Explain artificial intelligence like I’m 5

Nowadays, the term Artificial Intelligence (AI) is thrown everywhere. It’s very unlikely that you haven’t heard it before, but at the same time, not many resources explain AI in simple terms or they get complex rapidly. So, even if we are interested in learning more about it, we don’t have the right means around us, which feels like an irony. For that reason, this blog aims to inform about how AI works in layman's terms, define it, and explain some of its internals. At the end, we will also share our opinion around some popular questions, such as What’s the future of AI? and Can AI replace our jobs?

So what is Artificial Intelligence?

Artificial intelligence is what you get when software and mathematics have a baby. After many decades of research, scientists have found a way to use math and statistics to make computers (with the use of software code) solve problems and produce knowledge mimicking the way humans do it. Artificial intelligence is a general umbrella term that describes all the efforts of trying to teach machines to do human things. But, here is the catch. There are many ways you can teach a computer to think, just as there are many ways that humans learn. Someone may prefer to write a note on a piece of paper to memorize it, but someone else may prefer to repeat it to herself many times verbally.

An easy analogy to remember is that datasets are the ingredients used to solve a problem, and algorithms are like recipe instructions. Imagine a chef with a recipe in his hands. The recipe includes some ingredients and some steps to follow. The chef’s goal is to use all the information written in the recipe to produce a tasteful dish. Notice that, a recipe is not always the same. The ingredients may differ between different chefs, as well as the instructions they follow. However, in all cases, the end result is the same, a dish which will potentially satisfy the cravings of the customer. Similarly, a data scientist is someone who tries to solve a complex problem using datasets (ingredients) and algorithms (recipes). A nice thing about this is that different algorithms can be used to solve the same problem, similar to how chefs follow different steps to produce the same dish. As you can see, both professions focus on the end goal, which is the reason why we have so many variations of dishes/algorithms. Now, an AI model is just a recipe (algorithm) used by computers to solve a problem.

What about Machine Learning?

What about instances where we have a desired goal in mind but no instructions to follow? This is where Machine Learning comes in. Machine learning is a branch of artificial intelligence that produces algorithms/models by itself by recognising patterns after analyzing a vast amount of information. Essentially, when the model needs to answer a question, it remembers all the examples that it has been trained on and uses that information to produce an answer. This approach is very similar to how humans approach a problem as well. They recall all their previous experiences that are similar and approach the problem accordingly.

How does AI learn?

The most prominent machine learning techniques are supervised, non-supervised and reinforcement learning. Since the 1950s when AI was first introduced as a concept, supervised learning was the way to go. Let’s say you wanted to create an algorithm that can recognize whether a picture shows a dog or not. With supervised learning, you take many pictures that either contain dogs or not, label them accordingly and feed all those examples into the model. Then the model tries to find common patterns and characteristics in the photos that contain dogs so when you provide a not-previously-seen picture, it can recognize whether it shows a dog. However, there are limitations as to what sorts of problems can be solved with this technique. What if you have data (information) that you can’t label? Are there any hidden patterns in there that AI can help you unearth?

This takes us to the second technique. In non-supervised learning, data is passed into the system without labels and without any specific task in mind. You can for example feed all your restaurant’s customer data into the model and it will draw conclusions about the commonalities between these customers without any human intervention. This way the model will be able to group customers based on common characteristics.

Finally, reinforcement learning is when you train a machine through trial and error. After the model produces a solution, humans evaluate it and provide feedback on whether or not it has produced the desired outcome. Most of the known and widely used AI tools today use a combination of the three techniques.

In the same fashion, Large Language Models are the part of machine learning that specialize in finding patterns in language using statistics to produce text just like humans do. The way these models imitate how we talk is by choosing the most probable word next when you ask a question. Let’s take for example ChatGPT 3. It can answer questions in a fashion that is almost indistinguishable from a human. GPT stands for Generative Pre-trained Transformer, and as the name suggests, before it was able to answer anything, the model was trained on a huge amount of text found online until September 2021. More than 300 billion words have been fed into the model including the entire English version of Wikipedia, books, websites and Reddit posts among others. The more data you use to train a model, the better the result. Even though ChatGPT’s answers sound highly rational, the system does not ‘’understand’’ the actual meaning of the words. When you instruct ChatGPT to ‘’name an animal’’ and it answers ‘’Dog’’, it is not that it knows that a dog is a living creature that barks and has four legs and that it falls into the animal category. Rather, it has associated the words ‘’dog’’ and ‘’animal’’ strongly together as they appear very often next to each other within the vast amount of text it was pre-trained on.

Why all the hype now?

When you are watching an educational video on YouTube and it suggests a funny cat video, that’s AI. When you ask Google how to lose weight fast and it gives you 1 million relevant pages in result, that’s AI. YouTube and Google are 20 and 29 years old respectively, and both of them heavily use AI. As a matter of fact, AI has been in our everyday lives for a long time now. But why is everybody talking about it all of a sudden? Well, during the last few years, AI has gotten better than humans in simple tasks such as image recognition and language understanding, which wasn’t the case before. But to appreciate the speed at which AI has improved recently, take a look at the picture below showing fake faces generated by AI over the years.


Image source

The recent hype over AI was fuelled mainly by advances in deep learning algorithms (a sub-branch of machine learning), the availability of large amounts of online data for training algorithms and the increase of the computational power of GPUs (Graphics Processing Units). Previously, in order to bypass the computational limitation, we were using less memory & GPU-intensive algorithms, which although highly accurate in a lot of cases, on a few complex problems wouldn’t produce results that would satisfy our needs. These advancements have increased exponentially the amount of problems that can be solved with computers and more and more businesses are now integrating AI into their existing operations to increase productivity. Essentially, what technology has achieved during the entire human history is to enable humans to make more things, better, in less time with fewer resources. AI, with its use in almost all aspects of our lives, is already put in the same basket with technologies that had drastic impact on the progress of humanity such as the printing press, the Industrial Revolution and the internet.

What’s the future of Artificial Intelligence?

All the aforementioned technologies and tools are described as Weak (or narrow) AI, meaning models that perform very well in specific tasks but cannot do anything else. ChatGPT can write your homework essays but it cannot play chess. Similarly, Tesla’s Full Self-Driving algorithm cannot write songs or recommend what to see next on Netflix. Their ‘’intelligence’’ involves the particular task at hand they have been trained to do.

On the other hand, Strong AI or Artificial General Intelligence (AGI) is the ultimate goal of AI research: develop a machine that can solve a wide range of problems. There has been a long-standing debate about when (or if) AGI will be achieved but the rate of growth has definitely been growing substantially over the past few years.

Will AI replace my job?

For a number of our daily-life problems, AI manages to produce results that can match our precision, and on a few occasions to surpass it. Usually, those cases include boring, repetitive stuff for most humans. But, there are so many other problems that AI models don’t even scratch the tip of the iceberg. Take this simple example, catching a flying ball. According to the famous psychologist Gerd Gigerenzer, professional players in sports like baseball use something called the gaze heuristic. When the ball is high up in the air, they start running adjusting their speed so that the angle they see the ball is constant and that’s how they know exactly where the ball will land. On the other hand, computers cannot use rules of thumb like that to calculate the landing point of the ball. To predict that, a machine will need to know the ball’s distance from the ground, its travelling speed, the angle in which it was thrown, the ball’s spin and the speed and direction of the wind. Phew. That’s a lot of necessary information to perform a simple human task, catch a flying ball. This is a simple demonstration of how difficult it is for machines to match human complex and general intelligence which was optimized by thousands of years of evolution. What is interesting, is that tasks that are very difficult to almost impossible for humans, like dividing 5-figure numbers, are done in a split second by machines whereas simple tasks such as driving a car are extremely difficult for machines. Computer scientist Donald Knuth has noticed that ‘’AI has by now succeeded in doing essentially everything that requires 'thinking' but has failed to do most of what people and animals do 'without thinking’’

AI already revolutionizes many industries and automates many tasks that were previously done manually. It can create art, proofread documents, detect diseases or beat humans at chess. But as we have seen, no specific AI model can perform all those tasks. Rather, different models need to be programmed in specific ways to perform well at each particular task and they cannot use their training to generalize across different domains. AI models perform well in limited and certain types of tasks but our everyday jobs involve countless different tasks, therefore it is unlikely that an AI model will fully replace designers, electricians, architects or lawyers, at least for the upcoming decades. For the time being, AI models will be tireless assistants, helping with all those boring, repetitive jobs that humans hate.