what is AI: What is artificial general intelligence?

 

What is artificial intelligence?

The definition of the term AI

What is artificial intelligence? Simply put, AI is the attempt to transfer human learning and thinking to the computer and thus give it intelligence. Instead of being programmed for every purpose, an AI can find answers and solve problems on its own.

The aim of AI research has always been to understand the function of our brain and our mind on the one hand and to be able to artificially recreate it on the other.

Artificial Intelligence in Science Fiction and Reality

The dream of artificial intelligence is older than the computer itself – we know it from books and films, be it “Frankenstein’s monster” or artificially created people like the homunculus from the Middle Ages.

We have encountered the term "artificial intelligence" primarily in science fiction and usually refers to robots or computers that can think and act independently. Whether for good, like the android "Data" from "Star Trek" or for evil, like the computer HAL from the film "2001: A Space Odyssey". In art, they are a means of asking questions about ourselves: What makes a person? What is intelligence?

When we talk about AI in today's world, however, it has little to do with what we know from films and books. In real life, we have only encountered AI in a hidden way - when new products are recommended to us on Amazon, when people are automatically recognized in photos, or when we chat with "Alexa" or "Siri" on our cell phone. At the beginning of 2023, ChatGPT was the first AI that we actively call up to solve everyday or work problems. But is ChatGPT on the same level as "Data"?

Defining the concept of AI

So what is AI? This is difficult to answer clearly. There is no universally accepted definition of artificial intelligence - and the concept of intelligence itself has not yet been clearly defined.

That's why we're trying to approach the term differently: In German, a distinction is made between strong AI and weak AI when it comes to the definition of AI. Simply explained: Strong AI means what we know from science fiction. A machine that can solve problems of a general nature - in other words, any question that is asked of it. It is still pure fantasy and will remain so for decades. In English, the term AGI - Artificial General Intelligence - is often used here.

Weak AI, on the other hand, is what we deal with in everyday life: these are algorithms - and that is what an AI is, a very complex algorithm - that can handle specific tasks , the solutions to which they have previously learned independently . Even if ChatGPT can answer all questions, it cannot produce images or videos. A weak AI has no consciousness of its own and shows no understanding. (Well, it may share the latter with some strong AIs like the Terminator).

What defines an AI?

From now on, we will only talk about weak AI, since it is ultimately the only form that is commercially relevant today – we find weak AI in everyday life in our cell phones and computers.

So what distinguishes an AI from a simple program? Typically, a programmer writes code in a language of their choice, which consists of a set of arbitrarily complex instructions:

• If this, then that
• For example: When the user presses “ Send ”, send the email to server X

Such a system is also called rule-based. With artificial intelligence, the programmer does not specify every single program step, but writes an algorithm that is able to independently adapt its own parameters to a specific problem. An AI does not usually write its own program code (although there are already initial approaches here) but changes certain parameters within its code in order to find a general pattern in data, derive rules and then apply these to new data. 

Why is this important? Because certain problems are so complex that it is impossible to write code for them by hand. One example is image recognition, which is used in social media such as Facebook: No programmer in the world can write a set of instructions that always recognizes what I look like, regardless of whether the photo was taken at night, on the beach or in the car - in a rule-based system this would be completely impossible, because the programmer would have to know and be able to describe all possible images in advance.

A programmer now teaches an AI how to recognize people, but not how to recognize me . The AI ​​doesn't know every picture of me either, but it can learn what I look like from a number of existing pictures and then apply this rule to new pictures and recognize me.

And not just with me, but with billions of faces in fractions of a second. An AI is therefore able to deal with previously unknown data, find patterns or derive actions from them . It learns independently from the data it has - but what it learns is determined in advance by humans who design the AI. Humans program the AI, but the AI ​​learns independently how to carry out the task programmed into it. AI is therefore far more powerful than rule-based systems because it can react to - to a certain extent - previously unknown situations and learn from experience.

What can an AI do?

The potential uses of such AI systems are gigantic and not yet clear to most people. They are currently revolutionizing our economy. The Federal Network Agency, for example, expects value creation of 430 billion euros by 2030 from AI alone, while the market study by Allied Market Research expects the global market size for AI technologies to be 1.5 trillion euros by 2030.

AI is able to extract information from data that a human could never grasp, perhaps because it is too numerous or the underlying patterns are too complex - but which already exist.

Imagine if YouTube employees had to manually watch every uploaded video and check whether it contains banned or stolen content. 500 hours of material are uploaded to the platform every minute. The company would need 90,000 employees alone, watching videos non-stop for 8 hours a day, to keep up with the viewing! An AI does this during the upload process, almost in real time.

Artificial intelligences like these are very good at capturing what is known as unstructured data . These are, for example, images, videos or audio recordings - data that cannot be easily searched because it does not have a uniform form, unlike, for example, a table generated from sensor measurement data. A conventional search algorithm (such as when you type CTRL+F on this website) can find the title of an image (a structured piece of data), but not whether Susie Mustermann is depicted in the image - this information is not found anywhere, it is part of the image content. An AI can do that.

Of course, AI is also used to sort structured data and search for patterns. The current boom in AI is taking advantage of the fact that unstructured data is much more common: it makes up around 80 percent of all data and has only been available in such large quantities for a few years - with the rise of the Internet, Industry 4.0 and the mass availability of (cloud) storage. Many companies are unaware of the treasures of data they have and the potential for value creation that they have. Be it machine data, audio recordings of customer phone calls or recordings of transport routes. You will read a few examples of this later. It was only the mass availability of data in conjunction with the massive progress in computing speed that led to AI being usable on a large scale in recent years. One example: ChatGPT was trained with 300 billion words or parts of words from all over the Internet.

The Google AI "AlphaGo" became famous in 2016 when it defeated the world's best player in the board game Go with game strategies that were previously unknown and has since changed the way people approach the game . A new version, MuZero , is even able to learn the rules of a game on its own and optimize them. This ability to learn on its own means that the AI ​​can potentially be used in many areas - and is much more powerful than its predecessors, which previously had to be reprogrammed for every purpose.

What can't an AI do?

AI is not a general problem solver - not yet. It can process data incredibly well and recognize patterns, but it cannot understand them. Artificial intelligence has no "common sense". If it comes to wrong conclusions due to insufficient data or poor programming, it does not recognize this (see section "Artificial intelligence and humans"). It can only provide answers to the specific questions it was programmed to answer.

Examples of AI projects

AI has long since found its way into our everyday lives. The example of facial recognition on social networks is one among many. Another is voice assistants on our cell phones - we use Siri, Alexa and Co. as a matter of course in our everyday lives. Translators like Deepl can translate our words almost perfectly into other languages ​​in a matter of seconds. ChatGPT reached 1 million users within five days in November 2022 and is now likely to be in the three-digit million range.

When we surf the internet every day, the advertisements we see are selected by artificial intelligence that tries to show us the most attractive product based on our interests and activities. We encounter these so-called "recommendation systems" everywhere online: Amazon, Google, Netflix, Facebook. They are a very powerful system because more and more media are vying for our attention, there is more to discover online than we can ever perceive in our lives. Computers therefore have to make a pre-selection for us - and over time AIs learn to understand us better and better and to play our preferences (against us).

But AI is also moving into our everyday lives outside of the online world. Robot vacuum cleaners clean our floors and use algorithms to recognize their surroundings. Navigation systems find the optimal route. The greatest progress at the moment is being made by autonomous vehicles , which have collected millions of test kilometers on the road - even if they are still years away from widespread use. After all, Mercedes received model approval for autonomous driving on the highway up to 60 km/h in 2021 , making it the first manufacturer to reach level 3 of 5 on the autonomous driving scale.

A few more concrete examples :

• The Bremen start-up JUST ADD AI is working with the football club Werder Bremen to analyze reports from talent scouts using AI to find new football stars.
• Google ( Waymo ) is already testing the use of autonomous vehicles in practice – although currently with a driver as a last resort.
• Paypal uses AI to detect fraud attempts in the payment system .
• The Telekom AI “Tinka” processes 120,000 chat requests every month , it can resolve 80 percent of all customer requests and refers a fifth of them to human employees.

Even though AI has so far been used primarily by large corporations, medium-sized companies can also benefit from it. One example is wind power: the PiB research project aims to predict the icing of wind turbines . The medium-sized Bremen wind farm operator wpd windmanager is one of the companies involved.

Different types of AIs

The term AI covers a wide range of very different technologies that have been researched over the past 70 years. The examples and methods described so far relate to a specific area of ​​AI research, machine learning (ML). It stands for learning from experience. We have limited ourselves to this area so far because ML is the most relevant form of AI for companies in commercial use today and many of the latest researches come from this, whether it is about language (Large Language Models LLM) or image generation (generative AI such as Midjourney ) or image processing (using deep neural networks). Read more about this in our article on neural networks.

But there are also completely different approaches. These include so-called expert systems, which rely on a knowledge base compiled by experts in order to draw conclusions using certain rules - they are more or less the opposite of "learning from experience". The most famous example of an expert system is the chess computer "Deep Blue", which defeated the world chess champion Gary Kasparov in 1997.

Both approaches are often classified into different categories - symbolic and sub-symbolic AI. A symbolic AI comes to results in a comprehensible way by combining symbols (i.e. words, letters, numbers, etc.) according to pre-programmed rules in order to draw a conclusion. An example of this would be classical logic:

Symbol 1: "All men are mortal"
Symbol 2: "Socrates is a man"
Conclusion: "Socrates is mortal"

A subsymbolic AI, on the other hand, does not arrive at a result by combining symbols and rules. Instead, it breaks down information into mathematical formulas and optimizes these formulas until it produces the desired result. It is not possible to trace the path to the result directly from the formula afterwards. This is experiential learning - machine learning. The results are probabilistic, they are based on a probability. ChatGPT calculates what the most likely next word is when it is asked to complete the sentence "My favorite ice cream flavor is..." It does not know the answer, it guesses it. But it does so very well. For this reason, these types of AI can also give good-sounding but practically wrong answers, because they do not know what is right or wrong, they can only estimate it .

Both AI approaches are not mutually exclusive - there are efforts to combine them, or to use elements of one in the other. Read more about this in: Understanding the difference between Symbolic AI & Non Symbolic AI. One concrete attempt is the European research project Project MUHAI .

Using Artificial Intelligence in Business

The use of AI for their processes can be attractive for companies today. They promise great efficiency gains for certain problems. Companies should therefore ask themselves the question: What can I really achieve with AI? The first thing to look at is your own data - what data is already available in the company, what could still be recorded? An AI can draw conclusions from it that were previously not possible - for example because the analysis effort would be too high for humans or there was no way to get the right answers - or no one had thought of generating data from it at all, for example from an "old" machine that is not yet connected to a computer.

What companies can definitely expect: As soon as a job is found for the AI, it will do it better than any human. Not only is it faster, but the error rate continues to fall as the wealth of experience continues to grow. According to the company, the Google AI "Lyna" (LYmph Node Assistant) can detect breast cancer in images with 99 percent probability , a value that doctors dream of.

It is important to find a specific application, because AIs are not (yet) general problem-solving machines. One requirement would be, for example: "We want to check the quality of workpieces from the assembly line in real time using camera analysis, without having to resort to manual sampling."

Like all profound innovations, the successful implementation of AI in a company takes time. Roland Becker, managing director of Bremen-based AI expert JUST ADD AI , estimates that the return on investment for a project is between 12 and 18 months. In order for a project to be successful, both the available data and the relevant knowledge are necessary. In addition to hiring their own experts, SMEs in particular can work with partners in research projects ( such as the Bremen BIBA ). They introduce the topic carefully and enable you to get to know the new technology with relatively little expenditure of resources.

Training intelligent networks requires a lot of computing power, which can be achieved either through an investment or by renting cloud capacity - a partner who already has the capacity for this makes it much easier and cheaper.

So – do medium-sized companies now absolutely have to rely on AI in order to survive?

Small and medium-sized companies often find it difficult to adapt new technologies quickly. They lack the resources of large corporations for experiments and the agility of start-ups without ongoing costs. The same is true of AI: According to a survey by the industry association Bitkom, the majority of companies with fewer than 500 employees have so far refrained from investing in AI . The reasons are a lack of human resources, time, but often also competitive pressure. Many companies are still taking a wait-and-see approach.

So is it better to wait? The answer is clear: No. Large Language Models such as ChatGPT, Claude or Mistral offer a simple approach that anyone can learn in a few seconds ("prompting"). They cost little and quickly create noticeable effects for certain tasks, especially in the office area. In order to stay on the ball, every company should therefore look into the possibilities of these models. They don't have to be computer experts, anyone can experiment with them themselves.

If the AI ​​is also to improve or further develop the business model itself, more effort is required. For small companies, artificial intelligence is an investment at this point at the latest, because now they need developers who can also handle the systems technically. The first question should therefore be: How could an AI increase my sales? How could an AI reduce my costs and improve services? How can my customers benefit? It helps to familiarize yourself with the technology in order to get an overview of the possibilities. Free information offers, such as those from the Mittelstands-4.0 centers in Germany, help to accumulate knowledge. If a use case or an idea for a use is found, local partners and funding will help to implement it.

Although the large cloud companies such as IBM, Google and Amazon also offer AI solutions, these can quickly become oversized, especially since experts are still needed to implement them successfully. And specialists are rare, especially in the field of AI. Anyone who currently does not see a use for AI should stay on the ball: because one day the time will come when competitors will rely on it and by then it will be too late. And given the speed at which AI is currently evolving, this time will come sooner rather than later.

At the same time, the costs and resources required for the use of AI are falling rapidly. For several years now, there have been so-called frameworks that provide the basic tools for quickly setting up your own AI networks - TensorFLow and PyTorch are the most common. This means that even small companies can set up AI - the five-man company INnUP in Bremen is a perfect example of this . At the same time, work is also being done on systems that enable laypeople without programming experience to use AI.

And one more piece of advice: data is the oil of AI. Those who start collecting, storing and cataloging data today will benefit from it tomorrow.

Artificial Intelligence and Humans

Like many new technologies, AI also stirs up fears. A famous study by the University of Oxford in 2013 analyzed that 47 percent of all US jobs were at risk from automation, a significant proportion of these from AI . Such figures stir up fears that certainly lead to real actions: Waymo, the Google subsidiary for automated driving, reports that its test vehicles have been attacked with knives and stones on several occasions. So is AI a danger to people? A bitkom survey paints a mixed picture : 62 percent of Germans see AI primarily as an opportunity, 35 percent as a danger. A survey of managers also found that 42 percent of them observed reservations among their workforce.

The truth lies somewhere in the middle. AI will undoubtedly take over human labor, and when it does, it will do so entirely - that is, no human will be needed for this one specific task. These are usually tasks that are not very fun, monotonous and repetitive in nature: watching surveillance videos, answering standard queries on the phone, searching through documents.

In many areas, however, AI will take on more of an assistant role in the foreseeable future. It will support doctors in coming to the right conclusions or be part of a work chain - for example in the fields, where agricultural robots can take on subtasks autonomously or semi-autonomously, while other tasks continue to be carried out by humans.

At the same time, however, new jobs will be created that will be supported by the innovative AI business models. People will then have more time to use their labor for new tasks because they will be working with AI. Lawyers could spend more time with clients instead of spending hours poring over files. It is also clear that more education is needed to prepare people for their new tasks and to give them the skills to work with AI systems.

A 2023 study by the World Economic Forum  estimates that 89 million jobs will be displaced by computers, while 63 million will be created.

And, to be honest, we don't really have a choice. AI has long since become part of everyday life and almost everyone uses it today, whether consciously or unconsciously - in their cell phones, when making transfers or for navigation. It will still be some time before we encounter AI everywhere, but that time will come sooner rather than later, because as soon as an area benefits from AI, it will have massive advantages over its human counterparts and thus push them out of the market or open up new fields of activity for them.

Nevertheless, it is important to talk about it and ask ourselves where ethics lie in the machine. This is not just about responsibility ("Who is to blame if the machine has an accident?"), but also the question of how we want to organize work in the future.

The Natural Stupidity in Artificial Intelligence

AIs are made by humans - and are therefore subject to a natural problem: an intelligence that imitates humans is also subject to their mental limitations. One of these is bias .

An example: In 2014, Amazon's AI experts developed an AI that automatically evaluated and sorted application documents. To do this, they trained the neural network with applications from the past ten years. Once the AI ​​was trained, they discovered that the algorithm only selected those from men among new applications. The reason: There were an above-average number of men among those hired earlier, as is common in the tech industry. The AI ​​created the rule from this: only hire men. (Source) The error lay in the selection and preparation of the data. Amazon ultimately abandoned the experiment, and applications continued to be searched manually.

The example shows that when designing artificial intelligence, humans must attach great importance to the selection of representative data - and be aware that they may already be biased by selecting and processing the data. This dilemma cannot be easily solved and must be taken into account in every AI design. This is another reason why it is worth taking an outside perspective and working with a partner and expert in AI.

Ultimately, every AI is programmed by a human – and we know where our intelligence begins and ends.

Finally, a short summary of what artificial intelligence is:

• AI is the attempt to transfer human learning and thinking to the computer
• Strong AI, i.e. general problem-solving machines, belong to the realm of science fiction, while weak AI is finding increasingly widespread use in today's world, whether in mobile phones, websites, social media or self-driving cars
• AIs are valuable wherever a lot of data can be analyzed and searched for patterns
• Machine learning is currently the most commercially important subfield of AI
• AIs require data as a basis, which can be numbers, images, videos or sounds
• AIs can process data better, more accurately and faster than humans, but they cannot understand it
• AIs are programmed ("trained") only for very specific purposes and must be retrained for other purposes
• AIs will take over tasks from humans, but at the same time will also create new business areas and thus jobs
• AIs cannot understand the data; if they are fed with incorrect data, they deliver incorrect results

Now that we have dealt with the basics, let's get down to business: Find out what buzzwords like machine learning, neural networks or deep neural networks are all about and how an AI "thinks".