What is Artificial Intelligence?
 |
| What is Artificial Intelligence? |
Artificial Intelligence in Science Fiction and Reality:
The dream of artificial intelligence predates the computer itself; we encounter it in books and movies, whether it’s Frankenstein’s monster or artificially created beings like the homunculus from medieval times. In science fiction, AI usually refers to robots or computers capable of independent thought and action. Characters like the android Data from "Star Trek" or the malevolent HAL from "2001: A Space Odyssey" serve as tools to explore questions about human nature and the essence of intelligence.However, when we talk about AI in today’s world, it has little to do with the depictions we see in films and books. In real life, AI has been present in hidden forms, such as product recommendations on Amazon, automatic recognition of people in photos, or interactions with assistants like Alexa and Siri on our phones. In early 2023, ChatGPT became the first widely popular AI that we actively use to solve everyday or professional problems. But is ChatGPT on the same level as Data?
Defining the AI Concept:
So, what exactly is AI? This is difficult to answer definitively because there is no universally accepted definition of artificial intelligence, just as there is no clear definition of intelligence itself.To approach the term differently, in German, we distinguish between strong AI and weak AI when defining AI. Simply put, strong AI refers to what we know from science fiction: a machine capable of solving general problems and answering any question posed to it. This remains purely fictional and will likely stay so for decades. The term AGI (Artificial General Intelligence) is often used in English for this concept.
On the other hand, we encounter weak AI in our daily lives: these are algorithms essentially, a very complex set of instructions that handle specific tasks they have learned to solve independently. Even though ChatGPT can answer various questions, it cannot generate images or videos. A weak AI has no self-awareness and shows no understanding (which, arguably, it shares with some strong AIs like the Terminator).
Strong AI Belongs to Science Fiction, While Weak AI is Part of Our Everyday Lives
What sets AI apart?
In this discussion, we will only address weak AI, as it is the only commercially relevant form today found in our everyday devices like smartphones and computers.What sets AI apart from a simple program? Typically, a programmer writes code in a chosen language, consisting of a set of complex instructions, such as:
- If this, then that.
- For example: If the user presses "Send," transmit the email to server X
The Importance of AI:
Why is this significant? Some problems are so complex that it is impossible to write a code manually. Take image recognition used in social media like Facebook, for example, no programmer can write a set of instructions that consistently recognizes a person’s face in all conditions night, beach, or car photos. In a rule-based system, this would be impossible, as it would require the programmer to foresee and describe all possible images.Instead, a programmer teaches an AI to recognize people but not specifically how to recognize me. The AI does not know every picture of me but can learn from a set of existing photos and apply this understanding to new images to identify me. This capability extends to billions of faces in mere seconds. An AI can handle unknown data, find patterns, or derive actions from them. It learns autonomously from the data provided, though its learning scope is predetermined by humans who design the AI. Humans program the AI, but the AI independently learns how to perform its programmed task, making it far more powerful than rule-based systems by being able to react to previously unknown situations and learn from experience.
Capabilities of AI:
The potential applications of AI systems are vast and not yet fully recognized by most people. They are currently revolutionizing our economy. The German Federal Network Agency estimates that AI will generate 430 billion euros in added value by 2030. A market study by Allied Market Research projects a global market size for AI technologies of 1.5 trillion euros by 2030.AI can extract information from data that humans could never comprehend due to the volume or complexity of the underlying patterns, which already exist. Imagine if YouTube employees had to manually watch and check every uploaded video for prohibited or stolen content. Every minute, 500 hours of material are uploaded to the platform. The company would need 90,000 employees watching videos nonstop for 8 hours a day to keep up! AI can accomplish this during the upload process, almost in real-time.
Handling Unstructured Data:
Unstructured data constitutes most data and can now be automatically analyzed by AI for the first time. AI excels at capturing unstructured data, such as images, videos, or audio recordings data that cannot be easily searched due to the lack of a uniform format, unlike a table generated from sensor measurements. A conventional search algorithm (like using CTRL+F on this webpage) can find the title of an image (a structured datum) but not determine if Susie Mustermann is depicted in the image this information is part of the image content. An AI can do this.Of course, AI is also used to sort and search for patterns in structured data. The recent surge in AI utilization capitalizes on the prevalence of unstructured data, which makes up about 80% of all data and has only become available in large quantities in recent years, with the rise of the internet, Industry 4.0, and the widespread availability of (cloud) storage. Many companies are unaware of the treasure troves of data they possess and the value creation potential within them, such as machine data, audio recordings of customer calls, or transport route logs. Some examples will be discussed later. The mass availability of data, combined with significant advances in computing speed, has made large-scale AI applications possible in recent years. For instance, ChatGPT was trained on 300 billion words or word fragments from across the internet.
A well-known example is Google’s AI, AlphaGo, which in 2016 defeated the world’s best player in the board game Go using previously unknown strategies, changing how humans approach the game. A newer version, MuZero, can learn and optimize game rules independently. This self-learning ability makes AI potentially applicable in many fields, far surpassing its predecessors, which had to be reprogrammed for each new task.
What Can't AI Do? Limitations of AI:
AI is not yet a general problem solver. While it can process data and recognize patterns exceptionally well, it does not truly understand them. Artificial intelligence lacks "common sense" or human intuition. If it draws incorrect conclusions due to insufficient data or poor programming, it does not realize its mistake (see the section "Artificial Intelligence and Humans"). It can only provide answers to the specific questions it was programmed to address.Examples of AI Projects:
AI has already become a part of our everyday lives. Facial recognition on social networks is just one example. Another is voice assistants on our phones using Siri, Alexa, and others has become second nature. Translators like DeepL can almost perfectly translate our words into other languages in seconds. ChatGPT reached 1 million users within five days of its launch in November 2022, and the number is now likely in the hundreds of millions.When we surf the internet, the advertisements we see are selected by AI, which tries to present the most attractive products based on our interests and activities. These "recommendation systems" are ubiquitous online: Amazon, Google, Netflix, and Facebook. They are incredibly powerful because an ever-increasing amount of media competes for our attention, and there is more to discover online than we could ever perceive in a lifetime. Computers must make preliminary selections for us, and AI systems learn over time to understand us better and better, tailoring our preferences (often against our favor).
AI is also entering our offline lives. Robot vacuum cleaners use algorithms to recognize their surroundings and clean our floors. Navigation systems find the optimal routes. The most significant progress is in autonomous vehicles, which are accumulating millions of test kilometers on roads, though widespread use is still years away. In 2021, Mercedes received model approval for autonomous driving on highways at speeds up to 60 km/h, making it the first manufacturer to achieve Level 3 of 5 in autonomous driving.
Some Specific Examples:
- The Bremen start-up JUST ADD AI is collaborating with the football club Werder Bremen to use AI to analyze talent scout reports to discover new football stars.
- Google (Waymo) is testing the use of autonomous vehicles in real-world settings, though currently with a driver for safety.
- PayPal uses AI to detect fraud attempts in its payment system.
- The Telekom AI "Tinka" handles 120,000 chat queries monthly, resolving 80% of all customer inquiries and referring the remaining 20% to human employees.
Different Types of AI - Various AI Technologies
The umbrella term "AI" encompasses a wide range of distinct technologies developed over the past 70 years. The examples and methods previously described pertain to a specific area of AI research known as Machine Learning (ML), which focuses on learning from experience. We have focused on this field because ML is currently the most commercially relevant form of AI for businesses, driving many recent advancements, whether in language models (Large Language Models or LLMs), image generation (generative AI like Midjourney), or image processing (via Deep Neural Networks). For more on this topic, see our article on Neural Networks.However, there are entirely different approaches within AI. One such approach involves expert systems, which rely on a knowledge base compiled by experts and use specific rules to conclude. These systems essentially do the opposite of learning from experience. The most famous example of an expert system is the chess computer "Deep Blue," which defeated world chess champion Garry Kasparov in 1997.
These approaches are often categorized into symbolic and subsymbolic AI. Symbolic AI arrives at conclusions transparently transparently by combining symbols (such as words, letters, and numbers) according to predefined rules. A classic example would be:
- Symbol 1: "All humans are mortal"
- Symbol 2: "Socrates is a human"
- Conclusion: "Socrates is mortal"
In contrast, subsymbolic AI does not reach conclusions through symbol and rule combinations. Instead, it translates information into mathematical formulas and optimizes these formulas until the desired result is achieved. It is not possible to directly trace the path to the result from the formula afterward. This is experiential learning Machine Learning. The results are probabilistic, based on likelihood. For instance, when ChatGPT completes the sentence "My favorite ice cream flavor is ...," it calculates the most likely next word. It does not know the answer; it guesses it, albeit very accurately. This is why these types of AI can sometimes produce plausible but incorrect answers, as they don't truly understand right or wrong they can only estimate.
Both AI approaches are not mutually exclusive there are efforts to integrate them or use elements of one within the other. For more on this, read: "Understanding the Difference Between Symbolic AI & Non-Symbolic AI." A concrete attempt at this is the European research project MUHAI.
Companies can expect that once AI is assigned a task, it will perform it better than any human. Not only is AI faster, but its error rate decreases as its experience grows. Google's AI "Lyna" (LYmph Node Assistant), for example, can reportedly detect breast cancer in images with 99% accuracy a level doctors can only dream of.
Finding a specific use case is crucial, as AI is not yet a general problem-solving machine. An example goal might be: "We want to inspect parts on the production line in real-time using camera analysis without relying on manual sampling."
Like all profound innovations, successfully implementing AI in a company takes time. The return on investment for an AI project is typically estimated at 12 to 18 months, according to Roland Becker, CEO of Bremen AI expert JUST ADD AI. For a project to succeed, quality data is essential, along with the necessary expertise. For SMEs, partnering with research projects (such as with Bremen's BIBA) can be particularly beneficial, providing a gradual introduction to the new technology with relatively low resource investment.
Training intelligent networks requires significant computing power, which can be achieved either through investment or by renting cloud capacities. A partner who already has this capacity makes the process much easier and more cost-effective.
So, should they wait? The answer is clear: No. Large Language Models like ChatGPT, Claude, or Mistral provide easy access that anyone can learn in a matter of seconds ("prompting"). They are low-cost and can quickly deliver noticeable effects in certain tasks, especially in office environments. To stay competitive, every company should explore the potential of these models. This doesn't require computer experts; anyone can experiment with them.
However, improving or evolving the business model with AI demands more effort. For small companies, AI becomes an investment at this point, as it requires developers who can technically handle the systems. The first question should be: How could AI increase my revenue? How could AI reduce costs and improve services? How can my customers benefit? Understanding the technology and its possibilities is essential. Free information resources, such as those offered by the Mittelstand 4.0 centers in Germany, can help accumulate knowledge. Once a use case or idea for an application is identified, local partners and funding can assist in implementation.
Although large cloud companies like IBM, Google, or Amazon also offer AI solutions, these can be quickly oversized and still require experts for successful implementation. And AI experts are currently scarce. Those who do not yet see a use case for AI should stay engaged: competitors will adopt it, and by then it might be too late. With the rapid advancement of AI, this moment will come sooner rather than later.
At the same time, the costs and resources needed to implement AI are decreasing rapidly. In recent years, frameworks such as TensorFlow and PyTorch have emerged, providing the basic tools to quickly set up AI networks. Even small companies can now implement AI, with the five-person Bremen-based company INnUP being a perfect example. Moreover, systems are being developed to enable AI use by non-programmers.
Another piece of advice: data is the oil of AI. Those who start collecting, storing, and cataloging data today will benefit tomorrow.
The truth lies somewhere in between. AI will undoubtedly take over some human tasks completely meaning that for these specific tasks, humans will no longer be needed. These are often tasks that are monotonous and repetitive: watching surveillance videos, answering standard inquiries over the phone, and searching through documents.
In many areas, however, AI will likely play a more supportive role for the foreseeable future. It will assist doctors in reaching accurate conclusions or be part of a workflow such as in agriculture, where robots can autonomously or semi-autonomously handle certain tasks while humans continue to perform others.
Simultaneously, new jobs will emerge, driven by innovative AI business models. People will have more time to devote to new tasks because they will be collaborating with AI. For example, lawyers could spend more time with clients instead of spending hours searching through files. It is clear, therefore, that more education is needed to prepare people for their new roles and to equip them with the skills to work with AI systems.
A study by the World Economic Forum in 2023 predicts that while 89 million jobs will be displaced by computers, 63 million new jobs will be created.
To be honest, we don’t have a choice. AI has already been integrated into our daily lives, and almost everyone uses it in some form, whether knowingly or unknowingly through smartphones, bank transfers, or navigation systems. It will take some time before AI becomes ubiquitous, but that moment is approaching sooner rather than later. Once a sector benefits from AI, it will gain significant advantages over its human counterparts, either pushing them out of the market or opening up new areas of work for them.
An example: In 2014, Amazon's AI experts developed a system to automatically evaluate and sort job applications. They trained the neural network using resumes from the past decade. Once trained, they discovered that the algorithm started favoring resumes from men when presented with new applications. The reason was that historically, the tech industry had a higher proportion of male hires. The AI inferred a rule: only select male applicants. (Source) The flaw lies in the selection and preparation of data. Amazon ultimately scrapped the experiment, opting to continue the manual review of applications.
This example highlights the importance for humans to meticulously select representative data when designing artificial intelligence and to be aware that biases may already exist in the data chosen and prepared. This dilemma is not easily solved and must be carefully considered in every AI design. This underscores the value of external perspectives and collaboration with AI experts and partners.
Ultimately, every AI is programmed by humans and where our intelligence begins and ends, we are keenly aware.
Both AI approaches are not mutually exclusive there are efforts to integrate them or use elements of one within the other. For more on this, read: "Understanding the Difference Between Symbolic AI & Non-Symbolic AI." A concrete attempt at this is the European research project MUHAI.
Examples of Artificial Intelligence in SMEs:
Our retrospective article on the Bremen Innovative website provides numerous examples of AI use in small and medium-sized enterprises (SMEs).Implementing AI in Business:
For companies, integrating AI into their processes can be highly attractive today, offering significant efficiency gains for specific problems. Companies should ask themselves: What can I realistically achieve with AI? The first step is to examine existing data within the company and consider what additional data could be collected. AI can conclude this previously impossible data either because the analysis effort was too great for humans or because there was no way to obtain the right answers or no one thought to generate data from, say, an "old" machine not yet connected to a computer.Companies can expect that once AI is assigned a task, it will perform it better than any human. Not only is AI faster, but its error rate decreases as its experience grows. Google's AI "Lyna" (LYmph Node Assistant), for example, can reportedly detect breast cancer in images with 99% accuracy a level doctors can only dream of.
Finding a specific use case is crucial, as AI is not yet a general problem-solving machine. An example goal might be: "We want to inspect parts on the production line in real-time using camera analysis without relying on manual sampling."
Like all profound innovations, successfully implementing AI in a company takes time. The return on investment for an AI project is typically estimated at 12 to 18 months, according to Roland Becker, CEO of Bremen AI expert JUST ADD AI. For a project to succeed, quality data is essential, along with the necessary expertise. For SMEs, partnering with research projects (such as with Bremen's BIBA) can be particularly beneficial, providing a gradual introduction to the new technology with relatively low resource investment.
Training intelligent networks requires significant computing power, which can be achieved either through investment or by renting cloud capacities. A partner who already has this capacity makes the process much easier and more cost-effective.
Should SMEs Invest in AI Now to Survive?
Small and medium-sized enterprises (SMEs) often struggle to quickly adopt new technologies. Unlike large corporations, they lack the resources for experimentation and the agility of startups with minimal overhead. This is evident with AI as well. According to a survey by the industry association Bitkom, a majority of companies with fewer than 500 employees have so far avoided investing in AI. The reasons include a lack of personnel, time, and often competitive pressure. Many businesses remain hesitant.So, should they wait? The answer is clear: No. Large Language Models like ChatGPT, Claude, or Mistral provide easy access that anyone can learn in a matter of seconds ("prompting"). They are low-cost and can quickly deliver noticeable effects in certain tasks, especially in office environments. To stay competitive, every company should explore the potential of these models. This doesn't require computer experts; anyone can experiment with them.
However, improving or evolving the business model with AI demands more effort. For small companies, AI becomes an investment at this point, as it requires developers who can technically handle the systems. The first question should be: How could AI increase my revenue? How could AI reduce costs and improve services? How can my customers benefit? Understanding the technology and its possibilities is essential. Free information resources, such as those offered by the Mittelstand 4.0 centers in Germany, can help accumulate knowledge. Once a use case or idea for an application is identified, local partners and funding can assist in implementation.
Although large cloud companies like IBM, Google, or Amazon also offer AI solutions, these can be quickly oversized and still require experts for successful implementation. And AI experts are currently scarce. Those who do not yet see a use case for AI should stay engaged: competitors will adopt it, and by then it might be too late. With the rapid advancement of AI, this moment will come sooner rather than later.
At the same time, the costs and resources needed to implement AI are decreasing rapidly. In recent years, frameworks such as TensorFlow and PyTorch have emerged, providing the basic tools to quickly set up AI networks. Even small companies can now implement AI, with the five-person Bremen-based company INnUP being a perfect example. Moreover, systems are being developed to enable AI use by non-programmers.
Another piece of advice: data is the oil of AI. Those who start collecting, storing, and cataloging data today will benefit tomorrow.
Artificial Intelligence and Humans:
Like many new technologies, AI stirs up fears. A well-known study by Oxford University in 2013 analyzed that 47% of all US jobs were at risk due to automation, a significant portion of which was attributed to AI. Such statistics fuel fears that can lead to real-world actions. For instance, Waymo, Google's subsidiary for automated driving, reported that their test vehicles were attacked multiple times with knives and stones. So, is AI a threat to humans? A Bitkom survey paints a divided picture: 62% of Germans see AI primarily as an opportunity, while 35% view it as a danger. Another survey among managers found that 42% observed employee reservations.The truth lies somewhere in between. AI will undoubtedly take over some human tasks completely meaning that for these specific tasks, humans will no longer be needed. These are often tasks that are monotonous and repetitive: watching surveillance videos, answering standard inquiries over the phone, and searching through documents.
In many areas, however, AI will likely play a more supportive role for the foreseeable future. It will assist doctors in reaching accurate conclusions or be part of a workflow such as in agriculture, where robots can autonomously or semi-autonomously handle certain tasks while humans continue to perform others.
Simultaneously, new jobs will emerge, driven by innovative AI business models. People will have more time to devote to new tasks because they will be collaborating with AI. For example, lawyers could spend more time with clients instead of spending hours searching through files. It is clear, therefore, that more education is needed to prepare people for their new roles and to equip them with the skills to work with AI systems.
A study by the World Economic Forum in 2023 predicts that while 89 million jobs will be displaced by computers, 63 million new jobs will be created.
To be honest, we don’t have a choice. AI has already been integrated into our daily lives, and almost everyone uses it in some form, whether knowingly or unknowingly through smartphones, bank transfers, or navigation systems. It will take some time before AI becomes ubiquitous, but that moment is approaching sooner rather than later. Once a sector benefits from AI, it will gain significant advantages over its human counterparts, either pushing them out of the market or opening up new areas of work for them.
The Natural Imperfections in Artificial Intelligence:
AIs are created by humans and thus inherit a natural issue: an intelligence that mimics humans is also subject to their cognitive limitations. One such limitation is bias.An example: In 2014, Amazon's AI experts developed a system to automatically evaluate and sort job applications. They trained the neural network using resumes from the past decade. Once trained, they discovered that the algorithm started favoring resumes from men when presented with new applications. The reason was that historically, the tech industry had a higher proportion of male hires. The AI inferred a rule: only select male applicants. (Source) The flaw lies in the selection and preparation of data. Amazon ultimately scrapped the experiment, opting to continue the manual review of applications.
This example highlights the importance for humans to meticulously select representative data when designing artificial intelligence and to be aware that biases may already exist in the data chosen and prepared. This dilemma is not easily solved and must be carefully considered in every AI design. This underscores the value of external perspectives and collaboration with AI experts and partners.
Ultimately, every AI is programmed by humans and where our intelligence begins and ends, we are keenly aware.
To conclude with a summary of what artificial intelligence entails:
- AI aims to replicate human learning and cognition on computers.
- Strong AI, capable of general problem-solving like in science fiction, remains theoretical. Weak AI, however, is increasingly prevalent in today's world, found in smartphones, websites, social media platforms, and self-driving cars.
- AI is valuable wherever extensive data analysis and pattern recognition are required.
- Machine learning is currently the most commercially significant subset of AI.
- AIs process data faster and more accurately than humans but do not comprehend it.
- AIs are trained for specific tasks and need retraining for new purposes.
- AIs will assume human tasks while simultaneously creating new business sectors and jobs.
- Garbage in, garbage out: AIs produce flawed results if fed with flawed data.