form of AI. The reason for the confusion in responding to my question is that the definition of AI has changed over time; there is no standard definition of AI.
Connection Between Deep Blue and a Calculator?
When IBM created Deep Blue and used it to beat a chess player, Garry Kasparov, in 1996, it shocked the nation.1 This was the first time a computer was able to win against a champion chess player. This was a milestone in the history of AI, but here is the catch: all the moves Deep Blue made on the chessboard, its output, determined by its input, the state of the chessboard at the time, was programmed by IBM developers. It could all be found in the source code of Deep Blue, all the possible states of the chessboard and subsequent moves made by Deep Blue, it was all there.
This form of AI is called a reactive machine, one that reacts to its inputs.2 These reactive machines predictably respond to input; when they receive the same input, they output the same response. This is essentially how a calculator works. When you type in 5*5, you will expect and receive 25 every single time as an output. We don’t typically consider a calculator, a very simple machine created as early as 1623, to be a form of AI, which is why when initially asked “is a calculator and ai” Google’s AI responded in the negative.3 However, as it turns out, a calculator is a reactive machine, the same way Deep Blue is a reactive machine, which is why it can be considered a simple form of AI. I was wrong in my response to my colleague.
Modern Artificial Intelligence Models
What we typically think of as AI are machines like Gemini, Siri, Dalli, and ChatGPT. These forms of AI are created using artificial neural networks (ANNs), which are computational methods to simulate the neural networks within our brain’s visual cortex. This is where the fun begins, where you get to learn about modern AI models.
Artificial Intelligence Models that don’t Require Artificial Neural Networks
Not all modern AI models are required to use ANNs. I have created AI models that don’t touch an ANN, including models that use automatically created decision trees to decide an output, and a self-playing game using the Q-learning AI model. These models, even though they don’t use an ANN, can “learn” to connect input and output on their own. Their “learning” comes down to math, math which I will not get into during this article, but which you can learn more about by following these links: Math behind Decision Tree Algorithm: https://ankitnitjsr13.medium.com/math-behind-decision-tree-algorithm-2aa398561d6d, Math of Q-Learning — Python: https://medium.com/data-science/math-of-q-learning-python-code-5dcbdc49b6f6.
The Perceptron
The origin of the ANN is an interesting one. It was created to simulate the neural connections of the visual cortex on the computer. The original single-layer ANN was called a perceptron, and it was able to figure out the connection between input and output on its own using a perceptron learning rule. Introduced by Frank Rosenblatt in 1957, surprising how old the ANN is, huh, it laid the foundation for what would become our modern understanding of artificial intelligence.4
(I will be explaining how a perceptron works, referencing this site: https://www.geeksforgeeks.org/machine-learning/what-is-perceptron-the-simplest-artificial-neural-network/)
A perceptron is composed of 7 key components: an input layer, weights that are adjusted until an adequate output is reached, a summation function that adds all the inputs multiplied by their weights, an activation layer, which takes the weighted sum and if it is above a threshold number it outputs a 1, otherwise it outputs a 0, the output layer, a bias which helps the perceptron make adjustments independent of the input, and a learning algorithm which adjusts the weights until the appropriate output is achieved. This is only a rudimentary explanation of how a perceptron works, and if you want to explore this further, follow the link to my reference to not only further your understanding of how a perceptron works, but also program your own perceptron. Essentially, a perceptron uses an equation similar to the equation of a line, y=mx+b, but instead of a slope, m, it is your input, the x is the weight multiplied to your input, b is the bias, and y is the output. The actual equation used is more like Output = Activation Function(Input vectorT * Weight vector + bias). The output is corrected by adjusting the weights according to the perceptron learning algorithm until the correct output is found. This is called supervised learning, and it will be the discussion of the next section of this article.
Supervised Learning
Supervised learning is when an AI learns the connections between an input and a desired output on its own. The AI algorithm is given an input vector and an output vector and told to find out how to generate the given output if given a specific input. This is why it is called supervised learning, because it is supervised by the input/output data it is exposed to. Once it can learn those connections, if given an input it has not yet seen, the goal is that it can use the connections it has “learned” to identify an appropriate output. Using an example that can be solved by the single-layer perceptron, if trained to identify SPAM using a series of emails that are labeled as SPAM (1) and those that are labeled as Not SPAM (0), the perceptron should be able to identify an email that it has not seen before using the connections it has learned as SPAM (1) or Not SPAM (0). (I won’t discuss unsupervised learning here, but if you are interested, check out this link: https://www.geeksforgeeks.org/machine-learning/unsupervised-learning/)
Artificial Neural Networks
Artificial Neural Networks (ANNs) use supervised learning to "learn". They are networks comprised of multiple layers. They are also called multi-layered perceptrons, although the number of layers is not the only difference between a single-layer perceptron and an ANN. ANNs also “learn” differently; instead of using the perceptron learning algorithm, an ANN uses backpropagation, a method that makes use of gradient descent and several advanced mathematical concepts, including multi-variable derivatives and differential equations. Linear algebra is used to add vectors and multiply matrices during the ANN as well, so math is found throughout the entire workflow of an ANN. Yay math!
There is a fantastic YouTube series by 3Blue1Brown that explains exactly how an ANN works, and I recommend you watch it: https://youtu.be/aircAruvnKk?si=vrpwgV5A-XHMLjjD.
Conclusion
It turns out that a calculator can be considered a form of AI, even if it is just a very simple form of AI. If this is the case, however, everything we do on our computers and our phones can be considered AI since they are much more complicated than a calculator and have preprogrammed output for every input, just like Deep Blue did, and something seems off to me about that.
We have advanced so fast with computation, and what was once a surprising feat, beating a professional chess player with a computer, is now relatively easy to do. We have even created a program to beat a professional Go player, and that was definitely using a self-“learning” AI. Instead, I will define AI, and I can do so since there is no official definition or understanding of what AI is, as a software program that, by using data or a software architecture, can figure out how to connect a desired output with an input without being hard-programmed to do so. Although by using this definition, Deep Blue is no longer considered AI, neither is a calculator or word processor, so I am satisfied with that.
References (Links):
Comments