One of the worst-case scenarios we have considered for how AI will transform the world is a machine versus human war like in sci-fi movies such as The Terminator, which Arnold Schwarzenegger starred in 40 years ago. However, if you understand how AI technology works today, you will realize that these situations are very far away. Let me tell you about how AI works.
Before we dive into how AI works today, let me introduce you to what AI is.
AI, or Artificial Intelligence, is a research field that studies how to duplicate decision-making thinking of living things like humans, mice, and dogs. This research field has been around since the 1950s, and trust me, usable AI theory is math!
The easiest or least complex AI model is the linear equation that you studied in junior school. Because AI is used for automatic decision-making, in some cases, a linear equation can be used to tell the machine what to do next.
For example, an automatic light that has a sensor to detect how bright the environment is will give a value of 0.0 when there is no light at all and a value of 1.0 when it is really bright outside. As an engineer, it is very easy for me to set some value between 0.0 and 1.0 depending on my experiment to determine when the light should turn on and off. But if I add a clock to the light, how can I adjust the turn-on and turn-off conditions? Sometimes, there is rain at noon that drops a lot of sunlight. When should the light turn on? Manual adjustment is sometimes hard to determine, but with AI, I can do something simpler like inputting the conditions for this case, such as time and environment brightness, and train my AI to suggest when to turn the light on or off.
If the application is more complex and more important, like an automatic watering system, I will be clear on how important it is to use AI instead of rigid conditions. In that case, we should move to more complex use cases of AI.
Email filters are the tools that we use daily to split normal emails from spam, but we do not know how they work. But trust me, it's AI! Email filters use Naive Bayes classifiers that transform the body of an email into a vector and predict whether the email is spam or not. So if you have read up to this line, you should be familiar with the words "train" and "predict," which are the core of AI.
AI can make decisions like living things because of "training." Data scientists need to create a training dataset from living things (almost always humans) to train AI to "predict" something similar to the training dataset. This means that AI will do something like it has seen in the training dataset. That is the reason why every AI cannot rule the world by itself. AI cannot create something new that humans have never done before (with current technology). Now it's time to move on to the most complex AI models that tech companies are researching and announcing, such as Dall-E, Chat GPT, or Tesla Self-Drive.
Since 2019, the OpenAI lab has announced OpenAI's DotA AI, known as OpenAI Five, which uses deep reinforcement learning to master the complex game of DotA. The AI learned to play by competing against itself and learning from its mistakes, eventually becoming skilled enough to defeat some of the world's best human players. This achievement demonstrates the potential of AI to tackle complex problems and make significant contributions to a wide range of industries. In 2022, a research team from Tencent announced their Arena of Valor AI, which uses a combination of deep learning and reinforcement learning to achieve high-level gameplay. These advancements in AI technology show that the potential for AI is vast and can be applied to a variety of industries beyond just gaming. Trust me, I have read their paper and found out that it costs more than 30 million dollars for AI to play games in a project.
The technology behind this is Artificial Neural Network. The main idea of NN is to input data into a neuron node that has weight and bias and pass those values into another neuron node in the next layer until it arrives at an output layer like a real neuron network in living things. Actually, the NN theory was originally introduced in the 1980s, but at that time, computers did not have enough computing power and memory to make it possible. But currently, the resources are enough, but trust me, the number of neuron nodes has increased unimaginably.
In conclusion, AI needs to be trained with a training dataset to predict the result that the AI has seen in the training dataset. AI cannot do new things that we have never shown them. The new technology that we use in frontier research is Artificial Neural Network. And AI for doing things better is Reinforcement Learning, such as playing games or driving a car.