Science Fiction meets reality: Regardless of whether it was the computer HAL 9000 from A Space Odyssey or the T-800 in Terminator, AI has always been impressive (and linked to dubious intentions), but it has also been a distant reality. By the 1990s, however, technological development in this field began to accelerate. AI pioneers such as Yann LeCun, Geoffrey Hinton, and Yoshua Bengio laid down a number of scientific foundations, and starting from the 2010s these were gradually implemented in larger and larger AI training models, as the necessary computing power had become available by then. This led to some remarkable breakthroughs such as the AI-trained image recognition system AlexNet, which won the image recognition competition ImageNet in 2012, and the resounding victory of AlphaGo against the best human Go player, Lee Sedol, in 2016.

Nevertheless, the broader public was generally unaware of this rapid development, mainly because AI was only really able to shine in certain niche areas. The release of ChatGPT 3.5 in November 2022 changed all of this in an instant. ChatGPT and comparable large language models (LLM) are not only intuitive and useable without (almost) any previous knowledge, they deliver remarkable results. So remarkable indeed that they raise important questions, which we address in the following.

What exactly is artificial intelligence?

The world is a complex place. Just a few years back is was inconceivable that AI software would be able to interpret images correctly, understand language, and properly assess the context of words and texts. But trained AI software is able to do much more in the field of language than just generate poems and jokes and imitate styles.

AI training is based on mathematics. The key to understanding AI is neural networks (NN) and thus ultimately mathematical functions. Let’s go back to our schooldays for a second: Functions can be used to describe (almost) anything, as long as the relationship between two elements (numbers) is known. For example, the sound waves which reach our ears and are interpreted as voices and words by our eardrums, or light rays that reach our eyes and are transformed into images by our optic nerve. Or how one recognises the (well hidden) head of a lioness in a digital image. This kind of function is complex, but it is manageable for a neural network, insofar as enough data are available that can be used for testing.

Neural networks are universal function locators, or to be more precise, function approximators. This is valuable, because for example in contrast to describing a triangle, for which known sinus functions can be used, there are no pre-set functions available for most of the connections in the world (image recognition, language recognition, etc.). These must be learned by the NN in a process that is intensive (in terms of computing power as well), in order to produce the necessary connections.

Even experts were surprised how good NNs are at producing connections and contextual links by training. Because not all connections which AI models produce and which lead to correct results are logical for us. In other words, there are reasons to believe that AI will be able to help us solve (data-intensive) problems in the fields of science of research, because it recognises connections that have previously remained hidden to us. Speaking about this, Jeff Bezos recently said that AI is not only to be seen as an invention, but rather as a discovery as well, to a certain degree. A discovery, whose potential has not been fully recognised and researched.

Example with ChatGPT

AI, like ChatGPT 4.0 in this case, is not programmed to immediately recognise what a hidden lioness looks like. That would hardly be effective or practically possible, because “the whole world” is too complex to be packed into a piece of software, regardless of how big the programme might be. Instead, AI creates the connections (functions) itself using a large number of runs (trial, error & feedback) in a neural network.

Why are people also afraid of AI?

We use computers and mobile telephones as tools. It has become completely normal that technological progress results in the continuous development and improvement of the hardware and software for these tools. This trend does not scare us. But this is different to some degree when it comes to AI. The prospect that AI, which has already achieved significant results, may continue to develop at a similar pace in the coming years and decades, as we are used to seeing with semi-conductors and micro-processors, is unsettling to some people.

This feeling has deep roots: Books and films which address these fears, such as Brave New World and Metropolis, were already being made well before the first computers were even invented. They portend topics which are on people’s minds today as well when extreme, negative AI scenarios are being discussed: Are we on the verge of a new age, in which humans will be dominated by machines? Why should I, as a human, even bother to live up to my potentials and to expand my horizons at the expense of considerable time and effort, if AI is faster, more intelligent, and more creative than I could ever be.

That said, one need not even go to this philosophical level in order to understand that a revolutionary technology like AI must necessarily also have a darker side. On the one hand, because undesired consequences could occur, which are unforeseeable at the present time. On the other hand, because AI opens up new opportunities that can be exploited for criminal or otherwise harmful purposes.

One could say that as a society we were similarly unprepared for the advent of social media 10 to 15 years ago. In conjunction with the spread of smartphones, social media has led to significant changes in our world. Many of these are good, but obviously there also have been and continue to be negative consequences which either stem directly from this latest major innovation or have been amplified by it. The potential impact and change from AI is at least as significant as that of social media, if not greater. And thus the challenges are massive as well.

What can AI do and where are we heading?

In response to critics, many people point out that one of humankind’s key strengths is the ability to adapt. Transition periods triggered by technological leaps are often unsettled times, as they result in major changes. Over the long term, however, in the past we have learned to manage comparable disruptions, to adjust to the new opportunities and risks, and to set limits on the negative aspects. .

Positive aspects of AI are already reality or approaching maturity. We can see examples for this in a wide range of areas:

  • AI-supported robots are being used more and more for recycling waste materials. The task of recognising and differentiating various kinds of waste on a conveyor belt is a perfect job for image-recognising AI.

  • A market for AI-supported robots is now developing in agriculture. Image-recognition systems can distinguish between weeds and crops, so that only the weeds are sprayed by the robot. There are a number of start-ups in the so-called precision agriculture field, which are concentrating on planting, weeding, watering, and applying fertilisers in a targeted manner. Manufacturers of these systems say that CO2 emissions can be reduced by precision spraying, because up to 95% less herbicide is used compared to traditional, undifferentiated spraying.

  • Schools are another possible area for AI. Accommodating the learning needs of an average of 25 children with different strengths and weaknesses is a great challenge for teachers, and can only be met by making compromises. This often means that some students are overwhelmed, while others are bored and cannot progress as quickly as they could if they had more individualised learning conditions. AI-supported teaching programmes can quickly adjust to the performance level of a student. An increasing shift towards having AI systems teach content would result in a significant improvement in educational outcomes and also free up teachers, so that they could concentrate more on imparting and solidifying students’ social skills. Even grading can be supported by AI, which might help to improve fairness. Because in contrast to human teachers, AI only assesses the performance, free of any prejudice or stereotyping.

In summary, artificial intelligence is one of the most significant development topics for the immediate to long-term future. A wide range of economic sectors are impacted by this topic, especially in the fields of education, marketing and sales, commercial activities, and industrial production.

ESG assessment of the topic ai

„E“ (environment): From an environmental point of view, AI opens up new possibilities in the analysis of networked databanks, as a result of which measures for environmental protection can be developed. On the other hand, the very large amounts of energy required, in particular for the (advanced) development of AI models, is an extremely negative factor.

„S“ (social): From a social perspective, AI has positive potentials in the fields of medicine and pharmaceuticals, while the possible loss of jobs due to AI is a negative factor.

„G“ (governance): The governance aspects are extremely varied. Artificial intelligence can lead to better security due to improved solutions in relation to criminal acts or cyber threats, but the risk of data misuse and various forms of manipulation is very high.

Leopold Quell, Raiffeisen KAG

Author

Leopold Quell, Fundsmanager Raiffeisen Kapitalanlage GmbH

Further sources:

Watching Neural Networks Learn
A Hacker’s Guide to Language Models
chatbot-arena-leaderboard
Intro to Large Language Models
But what is a neural network?
How the Enlightenment Ends Henry A. Kissinger
Noam Chomsky on Artificial Intelligence, ChatGPT

This content is only intended for institutional customers.

More