AI does not exist!

AI doesn't actually exist. At least, not as a single “something.” AI is not just Siri or Alexa, not just a way to predict the weather, and certainly not just a Large Language Model (LLM). AI is an umbrella term for research within computer science. All kinds of processes, techniques and also goals fall within this research. According to Wikipedia, traditional goals of AI are learning, reasoning, knowledge representation, planning, language processing, perception, and also robotics support. And all with a computer. But that's not all.

Shifting definition

Funnily enough, the definition of the term AI is constantly shifting. Things that we are not at all surprised by today were first included in the field of AI. Consider, for example, computers that can play chess better than humans: groundbreaking in the nineties, but now dead normal. Another example. Your personal feed on social media like Instagram or X. These are created with so-called recommendation systems. Before this was AI, now this is standard Data Science and no one thinks of the word AI at all when they scroll on Instagram. This is also known in the field as the AI Effect or, sloppily translated, the Strange Paradox. Therefore, when an AI trick becomes widely used - and thus normal - it is no longer seen as AI. As such, Tesler's thesis states, “AI is whatever hasn't been done yet.”.

The AI hierarchy

AI is a very broad, mostly goal-oriented definition, constantly shifting. Check. Yet you constantly hear terms like Machine Learning and LLMs flying by. What about that?

AI encompasses all sorts of techniques that we describe here. Like AI, Machine Learning (ML) is also a generic term. There are all kinds of ML techniques, such as for classifiers and regressions. These are mostly still statistical models, which people are already finding increasingly normal. ML also includes artificial neural networks (ANNs), which are the basis for deep learning (DL). ANNs are, as the name implies, networks of artificial neurons. Yes, copied from how it works in humans! (See figure 1).

Figure 1: This figure displays how an artificial neuron works.

If you make these networks more and more complex, where there are a lot of neurons - and thus computations - taking place in the middle of the network, that is called deep learning (DL). Deep learning is another umbrella term in itself, and not everyone agrees on the definition. Usually we speak of DL when a ANN has at least three hidden layers (See figure 2). Within deep learning there are, you expect it, again all kinds of goals and directions. Generative AI (GenAI) falls within deep learning, with goals such as text generation, as well known by now. We also find here, among others: deep reinforcement learning, representation learning and discriminative tasks (as typically solved with statistical models within ML).

Figure 2: If you link multiple artificial neurons (as shown in figure 1) together in a network, you get an artificial neural network (ANN). With at least three hidden layers, it is called deep learning.

Once we arrive at GenAI, we are in the territory of LLMs. Traditional large language models (read: LLMs), as known from ChatGPT and Copilot, are used for text generation, and thus fall within GenAI. However, there are many other GenAI techniques. For example Generative Adverserial Neural Networks (GANs), Latent Diffusion, Transformers (the technique behind modern LLMs) and Variational Auto Encoders (VAEs).

See Figure 3 for a visual representation of this hierarchy of AI techniques.

Figure 3: The AI hierarchy visualized.

GOV-AI is explainable AI

What about the black box; a term often associated with AI? AI models are not transparent, it is said. But what models are we talking about exactly? Typically, models built from artificial neural networks are not very transparent. This is partly due to their fundamental design, but nowadays mainly due to their scale. As these models get bigger and bigger, it becomes more difficult to make sense of what exactly is happening. Sometimes this is fine, but for applications in the public sector, for example, it's not okay. And that's where GOV-AI (Governmental AI) comes in. GOV-AI is a collection of techniques and design principles, which ensure transparency, accountability and data security when developing AI systems. An important part of this is Explainable AI (XAI). XAI primarily revolves around exploring and developing techniques that allow people to hold intellectual control over AI algorithms and systems. Essential for sensitive industries such as government!

AI in Perspective

The recent AI hype, followed by the revolutionary product ChatGPT, has brought a lot of attention to the field of AI. Because this is so recent, and the field has been evolving for years, it is logical that misunderstandings arise. Having made it to the end of this article, you can finally explain at the dinner table how LLMs, AI, and Generative AI relate to each other. Not unimportant, as there are many developments. Also, by now we know one thing for sure: the definition of AI will be different in 10 years or so from today!