How to Teach AI to Understand Your Business

What is Artificial Intelligence?

Artificial intelligence is an area of informatics that attempts to generate machines or systems mirroring human intelligence. This does not mean, however, that AI thinks like a person, but simply that it fulfils tasks that would otherwise require human thinking. In short, AI is a machine or, more accurately, a program, which learns from data and makes decisions according to recognisable patterns. In contrast to a person, who can understand context, emotion, and the possible purpose of a situation, AI simply describes data. It does not understand motivations. It cannot sense changes nor take into account intuition or mood. It does not have emotions nor desires. Instead, it is able to analyse massive amounts of data, identify patterns and extrapolate trends, for example. And it can do all this much faster than a human. Let’s go through a few examples.

A person looking a picture of an older, smiling gentleman sitting at a table with other people and holding a glass of wine would ascertain that the gentleman is most likely at some type of celebration. However, AI would simply the visual elements it sees: an older man is sitting down, is holding a glass of wine, is smiling, and there are other people there with him. It cannot assume emotions nor the purpose of the situation. In contrast, a businessman predicting price fluctuations might factor in not only facts and statistics but also the geopolitical situation or market sentiment, or he might rely on his own experience and intuition. AI is able to analyse data, find patterns and extrapolate trends, providing humans a basis for decision-making.

AI is not magic. It is a powerful mathematical tool that learns from individual cases, processes inputs, and generates answers based on probability. It may not understand what it is saying, but it can evaluate what could be said in a similar situation.

How Artificial Intelligence Works

AI is not just a smart program. It is a system comprised of multiple parts. The brain of AI is known as its model. It contains the formulae and rules which the AI has learned. Training data is used to teach the model. This refers to the examples, texts, images or audio samples from which the model learns. The way this model learns from data is determined by its learning algorithm. Another part of AI, called inference, dictates how to use the taught model to react to a new situation, like how to respond to a question.

Neural Networks and Deep Learning

The foundation of AI is its so-called neural network. A neural network is a mathematical model (computational system), which resembles the workings of the brain. Like a brain, a neural network is composed of many interconnected computing nodes (neurons), each of which performs a small number of computations. Also like brain, a neural network processes information in layers. Data such as text, for example, travels through the network, and each layer understands something deeper. The first layer might recognize the words, the second relationships between them, and the last might construct a reply. The more layers a neural network has, the more complex tasks it can handle. This is known as deep learning.

Large Language Models (LLMs)

The large language model (LLM), is a type of neural network that processes text, such as in the form of questions, sentences or conversations. It can learn from a number of different sources—from texts, books, articles, or the web. Whatever training data is presented to it, it will learn. The LLM then generates new responses based on what it has learned.

Let’s take another example. If you were to ask AI, “Who was Jan Nepomucký?,” it will recall all the sentences in which the phrase Jan Nepomucký appears, find the most likely connections and, on that basis, builds its response. AI does not know who Jan Nepomucký was, but it knows the probabilities of what has been said about him and can thus craft a response that a human could probably provide.

How AI Learns

The principle of AI learning is relatively simple:

1. AI receives input data, e.g. several million articles.
2. It reads them and tries to guess other words in the text.
3. If it makes a mistake, the learning algorithm updates the relationships (weights) of the neurons in the neural network.
4. After millions and millions of attempts and corrections, the network learns language patterns, and a model that “knows how to speak” is created.

The fact that this principle is simple does not mean, however, that it is simplistic in its realisation. Executing these steps requires both time and highly powerful computing equipment.

The Three Levels of Artificial Intelligence

The levels of AI demonstrate how close artificial intelligence is to real human intelligence. Experts today normally only refer to narrow AI, are working towards general AI, and are contemplating what will happen if they can succeed in creating a super intelligence. While the levels of AI are a common topic of conversation, they are rarely explained correctly and understandably, so let’s take a brief look at them in more detail.

Narrow AI

Dubbed ANI (Artificial Narrow Intelligence), narrow AI is single purpose. That means that it can fulfil only one class of functions, such as generating texts. It does not understand contexts and it cannot learn anything without new training. Typical examples of this type of AI include most contemporary models such as ChatGPT, Google Translate, etc. Practical uses of ANI are already massive, extending from cars, mobile phones, and translators to complex analytical and expert systems.

General AI

Artificial General Intelligence, or AGI, should be capable of learning, thinking, and understanding like a human. It should be able to manage any intellectual that a human can. It should be able to migrate between disciplines, understand new situations, and even pose its own questions. Many consider AGI as a turning point in the history of humanity. AGI can either significantly move humanity forward or, according to some, endanger it. AGI does not exist yet, but development is already underway. However, experts differ widely in their estimates of when humanity will succeed in creating AGI. Optimists believe that it will come before the year 2030, while the majority assume AGI will not make its real appearance until sometime around 2050.

Super intelligence

Artificial Superintelligence, ASI, is often referenced in sci-fi films and novels, but even in serious discussions by scientists, this AI is intended to be smarter than any human. It should have better memory, faster thinking, more precise logic, and the ability to plan, understand, and create. Though it does not exist yet, it should theoretically be able to deal with scientific problems, including medical or physical.

The levels of AI are not simply theories. Narrow AI (ANI) is quite common and commercially used today. General AI (AGI) constitutes a goal of many companies and would be considered a great hope as well as a real threat. Super intelligence (ASI) remains fixed in the philosophical realm, though its develop is progressing quite quickly, with potential ethical, legal, and security issues already being addressed.

Dispelling Myths

As with all new things, artificial intelligence subconsciously awakens in people uncertainty and fear, but also resistance or even curiosity. There are thus many myths, half-truths, speculations, and conspiracy theories about AI and its abilities or inabilities circulating around the world. But AI is neither a demon nor a miracle. It is simply a powerful tool that reflects the intentions and ethics of those who create and use it. In general:

• AI has no consciousness, emotion, or understanding. It does not understand the meanings of words; it only knows the best way to order them in order to produce the response that a user is likely expecting.
• AI can elicit a feeling that it understands someone, though in actuality, it simply mimics natural language well.
• AI does not have its own goals nor desires. It does not want to, nor can it, do harm on its own. Instead, its responses depend on what data it was trained on and what security mechanisms its creator installed.
• AI will not replace the jobs of most people and will not cause mass unemployment. It will change the character of routine functions, but it will never replace creativity, for example. On the contrary, like any new field, it is generating new positions, such as AI Trainer or Prompt Engineer (specialists who know how to construct questions for AI in order to generate the most precise responses possible).
• AI does not have prejudices, it only repeats what it has learned from data, which was presented to it by humans. If that data contains prejudices, then those will show in the AI’s responses.
• AI is not a fad; it is a real, technological structural change similar to the invention of the steam engine, electricity, or the Internet.
• AI does not always provide correct nor incorrect results. AI only provides the most likely response depending on the data on which it was trained. In other words, it can hallucinate, which is to say it can create completely nonsensical responses or half-truths, as if making up information. It all depends on the training and set-up.

AI is a tool, not an authority. And just like any tool, it can be either a faithful servant or an evil master. It depends on who is in control. A correctly trained AI can help large companies save massive amounts of time and effort, such as that spent on routine activities.

What Does AI Actually Need?

For AI to function well, it primarily needs adequate computing power, sufficiently large operating memory, and available storage capable of holding all its data.

In comparison to typical business applications, AI requires enormous computing power. Servers with typical processors are not enough even for a small corporate AI, which will also require GPU processors (graphics cards), TPUs (tensor processors), or specialized AI chips. All this is accompanied by high electricity consumption and increased demands on the environment, especially due to cooling needs and air conditioning.

In practice, even a small corporate AI may require several high-performance GPU processors, such as NVIDIA A100, H100, or RTX 4090, RAM memory in the order of hundreds of gigabytes, sufficiently and ideally redundant power supply, reliable network connection and powerful cooling. More extensive AI models are practically unrealisable in typical companies and certainly not in home-like conditions.

AI Cloud: Artificial Intelligence as a Service

Despite this, there is no need to throw in the towel just yet. Even extensive AI solutions are made possible, simple, and with significantly lower operational costs via cloud. One does not need to purchase extremely extensive servers with GPU processors and other hardware. In the cloud, AI can be launched with almost no delay, with both empty models and models pre-loaded with the most frequently used applications available. Unlike an on-site solution, the cloud easily adapts to increasing workloads or expanding AI capabilities. Open-source interfaces (e.g. Open AI API) provide the possibility of seamless integration with other systems, and not only corporate ones.

Provider specialists care for the operation and maintenance of cloud infrastructure. Most often, you pay either for the operation of the AI models used or for the volume of processed data, measured in tokens. For companies who are looking to begin using AI or who would at least like to try it out, AI cloud offers an invaluable opportunity to experiment with high-demand AI models without having to invest in expensive infrastructure.

AI as a service is a practical, scalable, nearly immediately available and, in the vast majority of cases, a significantly cheaper solution than its own infrastructure.

How to Turn Empty AI into an Expert?

Today, nearly everything from automatic responders appears AI-generated, whether from various smart chatbots or extremely intelligent legal or medical knowledge systems. But what does a company or organisation really need in order to fill an “empty” AI and turn it into an Internet expert?

Empty AI is typically a language model that knows how to speak, understand language, and understand common expressions but is not filled with any knowledge or information. Think of it like an executive trainee who has just joined the company but knows nothing about his job.

Knowledge can be uploaded to AI in a variety of ways, with one of three paths typically used:

1. Finetuning – the AI learns a new model based on relevant data using the aforementioned method of trial and correction. Finetuning is very computationally demanding and expensive, making it suitable primarily for large companies with their own AI team.
2. Embedding – AI is not hard-coded, but the data is encoded into a knowledge base within which the AI then dynamically searches each query. Embedding is modern, flexible, and currently the most common means of quickly launching a corporate AI.
3. Prompt Engineering – AI learns from a carefully prepared prompt, such as “You are an expert in kicking balls, answer according to these rules…”. Prompt engineering is suitable for smaller companies or companies with a low response capacity, but it is important that the prompt is prepared carefully. Otherwise, the AI’s answers may annoy the client, rather than help or guide them.

Teaching AI is far from just a technical matter. It also requires very thorough, careful preparation and a step-by-step procedure from the trainer. Primarily, it is necessary to collect, clean, and correctly structure the inserted data, design the AI personality (i.e. in what tone it will communicate with clients), choose the appropriate learning or insertion approach, and design and implement a method of connecting the AI with your own systems. Without correctly structuring data, AI can hallucinate at any point when attempting to respond.

It is always important to keep in mind that AI is not able to understand contexts, such as business strategy, on its own. This you will need to upload into it. And no AI can replace communication between people. It can significantly reduce time informing clients, speed up response times, filter requests or even analyse client needs or problem descriptions, and serve as a foundation for improving communication.

An important component of AI implementation is testing and securing it. You will need to verify that the AI does not give false or overly vague responses and set certain limits, such as when to switch to a human colleague, for example. The more thorough an AI is prepared and tuned, the more satisfied its users, clients, or corporate or institutional partners will be. High-quality and well-functioning AI is, and increasingly will be in the future, a part of a company’s overall image.

Practical Uses of AI

While general AI (AGI) remains a thing of the future, narrow AI is experiencing a rapid and practical boom. In the next two to five years, we can expect a huge increase of its usage in literally all areas of business and everyday life, especially:

• Expansion into common business processes,
• Usage in industry and technical fields,
• Deployment in administration, public, and state sectors,
• Personalisation of content and communication,
• Education, medicine, science, and many other areas

An even greater emphasis will be placed on security, auditing, and ethics. AI must be transparent, not an invisible black box. This will mainly concern visibility (i.e. why the AI told me or recommended me this or that), auditability, and ethics. AI must also not discriminate or worsen prejudices and must be compliant with legislation such as NIS 2, GDPR, and other new laws directly relating to AI.

In the near future, AI will be a common component of offices, production halls, and companies. These are not sci-fi thinking machines but practical tools that will speed up work, make information more accurate, and help people focus on the essentials. The key is to know the goal, have the data, and choose the right approach.

If you would like to know more about the fields using AI, about AI cloud and AI services, visit the České Radiokomunikace website AI Cloud  to discover more.