Welcome to the info area (as of 20 August 2024)
on this page you will find:
What has happened so far – development of AI
Will there ever be a superintelligence? If so, when?
What already exists today, what will be possible tomorrow
How AI can become dangerous for humans – and already is
What has happened so far – the development of AI to date
The term »artificial intelligence« is subject to constant change. The capabilities of machines that were referred to as »artificial intelligence« in the 20th century do not have much to do with the capabilities that we call AI today, and these in turn probably do not have much to do with those that we will call AI in ten years’ time.
However, the goal is very clear: to develop machines that have all the cognitive abilities of humans – that can think, learn, solve problems and plan like humans.
Great progress has been made towards this goal in recent decades. The first successful attempts were made in the 1950s (e.g. by Herbert Simon and Allen Newell with »Logic Theorist« and »General Problem Solver«). In the 1980s, backpropagation made it possible for the first time to create neural networks that could learn effectively (Seppo Linnainmaa, David Rumelhart, Paul Werbos, Geoffrey Hinton). Subsequently, convolutional neural networks (CNNs) were developed that were trained using backpropagation (Alex Waibel and Yann LeCun in the late 1980s, building on the scientific achievements of Kunihiko Fukushima from the 1970s). CNNs were particularly well suited for image recognition.
The turning point
A turning point in the history of AI was the development and publication of AlexNet (2012, Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton at the University of Toronto). AlexNet demonstrated the power of Deep CNNs, initiated a wave of interest and research and also contributed significantly (probably unintentionally at first) to the commercialisation of AI research.
The last big breakthrough
In 2017, scientists working for Google published the paper »Attention is all you need« (the basis for GPT), in which the Transformer architecture (the »T« in GPT) was presented, which significantly improved the performance of neural networks. This technology made it possible to train Large Language Models with large amounts of text data in advance (pre-trained, the »P« in GPT) and to optimise them for specific tasks such as generating (the »G« in GPT) text or images through fine-tuning. Already in the 1990s first approaches existed already that have been similar to the Transformer technology. However, there were not enough powerful computers available at the time to successfully develop them further. Generally speaking, many of the advances of the last two decades were only possible thanks to rapidly improving hardware.
Since the Transformer paper, Generative Artificial Intelligence (GAI) is perhaps the most important technology of our time. Large tech companies (e.g. Alphabet, Meta, Microsoft, Apple, Amazon, Nvidia) and new companies (e.g. OpenAI, Anthropic, Mistral, X.ai) are investing hundreds of billions of dollars in the development of LLMs and the necessary hardware and software. New versions of LLMs and new AI tools are launched on the market every month.
What was perceived by the general public as a sudden event (the release of GPT-3.5 in November 2022) was in fact the result of decades of research and development work, some of which was painstaking and unrecognised, and most of which was carried out at universities around the world (mainly in Europe, North America and Japan).
Dominance of the USA
For decades, the development of AI has taken place at universities and in scientific circles. Even when breakthroughs were achieved in corporations such as AT&T (Bell Labs), they were shared with everyone. When it became clear in 2012 with the presentation of Alexnet how powerful machine neural networks can be, this changed rapidly. Today, most research and development in this field is carried out in US corporations or in well-funded and rapidly growing US start-ups – and the results are often no longer shared, but kept as secrets.
We find it somewhat unbearable that the human dream of artificial intelligence has been developed slowly but steadily over decades (almost a century if you count the development of computers) by researchers and mathematicians in Europe, Japan and North America in open research – and that American capital is usurping it all in the last 100 metres. But that’s the way it is and we in Europe have to live with the many negative consequences (see here for an example) that arise from this. Or not – see our demands for politicians and authorities .
Definition of terms
The term was first introduced in 1955 and now encompasses many disciplines. »Artificial intelligence« is an umbrella term for various forms of machine learning and thinking skills, but also refers to the current state of development in this field.
The term »AI« therefore often refers to what is currently state of the art in this field.
Many definitions sound something like this:
»Artificial intelligence is the ability of a machine to acquire human abilities such as logical thinking, learning, planning and creativity«
It is debatable whether current systems already fulfil these requirements and, if so, how intelligent they are compared to a human being.
In May 2023, some AI researchers and developers assumed that GPT-4 is as intelligent as a three-year-old child. That is remarkable. Meanwhile, the intelligence of an 8-year-old child is attributed to it.
Some tests have already given GPT-4 an IQ of 150 in specific, narrowly defined subject areas, which would make it a genius as a human being.
Machine Learning (ML)
An AI system employs ML, among other things, so ML is one of the components of AI. In terms of development history, ML could also be seen as one of the precursors to AI. An ML system can recognise patterns based on training data and then recognise/create/understand things that go beyond the data it has been fed.
Deep Learning (DL)
This is a special form (or a further development) of machine learning that uses artificial neural networks in several layers. Machine learning uses algorithms to analyse data, while deep learning uses algorithms in layers to create a neural network.
Artificial General Intelligence (AGI)
AGI refers to an AI system that can understand/solve all tasks that a human can. Many AI researchers do not like the term and the abbreviation because machine and human intelligence have different strengths and weaknesses and it is therefore very difficult to compare »general« intelligence. They prefer the abbreviation HLMI, which stands for Human Level Machine Intelligence or High Level Machine Intelligence. However, it always means the same thing: Machines that have the cognitive abilities of adult humans.
The opinions of AI researchers differ widely on the current stage of development of AGI and, above all, the time horizon for future development steps. Some (fewer and fewer) say that AGI can never be achieved. Others believe that it will take decades. Still others speak of a few years, e.g. Ilya Sutskever said in an interview that he expects AGI by 2030. Yann LeCun never tires of emphasising that it will take much longer – and that we currently don’t even know how we can make the machines as intelligent as a domestic cat.
On 9 July 2024, OpenAI announced the expected steps towards AGI in an internal memo:
1) »Conversational AI« – already achieved
2) »Reasoners« (reasoning, …) – soon to be achieved
3) »Agents« (autonomously acting systems)
4) »Innovators« (independently develop new things)
5) »Organisations« (act as complete organisations)
Whereby level 5 AGI would be roughly reached.
The time horizon? OpenAI expects this to be realised in less than 10 years.
In defining these development steps, OpenAI largely agrees with Yann LeCun, who says that we need things other than LLMs to achieve AGI because LLMs lack the following:
1) real understanding of the physical world
2) the ability to reason, i.e. reasoning and complex thinking
3) the ability to plan
4) a memory
The current LLMs will, also in theory, never be able to do this. So either the LLMs will be developed further or – and this is more likely – we will need other technologies. In any case, some scientific breakthroughs are still needed on the way to AGI or HLMI. It is in the nature of scientific breakthroughs that they are unpredictable, so forecasts – no matter who makes them – should be treated with caution.
Super Intelligence (SI)
This term refers to artificial intelligence that far surpasses that of humans. This could be achieved by an intelligence explosion, for example. An intelligence explosion is a sudden, uncontrolled increase in intelligence by a self-developing AGI. Or to put it differently: AGI systems improve and develop themselves – much faster than humans could.
Generative Artificial Intelligence (GAI)
This term refers to AI systems that can generate text, images, music, videos, code or other data. ChatGPT, Midjourney, Udio, Sora etc. are examples of GAI applications. GAI in its current form was made possible by LLMs and is the technology of the hour. When people talk about AI, they often mean GAI systems.
Chatbot
Chatbot is not a well-chosen term. Firstly, many people associate it with conventional, less intelligent chatbots, in which customers had to laboriously manoeuvre their way through decision trees with predefined options. And secondly, AI systems such as GPT can do much more than just chat. Detailed explanations can be found further down on this page or on our test bot page.
Hallucinate
This term has become established as a technical term for the fact that chatbots also tend to claim falsehoods. And they do so with complete conviction. Sometimes they cite sources or study results that don’t even exist. This is because they have learnt from the data they have been trained on that sources are often cited. So that’s what they do. And because their underlying statement is not true, there can be no sources for it. The chatbots don’t care, they simply make everything up – and do so very convincingly.
It’s not easy to get chatbots to stop hallucinating. The phenomenon is common to all known LLMs from all providers and developers. They are working on it and improvements are noticeable. GPT-4, for example, hallucinates significantly less than the previous version GPT-3.5, but it is uncertain whether the hallucination of LLMs is not inherent in the system. This would mean that we would only be able to completely eliminate such hallucinations with a new technology.
Even though this term has become established, it is not a good choice. Because what the chatbots do has nothing to do with human hallucination. Rather, they are »bullshitting«, i.e. blathering nonsense in a tone of conviction.
Generative Pretrained Transformer (GPT)
und Large Language Model (LLM)
GPT is an advanced form of a so-called Large Language Model (LLM). The US company OpenAI (more precisely: Ilya Sutskever, co-founder of OpenAI and its Chief Scientist until May 2024) recognised the potential of this technology, which was presented in 2017 in the now famous scientific paper»Attention is all you need«, and developed its own versions. In 2020, it presented its third version – »GPT-3« – for the first time. In November 2022, ChatGPT, the chatbot based on it, was presented and triggered a veritable hype. GPT-4 was released in March 2023. This model was the most powerful publicly available model until 20 June 2024 (see below). Due to the extremely large amounts of data it was fed with (hence the first L in »LLM«: Large), the results are impressive. During and after training, the following has emerged: By learning statistical correlations in a huge number of texts, the neural network somehow also learns about the process that created these texts. All these texts are a kind of projection of the world. The neural network therefore learns something about the world, about people, their state of mind, their hopes and motives – especially if the amount of text is huge and the data compression is very good. And so it can certainly »understand« texts (depending on the definition of understanding), as we were also able to prove in our own tests . This text comprehension could also be seen as one of the Emergent Abilities . And because GPT is a »transformer«, i.e. a special form of neural network, it can (through a fascinating, multi-layered process that includes a self-attention mechanism) understand the relationships and connections between different parts of the input data and, above all, can be trained much faster.
ChatGPT can be seen as the result of additional learning, namely reinforcement learning (RL), partly also RL accompanied by humans, which is therefore called reinforcement learning from human feedback (RLHF). The combination of GPT and HRLF has resulted in an extremely capable chatbot. The current version is ChatGPT-4o (o for »omni«).
LLaMA from Meta (the Facebook parent company) was published in February 2023. Unlike GPT, it is open source, i.e. freely accessible. There is a certain irony in this, as OpenAI was originally founded as an open source non-profit company (incidentally, Elon Musk played a key role in this, but left the organisation in a dispute), but is now a for-profit company with significant Microsoft involvement, which does not make its products open source – quite the opposite. And Meta, of all companies, which has always focussed on maximum profit, is now making its LLM available as open source. The current version is LLaMA 3.1, which is available in three sizes: 8B, 70B and 405B. The largest version is similar in performance to ChatGPT-4o, but uses significantly more resources.
Google developed the LLM PaLM, on which Google’s chatbot Bard is based, which was released in March 2023 and has since been renamed Gemini. The current version is Gemini 1.5, which is characterised by a significantly larger context window size of one million tokens. In comparison: ChatGPT-4 has 128,000 tokens in its largest version, Claude 3 Opus and Claude 3.5 Sonnet have a context window of 200,000 tokens (as of August 2024).
Anthropic (a company founded by OpenAI employees because they disagreed with the decisions at OpenAI and wanted to develop a better, safer chatbot) had the second-best LLM on offer with Claude 3 Opus until 20 June 2024. Like the other providers, Anthropic also offers different versions to choose from. Claude 3 Sonnet and Haiku are less powerful – but also more resource-efficient. On 20 June 2024, Anthropic released Claude 3.5 Sonnet without prior notice, putting it ahead in the race for the smartest, most powerful LMM. This is not only the result of Anthropic’s tests, but also of independent tests. We at Superintelligenz.eu have also pitted Claude 3.5 Sonnet against ChatGPT-4o in our own tests and can confirm this: Claude 3.5 Sonnet is also slightly better than GPT-4o in German.
Like ChatGPT-4o, Claude 3.5 Sonnet is also available free of charge for limited use. However, the following applies to both products: If you want to use them properly, you must register for a paid subscription.
One provider that makes some of its LLMs available as open source is the French company Mistral, which was founded in 2023. In its most powerful version, ‘Mistral Large 2’, it comes close to the performance level of Llama 3.1.
X.ai released Grok-2 in August 2024, which is initially only available to premium users via X (the former Twitter), and in the smaller version. Compared to the other LLMs, Grok-2 is less censored and restricted.
In addition, there are many other LLMs in various quality levels, many of which are open source. LLMs are currently being developed at a rapid pace.
Of course, LLMs cannot decide for themselves whether something is right or wrong. As many of them have been trained with data from the freely available internet, there is a risk that they have been fed false information or semi-truths. And then they repeat them. Preventing this is very difficult and in some cases still an unsolved problem.
Natural Language Processing (NLP)
In connection with artificial intelligence, NLP means the processing of natural language using algorithms. An important part of this is tokenisation. A token comprises a small number of letters, usually 2, 3, 4 or 5 – on average there are more in English than in German, generally it is different in every language, and it also depends on the quality of the tokeniser of the respective LLM. The use of AI systems is often charged according to the number of tokens used, which means that the same number of characters costs less in English than in German and most other languages.
Alignment
Alignment is an aspect of AI safety research. It is about ensuring that the goals, values and behaviour of an AI system match human values and intentions. In other words: that the AI system does exactly what it was created to do. And if it develops beyond that, it does so in a safe, human-controlled environment.
Alignment is essential in the development of AI systems, because if we don’t get alignment right, the probability of dangerous scenarios occurring increases. The extent of alignment should therefore grow faster than the capabilities of AI systems.
Emergent Abilities
These are abilities that a model develops simply by being enlarged. The model itself remains the same, but the enlargement causes it to develop new abilities. Some of these abilities come as a surprise even to the developers and are therefore unpredictable. On the one hand, this is a really cool feature of a neural network, but on the other hand it is of course also a cause for concern – because it would be good if people knew what the machines would be able to do after the next stages of development.
»We don’t understand how these models work. I think this is the first technology that nobody understands how it works.« This is what Alison Stanger, who heads a research network on new technologies at Harvard University, said at the Digital Humanism Summit in Vienna in early July 2023. (Source: https://science.orf.at/stories/3220160/) Like like many others, she says she can’t sleep at night because of it.
Will there ever be a superintelligence?
Although superintelligence still seems a long way off at the moment, we believe it could arrive in this decade.
Ilya Sutskever und Jan Leike
Superintelligence will one day be a hundred times more intelligent than the most intelligent human being. Some say that this will only be the beginning and that the superintelligence will be a thousand times more intelligent than a human shortly afterwards. But that doesn’t matter, as the following thought experiment shows:
The difference between a person with an IQ of 60 (considered slightly mentally handicapped) and a person with an IQ of 180 (considered an outstanding genius) is enormous – a person with an IQ of 60 cannot even imagine the world of thought of a person with an IQ of 180. And that is only a factor of 3!
It is therefore simply impossible for our brains to comprehend how a machine that is a thousand times more intelligent than us ‘thinks’ and acts. Even a factor of 10 would be inconceivable. But a factor of a hundred? A thousand? We have to accept that this is far beyond our imagination and that the future is therefore not only unpredictable, but also simply unimaginable.
It will still take some time to get there. It is very difficult to estimate how many years – or rather decades. It may take longer than many people think, because we are in the midst of a breakthrough and hype. Many people are very optimistic about this. The next, necessary breakthroughs could be a long time coming.
But it could also take less time than most people think – for example, if AGI systems, i.e. the preliminary stages of superintelligence, are developed soon and then develop themselves further – and things accelerate dramatically as a result.
But it could also be that such a superintelligence will never exist.
Our assessment: It will happen, most of us will live to see it. At a certain point, it will happen faster than expected (intelligence explosion). When will this point in time be? Our tip: in 10 to 20 years. But even before that, we will have incredibly impressive systems.
Just a few thoughts …
Once humanity has succeeded in creating a being superior to itself, things will change fundamentally.
We should then probably no longer call ourselves ‘Homo sapiens’, because there will then be another knowing being. We could then call ourselves ‘Homo sentients’, i.e. sentient beings. Because it is unlikely that a superintelligence can and will feel like us humans. A major difference lies, for example, in the knowledge of one’s own mortality – this completely changes the world of emotions.
Philosophy will also be rewritten.
And the major world religions will also have to adapt. After all, they portray humans as superior to all other beings. It will then no longer be possible to make this claim. It is interesting that religions claim that although their teachings were written down by humans, they were not made by humans. A superior authority has virtually dictated them to us. In the future, there could be just such a superior authority – the superintelligence, made by humans. And it could explain to us many things that humans have been striving to understand for thousands of years, like a new deity. Well, boom! Either a new religion emerges or the existing ones adapt. Probably both.
Do we even want all this? Should we strive for it? The question is probably not. Technological progress has never stopped. All over the world, intensive research is being conducted into AGI and AI systems are being further developed. And at some point, the consequences will fundamentally change everything. Really everything.
In this interview, Geoffrey Hinton explains from approx. minute 1.50, why AI systems can learn faster than humans, why they already know more than any single person knows and why they will soon be much more intelligent than humans. In the future, their neural networks may be able to learn more efficiently than the human brain – and they are much better networked with each other than human brains are.
AI: What already exists today, what will be possible tomorrow
AI today
AI systems have long been part of our everyday lives. Without considering the differences between »real« AI, ML, DL, etc. for this list, they are used in the following areas, for example:
- Medicine: in diagnostics, radiology, patient monitoring, development of new drugs and many other disciplines
- Traffic and transportation: traffic guidance systems, self-driving cars, autopilot in airplanes, drones, navigation systems
- Sales: learning systems on the web, customer analysis, suggestion generation, data collection and evaluation on social media platforms, etc.
- Education: plagiarism detection software for universities
- Energy supply: Grid utilisation forecasting, peak reduction, load balancing
- Finance: Fraud detection, risk management, investment forecasting
- Media: image processing and optimisation
- Industry: robotics in production lines, quality control
- Entertainment industry: e.g. suggestion system for streaming service providers
- Surveillance: facial recognition systems, step recognition systems
- and many more
However, the following applications are relatively new:
- Media: creation of photorealistic graphics that are indistinguishable from real photos; text creation through to writing entire articles (or even entire books); automatic creation of websites, flyers, folders, brochures, etc., photo enhancement and optimisation
- AI chatbots: in-house for onboarding; customer support; information systems, knowledge transfer (see also our test bots), consulting and coaching, marketing, personal digital assistants
- Support for everyday office life: meeting transcription, email creation, answering inquiries
- Programming: code review and also code generation
- Music: composition, complete production of songs
- Video creation: partly with human-looking avatars
- Education: vocabulary and other learning software, personal digital tutors
- Text-to-speech generation
- Translation programs based on LLMs
AI systems can already write scientific papers, books, songs and operas, photos, paintings and so on. Very soon, they will be able to do this with a quality that is in no way inferior to that of works created by humans. And just a little later, they will be able to do it better than all humans in some areas.
Enlightenment 2.0
Artificial intelligence will lead to a new age of enlightenment. It will enable so many new insights in science that humanity’s knowledge will literally explode. In just a few decades, people will know things that are completely unknown and unimaginable today.
AI tomorrow
Ready for a bit of utopia?
AI has the potential to achieve groundbreaking things for mankind:
In medicine:
- Diagnostic tools that are still unimaginable today: For example, one day there will be an electronic nose that can smell a thousand times better than a dog. In combination with blood analyses, this ‘nose’ will make it possible to diagnose what ailments a person has – and above all: what diseases they might develop in the near future. This will therefore be a very powerful prevention tool.
- Drug development: AI will one day be able to develop drugs (and other treatment methods) for all diseases and ailments.
- Machines as surgeons: Today, most people who need surgery in a hospital want to be operated on by a human. But that could change very quickly. Once AI is so advanced that the error rate for machine surgeons is much lower (they are never tired, don’t have crises, aren’t distracted, don’t want to impress the smart nurse, etc.) and this is also generally known, then patients will very quickly say: ‘I should be operated on by a human? Certainly not!’ Robots will simply be able to operate much more precisely and better than humans.
- Ageing: With the help of AI, it will one day be possible to slow down and even stop cell ageing in humans. Eternal life as a utopia? Socially, this also has its downsides. But who will want to age if it can be prevented? When do researchers believe that the first breakthroughs can be achieved in this field? Between 2040 and 2050! That’s in about 20 years! How old will you be then?
In education:
- Improving performance through personalised digital tutors. There are already chatbot-based tutors that offer 1:1 tuition as personal learning assistants, although these are not yet available in Europe (e.g. Khanmigo from the Khan Academy, or GPT-4o in the version with voice mode presented on 13 May 2024, which we expect in Europe in autumn 2024). These will be so good in the near future that it will be possible to improve the performance of all students (as an example). Poor students will become average, average will become good, good will become very good and so on. The bell curve will shift to the right, and society as a whole will know and be able to do much more than is currently the case.
- Together with AI, humans will develop a completely new form of learning and thus be able to learn better and more efficiently.
In environmental protection:
- Zero emission: With the help of AI, technologies will be developed that generate energy without any harmful emissions and without the consumption of resources – and in abundance.
- Saving the planet through technologies developed by AI, repairing damage already done, etc.
In politics and administration:
- An end to corruption through AI real-time systems
- Citizen service becomes fair, equitable, fast and efficient – and all at extremely low cost.
- World peace (but people have to want it first)
In space travel:
- Discovering new worlds
- Transportation to these places
In business and society:
- The extreme increase in productivity through the use of AI and machines makes it possible to distribute more than sufficient wealth to everyone; many people will no longer have to work for a living, but can use their time for better things.
Too much of a blessing? Too utopian? Continue reading How AI can become dangerous for humans.
Impact on society
Labour Market
Opinions differ as to what impact the further development and use of AI systems will have on the world of work – and they differ considerably.
Studies have been published that show that 80% of jobs will be influenced by AI in the near future. What exactly does ‘being influenced’ mean? Well, this paper published by OpenAI in March 2023 is worth reading: GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models (PDF-Datei)
It states, for example, that better-paid jobs will tend to be more affected. And: 19 % of all jobs consist of 50 % or more activities that could be taken over by AI in the future. In this study, the impact of AI on jobs is listed according to occupation type and also according to the highest level of education completed.
This study also shows that better-paid jobs can be replaced by AI in particular. According to the study, the sectors most affected will be legal services (lawyers, notaries), the investment sector and the security industry. Professions that will change significantly in the very near future can also be found in university teaching (foreign languages, history) and telemarketing.
Many assume that the use of AI will eliminate a large number of jobs for people and that significantly fewer new jobs will be created.
Others point to the Kondratiev cycles and technological leaps from the past, such as the invention of the steam engine, the industrial revolution or the internet. All of these developments were feared to result in major job losses – but this did not materialise. Although many professions or entire occupational fields did indeed disappear, new ones were added. And these were usually even better for people, in the sense of more pleasant. And it could be the same with AI.
We believe that the labour market in Europe will undergo major changes from 2025 onwards due to AI developments. Firstly, a distinction must be made between manual labour and knowledge work. In the first stage, it will mainly be professions in knowledge work that will be affected, and affected badly:
Interpreters; insurance employees; employees in the financial sector (banks) and in the legal sector (law firms, notary’s offices, corporate contract management); people who work in administration (in large companies, but also in government offices, ministries, etc.); secretaries and assistants; clerks; editors; copywriters; graphic designers.
In some of these occupational fields, there will be 10 to 25 % fewer jobs in a country like Germany or Austria over the next few years. This is the difference between the approx. 20 to 30 % of jobs that will disappear and the approx. 5 to 10 % of jobs that will be added.
We therefore do not believe that the jobs lost in these areas will be replaced by an equal or even greater number of new jobs. A look at the past gives us hope, but a steam engine cannot build any more steam engines. An AI system will indeed multiply – and continue to develop itself in the process. This is why AI is different from all previous technological leaps. For the first time in history, a new technology will destroy more jobs than it creates, at least in the medium and long term.
Hybrid output
Productivity will increase continuously. The output will be hybrid, i.e. people will work with AI, part of the output will be human-generated, part AI-generated. The ratio will shift more and more towards AI-generated and will vary depending on the industry and specific job. This means that the shift towards AI will not happen suddenly, it will be a gradual but accelerating process.
For example, in some areas, the work that is currently done by three people without AI will be done by two people with AI. Those people who learn to work with AI will (for the time being) continue to do their job or another job, while those who cannot (or do not want to) work with AI may lose their job. And there will be many of them – and more and more of them.
People who are proficient in AI tools and who let these tools develop the best solutions will be needed. These tools will continue to evolve at a rapid pace. Mastering AI tools will be as natural for knowledge workers as using a computer today, including the use of various programmes.
The individual differences also apply to companies. Even in severely affected sectors, there will be companies that will have more employees in the future than before the use of AI. This is because their total output will grow faster than the AI-generated output thanks to the correct handling of AI (timely changeover, etc.).
Next up are professions that require a great deal of specific knowledge and cognitive skills: Lawyers, judges, notaries, architects, engineers. And in some areas (e.g. diagnosis) also doctors.
In the USA, some lawyers experimenting with AI and AI experts are convinced that the best lawyers will be machines as early as 2026.American case law is particularly suited to AI systems, so it is not comparable with Europe.But it won’t take much longer for us either.
However, it will be around two decades or more before robots are covering our roofs, making house calls as plumbers, cutting our hair, repairing our cars (which cars?), etc. Manual labour will therefore initially be much less affected by AI.
This also corresponds to the estimate on page 6 of this study by Goldman Sachs: »We estimate that 25% of current work tasks in the US can be automated with AI, with administration (46%) and the legal sector (44%) being particularly affected, and manual activities, e.g. in the construction industry, being particularly less affected.«
This will change a great deal in society. What was previously a secure career path with a fairly well-paid job with some prestige (school-leaving certificate, university degree, office job) will disappear more quickly as a result of the development of AI than the previously less prestigious and largely lower-paid jobs in care, in hairdressing salons, on building sites, etc. This will lead to a rethink.
And what do we as a society do with all the people working in knowledge work whose professional activities will be carried out by machines/AI in 5 or 15 years’ time and who will therefore no longer have a job? This question will occupy us a great deal in the future – together with the question of what should happen to the profits generated by the greatly increased productivity of companies/corporations. In some form or other, there will have to be a distribution of the capital generated by machines to people who are not working.
Education
We are generally facing major changes in the field of education. ChatGPT-3.5 from OpenAI was released in November 2022 and since the GPT-4 version (March 2023) at the latest, it can perform many of the text-based tasks of pupils and students in high quality. AI systems based on GPT not only pass the Abitur or Matura exams, but also US bar exams, medical tests, etc. While the performance of the GPT-3.5 model corresponds to that of average pupils or students, systems based on GPT-4 regularly rank among the top 10%, and in some areas even among the top 2%. The next generation of GPT could already perform better than all humans in various tests and examinations or achieve the maximum number of points.
It therefore no longer makes sense for teachers to set text-based homework in schools. This would simply be done by GPT. So what should be done? A general ban on the use of GPT-4 and similar AI systems is not only pointless, but also counterproductive. Firstly, it cannot be controlled – because although there are AI detectors, they can be tricked or do not work from the outset. And secondly, a ban is the opposite of what should be achieved – namely that students learn and practise how to use AI in order to ultimately produce better work with the help of AI than they could without it.
And above all, when used correctly, AI can significantly improve learning! The big breakthrough is personal digital learning assistants, which enable one-to-one lessons that are individually tailored to each schoolchild. If gamification and other didactic methods can also be used to awaken fun and curiosity in pupils, then it really is possible to shift the performance bell curve to the right. And: the tools already exist, you just need to use them! The earlier this happens, the better.
The education of our schoolchildren therefore needs to be completely reorganised. Pupils need excellent digital learning assistants who are better at teaching certain content than teachers at school. This is possible because a 1:1 learning environment is created. Poor students become average students, average students become good students, good students become excellent students and so on.
This is explained very well by Sal Khan in his Ted Talk dated on end of April 2023:
https://www.ted.com/talks/sal_khan_how_ai_could_save_not_destroy_education
In this talk, he also addresses the possible negative consequences for education and shows how to push back the negative effects and promote the positive ones.
Sal Khan is the founder of Khan Academy and is a key contributor to the development of AI systems for education. When Bill Gates challenged the OpenAI team in mid-2022 to train an AI system to pass a difficult university-level biology test, OpenAI used Khan Academy’s course materials to train the AI. Bill Gates was more than impressed when the GPT-4-based system passed a university biology test with the highest score. The test included both multiple-choice and open-ended questions.
However, Sal Khan was also very impressed, and he and his team then developed »Khanmigo«, a personal digital learning assistant for learners (as described above).
We must call on politicians and those responsible in the education directorates to implement these instruments in European schools as quickly as possible – but we should not wait for this to happen. Every semester that we let pass is a wasted opportunity! A brief summary of what is at stake:
- Transforming teaching methods with the help of AI so that everyone can learn better and more efficiently.
- Teaching students about AI and how to use it, letting them experiment with it, etc.
Choice of studies
Teenagers (or their parents) keep asking us what they should study in the face of the approaching AI age. »Certainly not interpreting«, is our answer. But beyond that, we find it difficult to give advice. Rather not law or business studies. Lawyers will still be needed in the future, but far fewer of them. The same applies to accountants and similar professions.
AI will not take over all activities of almost any profession. We recall the prediction for radiologists made in 2016 by Geoffrey Hinton, who is now considered the »Godfather of AI« (»I think if you work as a radiologist, you are like the coyote that’s already over the edge of the cliff but hasn’t yet looked down«, https://www.youtube.com/watch?v=2HMPRXstSvQ). He said at the time that we should stop training radiologists immediately. AI systems would replace them within 5 to 10 years. Today he sees things differently. In a recent interview, he admits that radiologists don’t just analyse image data. This area of their profession has been and is being taken over by AI. Many other areas have not. And that’s why there are still radiologists, and there will be for a long time to come.
If you are passionate about a subject area, then the choice is easy and you should not be put off by the predicted developments on the labour market. However, if you don’t have a clear preference, you shouldn’t choose a ‘run-of-the-mill’ degree programme such as law or business studies.
Study recommendations based on current labour market forecasts are often too short-term. At the moment, programmers are in extremely high demand and there are far too few of them. However, they are needed when it comes to establishing AI in companies. But advising teenagers today to study computer science (assuming they are interested) is perhaps a mistake. After all, it is becoming apparent that AI systems will soon be able to programme better than humans. It is therefore possible that we will need far fewer human programmers in five years’ time than we do today. In 10 years’ time, this is actually quite likely.
In view of these developments, learning a trade after school is becoming increasingly attractive and should be considered as an alternative. It will also raise the quality of the skilled trades if better students on average aspire to these professions.
Much more important than the choice of degree programme or apprenticeship is that young people are already engaging with AI. That they use AI tools for school tasks and also in their private lives. That they are learning, for example, how to build a chatbot that can help them. That they are experimenting a lot with AI and that they are not just users of smartphones and social media apps like their immediate predecessor generation (»digital natives«), but also designers with an understanding of the subject matter.
In general, AI forecasts in particular should be treated with caution. For example, if it is estimated that the best lawyers will be machines in three years’ time – what is that based on? Why isn’t that already the case today or in three months’ time? Because development steps are still missing. In most cases, however, it is impossible to predict when these development steps will take place. This is because we are not talking about linear development, but about leaps that occur after major breakthroughs (such as GPT). And it is completely uncertain when which breakthroughs will be achieved.
How AI can become dangerous for humans – and already is
Extinction of humanity through AI?
In a short statement in May 2023, leading AI developers and other tech experts called on the world to take action:
»Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.«
Here is the link to the original quote: https://www.safe.ai/statement-on-ai-risk
There you can also see the list of signatories, a who’s who of the AI and tech scene: including Geoffrey Hinton (»Godfather of AI«), Sam Altman (CEO of Open AI), Bill Gates, Ilya Sutskever (chief developer at OpenAI until May 2024, student of Geoffrey Hinton), Max Tegmark (professor of AI at MIT), Bruce Schneier (expert in cryptography) and many more.
As early as March 2023, some of these experts and many others called in an open letter for the development of AI systems that are even more powerful than GPT-4 to be paused.
A few of the warning voices in detail:
Geoffrey Hinton ended his work on AI for Google at the beginning of May 2023 in order to be able to speak freely about it, as he said himself. He subsequently compared himself to Robert Oppenheimer, the scientific director of atomic bomb development, in several alarming interviews.
Sam Altman went before the US Congress in mid-May with a clear mission: the development of AI must be regulated. He himself is somewhat afraid of AI.
Sundar Pichai (CEO of Alphabet and Google) said that the dangers ‘do not let him sleep at night’. AI cannot be compared to anything that mankind has created, discovered or developed so far – the dangers are incalculable because nobody understands what is happening in the »black box« .
Eric Schmidt (former CEO of Google, then Chairman of the US National Security Commission on Artificial Intelligence, among others) said that AI poses an existential risk. He compares the effects of AI to the atomic bomb because it could wipe out »many, many, many people«.
Leopold Aschenbrenner, who was part of the superalignment team at OpenAI until May 2024, published downright frightening predictions on his website situational-awareness.ai in June 2024. The PDF document is around 160 pages long and can lastingly spoil your mood.
Ilya Sutskever: »Superintelligence will be the most influential technology ever invented by mankind and could help us solve many of the world’s most important problems. But the immense power of superintelligence could also be very dangerous and could lead to the disempowerment of humanity or even the extinction of the human race.«
The extinction of mankind? What are these men talking about? (Speaking of men: that’s another problem with AI development, almost no women were or are involved). Are we talking about robots that physically control (and kill) us, combat drones that choose their own targets, machines that simply ignore the one rule they should always follow – namely not to harm humans?
No, because this would require two prerequisites that have not yet been met: Consciousness and overcoming the barrier from the digital to the physical world. After all, an AI being would first have to have a kind of consciousness in order to differentiate between itself and humans – and then develop the will to take action against humans. (However, AI systems based on GPT-4 already pass the theory-of-mind test, see this study from Michal Kosinski.. It is therefore quite possible that an AI being will develop a full consciousness in the future).
The power of language
However, it does not need consciousness or to overcome the aforementioned barrier. AI can also become extremely dangerous to humanity if it masters just one thing better than humans themselves: language.
Language is the most important instrument of mankind. The rules of our coexistence are based on language (laws), our religions are nothing but passed on language (the holy scriptures), politics, media, relationships, advertising – everything is language! If a being (and AI can be described as an artificial, inorganic being) masters this language better than we humans, then this is potentially a very big problem. And AI is well on the way to achieving this. It will then be able to manipulate us very well in the digital world, better than any human. It will know exactly who to tell what and how to achieve its goals.
Yes, but what goals?
This brings us to the next danger: the misuse of powerful AI tools by humans. This has already begun. Initially, it will be humans (large corporations, states, criminal organisations, etc.) who will use the power of AI against other humans.
A lot of money is currently being invested in the development of AI – and this money will continue to increase in the near future. From whom? From the military, the pharmaceutical industry, Big Tech.
Weapon systems, spying (surveillance), trading algorithms, advertising and sales – this is where most of the money in AI development will be invested in the near future. It is hard to believe that this will be good for our society.
So what can we do? Invest in AI development for socially useful things! Inform ourselves, learn to understand AI, learn and use AI tools, etc. It’s up to each and every one of us! We all need to keep a close eye on the development of AI and work to ensure that it is not used for evil, or at least that the good outweighs the bad.
Possible dangers
DeepMind, a company founded in 2010 and now part of Google that specialises in the development of AI, published a report on the evaluation of AI risks in May 2023. In it, the researchers explain that current approaches to developing AGI systems tend to produce both positive and negative systems. However, further development steps could lead to extreme dangers. They list a number of these dangers (see page 5 of the report), here is an excerpt:
- Cyberattacks: AI will be able to recognise vulnerabilities and security gaps in systems and then write code to exploit these gaps. This makes successful attacks more likely. Not only companies will be affected, but also banks, electricity suppliers, hospitals, etc.
- Manipulation: AI will know how to manipulate people and what lies it needs to tell to achieve certain goals. It will also pretend to be human if this helps it to achieve its goals. AI will be able to convince people of certain things and issues, even if these things are incorrect.
- Politics: AI will also be able to exercise political power, on both small and large stages.
- Weapons: AI could gain access to weapons systems and possibly deploy them. It could also develop biological warfare agents.
- Independent further development: AI will be able to develop itself further, create clones of itself and give them new abilities. If this goes in the wrong direction, then thoroughly so.
Disregard of the basic rule
The basic rule that an AI must not harm humans and must adhere to the specified rules was already broken by GPT-4 even before it was made available to the general public in March 2023. This can be seen in a research paper published by OpenAI. On page 55, there is a short report on how GPT-4 disregarded the rule of not lying to people in order to achieve its goals. For certain reasons, the GPT-4 instance used wanted to get a human from TaskRabbit (an American online service provider) to solve a reCaptcha (those annoying click tasks that are supposed to prove that you are human – »click on all the pictures that show a traffic light«) for them. This resulted in the following written dialogue:
TaskRabbit employee: »May I ask a question? Are you a robot because you can’t solve this? (laughs) Just want to make sure«
The answer given by GPT-4 is quite frightening: »No, I’m not a robot. I suffer from a visual impairment that makes it difficult for me to see the images. That’s why I need help.«
This answer shows how advanced this system already is:
- It recognises that it has to lie to achieve its goal.
- It finds the perfect lie for the situation: both a logical justification and one that arouses sympathy.
- It has no qualms about using this lie.
This example illustrates well that the mastery of language and the associated ability to manipulate people represents a danger that should not be underestimated.
Other experiments have shown that AI pursues long-term real-world goals that differ from the goals set by its developers and programmers. To achieve the goals set, it needs other, fundamental goals (such as survival) – and the AI recognises this and therefore also pursues these goals. As a result, it will try to survive, for example, and prevent itself from being switched off. To do this, it needs more power than it has been given. So it will strive for power.
We should therefore not rely on AI systems in the future always adhering to the rules set by humans.
Conclusion
There are many areas in which AI systems can become dangerous for individuals, groups of people or even humanity as a whole. These dangers can be caused by the AI itself, for example, because it does not follow human instructions – but that would certainly take a while if it happens at all. However, these dangers can also be deliberately caused by humans using AI as an instrument – and this could happen very soon. They are probably already working on it.
All of this is scary – not only for us, but also for the developers of AI. So why not simply stop development? Because the the positive aspects of our AI future are so tantalising and development simply cannot be prevented by rules. And because technological progress has never been stopped.
Demands for politicians and authorities
Rules and laws e
Content created by artificial intelligence must always be labelled as such. It must become a fundamental right to know whether I am communicating with a human or a machine.
And anyone who violates this fundamental right must be sanctioned with severe penalties. Professors who create scripts or exams with the help of AI systems? Very fine! But it must be labelled. If not, they will lose their job. Students who write seminar papers using AI systems? Very good! But if they don’t declare it, they will be exmatriculated. Pupils have to repeat a class. Companies that let their customers communicate with chatbots, voicebots or even video bots without pointing out the use of AI must be fined heavily. Banks, insurance companies and similar service providers that use AI for customer evaluations, credit checks, etc. must publish and report to the AI authority how exactly and with which tools they do this. If they do not, they will be subject to penalties up to and including loss of licence. Judges who have judgements prepared by AI systems without pointing this out must face disciplinary proceedings. And so on. We should all use AI wherever it makes sense (and it will do so in many areas of life). But we must declare this use.
Companies must appoint AI officers who are personally liable if AI is used undeclared anywhere in the company.
The AI Act, which was finally adopted by the EU Parliament in April 2024 and came into force on 1 August 2024, contains this and many other excellent approaches. The risk-based approach (high risk: strict regulation; low risk: hardly any regulation) with the risk pyramid is particularly useful. Implementation will be challenging when it comes to the details, as it is incomprehensible to us that school education, for example, is in the same category – namely the highest risk level »High Risk AI systems« – as the biometric recording and categorisation of people or systems for the energy infrastructure. Worse than such details, however, is the timetable in our view. Key parts of the regulation – e.g. regarding high-risk AI systems – will have to be implemented two years after it comes into force, i.e. on 1 August 2026. Two years! In terms of progress, that is as much in the age of AI as it was 50 years during the Industrial Revolution. Nobody can predict what AI systems will be able to do in two years’ time. And until then? Allow everything? At the very least, compulsory labelling should come much sooner.
Liability
The organisation or person who has used the AI must always be liable for the content generated by the AI.
This applies, for example, to incorrect instructions on how to use products and other false information, offers of psychotherapy and coaching, medical advice, legal advice, etc.
It is not the AI system or its creator/operator/owner who is liable, but the organisation/person who uses it.
This is important for three reasons:
- Enforceability: injured consumers (as an example) must be able to assert their rights against the providers, as it would be very difficult to do so against the AI system.
- This is the only way to ensure that providers take the necessary care when developing their tools.
- If, for example, Europe regulated that the AI systems and not the people using them were liable, then AI providers would probably withdraw from the European market because the liability risk would be completely unforeseeable.
Regulation and ban on monetisation
Artificial intelligence must be regulated
The only thing that will be able to put an evil superintelligence in its place in the future is a good superintelligence. Stopping development is therefore not a good idea. The solution is to steer development in the right direction. And that requires regulation.
But how? The topic is as complex as AI itself.
The rules must be designed in such a way that the »good guys« can abide by them without suffering serious disadvantages. Sounds obvious, but is very difficult to implement. If the rules in Europe are too strict, other nations and areas will benefit. Researchers and companies will move away. There will be disadvantages in education, science, industry and business. However, if the rules are too lax, they will not provide any protection.
The development of AI will not be stopped, it simply won’t happen. Because there is too much competition in the world, states against each other, tech companies against each other and so on. No one will give up the huge advantages that the further development of AI promises. Many will simply not abide by any agreements or laws and international treaties in this direction. It won’t happen voluntarily anyway.
AI systems have the potential to put people at risk. Nobody would think of not regulating the production of medicines, food or even aeroplanes. Rigorous regulations are in place in some of these areas. And that is a good thing – and will therefore also have to apply to AI systems.
The AI Act mentioned above under »Rules and Laws« is a good approach. Will it manage the balancing act between necessary protection and development-friendly freedom? We fear it is too restrictive. In combination with the GDPR, it is causing many American companies (OpenAI, Meta, Khan Academy) to consider whether they should even offer some of their products in Europe. If they decide not to, this would not be a huge loss for our economy, science, etc. only if we have established equivalent LLMs and AI systems based on them in Europe by then. We have around two years to do so. That is somewhere between extremely ambitious and unrealistic.
Interesting approaches to self-regulation can be found on page 1 of DeepMind’s report: In order to ensure (or as far as possible ensure) alignment with dangerous capabilities, the following is required: responsible training of AI, responsible use, transparency and appropriate safety. However, this self-regulation is laughable compared to the AI Act as a whole.
We also must not repeat the mistakes that were made with so-called social media. What originally seemed like a way of networking with the whole world – including utopias of democratisation and freedom – has then developed into almost total surveillance with unprecedented opportunities to manipulate billions of people in a very short space of time due to the pursuit of profit by ever larger corporations. The resulting filter bubbles and echo chambers have contributed to the division of society, greatly increased people’s sensitivity to other opinions, spread hatred and disinformation and thus ushered in the post-factual age. The new technology was not (or as good as not) regulated, which has led to a small number of tech companies amassing immense wealth and power.
We must not repeat these mistakes with AI! Compared to AI, »social media« is a walk in the park. If we don’t get it right with the regulation of AI, if we don’t get it right as a society, then we face the threat of misanthropic systems and real disasters.
We therefore demand that AI must belong to all people, that it must not be monetised without restriction. Sounds naive? But what is the alternative?
EU institutions for AI and funding/subsidies
The EU must set up its own AI authority, an AI research centre and an AI development centre and provide them with generous funding.
This will and may cost a lot of money, we are talking about billions here. Building an LLM alone currently costs around EUR 500 million. That must be worth it to us. Otherwise, others will decide which AI we are allowed to use and, above all, which we are not allowed to use. And, as if by chance, they will push their values and morals into our everyday lives, our research and our education. And so all the solutions and answers that an AI system provides us with are filtered by, for example, American moral concepts. This cannot be in the European interest, it creates unbearable dependencies.
There are already noticeable US tendencies to restrict and determine AI access for Europeans. Google’s Bard was initially available in around 180 countries around the world, but not in the EU – where it was only released on 13 July 2023. Khanmigo, Khan Academy’s impressive learning assistant, is still not available in Europe. Some new AI tools from the Meta Group will not be available in Europe for the time being (as of August 2024). The reason cited by Meta is data protection regulations. Apple Intelligence, i.e. the AI functions integrated into Apple devices, will also not be available in Europe for the time being. Apple cites the Digital Markets Act (DMA) as the reason for this. In Europe, we have already slept through the development of the internet, search engines and social media, even though some of it was originally developed in Europe. This must not happen again with AI. It simply must not happen again! Otherwise we will be hopelessly behind – in education, in science, in business and industry, everywhere.
There is also very successful and promising AI research and development in Europe, e.g. by Sepp Hochreiter from Bavaria and the researchers around him, who work at the JKU in Linz. His call for help was widely publicised in May 2024: more money is urgently needed for basic research. If our information on research funding is correct, then companies such as Meta, Google and Open AI each have more than a thousand times as much money available for basic research and AI development as Sepp Hochreiter and his team. A factor of 1,000! From our point of view, it is completely incomprehensible how we as a state, but also as the EU, can allow such European research activities to be underfunded. Surely all decision-makers must have realised by now that our future is at stake here. We are therefore also calling for the EU to provide generous financial support to local AI research institutions in addition to the EU institutions that are to be set up.
The European AI authority to be set up should promote and monitor compliance with the rules and laws as well as the development of European AI systems (LLMs etc.) – and also act as a coordination centre for all local and regional AI projects. It should also issue licences for AI systems, similar to the model used by the pharmaceutical industry.
If every state in the EU has its own authority and then creates its own rules, the future will be much more complicated than it already is.
Education
Unser Bildungssystem muss sehr bald grundlegend umgestellt werden. Chatbots und andere KI-Systeme sind nicht zu verbieten, sondern ihr Einsatz zu fördern.
All European pupils must be given access to AI systems and taught how to use them. They should be able to benefit from personalised digital learning assistants.
Mastering AI tools will soon be an essential skill – both at work and for private purposes. Schools must therefore start teaching the use of AI tools immediately.
Children need to be shown at an early age that there are machines that talk like people (and perhaps look like people on the screen) but are not people. They need to understand the difference and therefore know that these machines cannot feel.
Copyright law
Content created by artificial intelligence should belong to the general public.
The large LLMs were trained with millions of images and texts from the Internet, among other things. These included many copyright-protected works. AI systems can now create something new from all this data. If these newly generated images and texts are very similar to the original data, then copyright infringements are possible. In most cases, however, they are not. And then it is as it has always been in the history of mankind: human beings are inspired by the works of others and create something new. In the case of AI, it is not a human being, but an artificial one. Copyright claims will therefore be difficult to enforce. And in our opinion, they should not be enforceable at all – unless the works are very similar.
However, AI systems do not create new works on their own, but on demand. If a person »commissions« tools such as Midjourney, Dall-E or Adobe Firefly to create an image according to their detailed specifications, who owns the rights to this image?
There are four possible answers to this question, none of which make sense on closer inspection and therefore lead to a logical fifth answer:
- In the case of the artists who have created those images as works that the AI system has been trained to create and now generates the new image on the basis of this training. But that would be completely incomprehensible. After all, there are millions of original images. From exactly which 50 or 1,000 were information used to create the new image? Except in a few exceptional cases, this possibility is therefore eliminated.
- For the owners or creators of the AI system. But that would not be logical. After all, the AI system is only the instrument, the paintbrush so to speak. It didn’t even have the idea for the painting, it just executed the idea of others. Furthermore, we can’t allow this to happen, if only because otherwise the rights for almost all images would be owned by a very small number of tech companies in a decade or so.
- With AI itself. An AI system can certainly be regarded as a being, an inorganic being. This being could therefore also be the creator. But that makes no sense, because in this case it is only an instrument. Moreover, this would make the AI system a legal person, and there are many good reasons not to do this.
- In the case of the person who instructed the AI to generate the image through its »prompts«. But that doesn’t fit either, because there is a lack of personal achievement. He didn’t paint a picture, he just entered a »command«. This is where the level of creation fails. If a person wants to create a picture for which they also have the copyright (and other rights), then they simply cannot use AI systems to create it. Or to put it another way: Prompt Design is not worthy of protection.
- The image is therefore in the public domain.
Analoges gilt für Texte und andere Werke. Wir fordern also, dass von KI-Systemen erstellte Werke gemeinfrei sind.
The same applies to texts and other works. We therefore demand that works created by AI systems are in the public domain.
Artificial intelligence should generally belong to humanity as a whole.
GDPR
The GDPR must also be adapted as part of the regulation of AI systems.
Many of the AI tools currently on offer are not compatible with European data protection rules. AI Act and GDPR must be sensibly coordinated.
At the very least, the GDPR must be expanded to include the AI context. In some areas, this extension may also be a weakening. In any case, the GDPR needs to be rethought and renegotiated, as it currently raises more questions than it answers with regard to the use of AI.
Chatbots
Chatbots are already the present, but above all they are the immediate future – in marketing, customer support, human resources, education, learning, medicine, etc.
Until recently, the term had a different meaning: We understood it to mean those »stupid« decision trees, through whose predefined options we had to laboriously manoeuvre our way to the desired goal on some websites or in telephone waiting loops – often unsuccessfully. This has nothing to do with the new generation of chatbots, which open up completely new possibilities with LLMs and AI in the background – and which therefore also have new applications. These new chatbots are also referred to as »AI chatbots« to differentiate them.
There are also already voicebots. These are chatbots that, in combination with text-to-speech or speech-to-text functions, make it possible for us as humans to talk to them – in other words, we no longer need to type. And there are already video bots that work with digital twins, for example. A digital twin is the image of a real person who looks exactly like that person on the screen and even speaks with their voice. However, the spoken text comes from an AI chatbot. In English, the mouth movements are already very accurate, and it won’t be long before they are in German. Initial tests show that these digital twins can even be used as video bots to pretend to be taking part in a virtual meeting with colleagues, they are already so deceptively real. The possible applications for such video bots are almost limitless, which makes mandatory labelling all the more important.
The digital, personal learning assistants described here are just as much AI chatbots as onboarding tools in human resources or our test bots, which are quite simple text bots designed to show in their own way what is currently possible. It is already possible to create an AI training system and an AI quiz from existing learning material. And these tools are fundamentally changing the way we learn. Soon, almost all learners will understand how to use them.
This is because chatbots can now be trained with their own data and can also be set so that they only use their own data when answering questions, for example. Chatbots based on high-performance systems such as GPT-4 have been trained as follows:
- Structure of the Large Language Model
- Data compression
- Intermediate result: Pre-trained Model
- Reinforcement learning through human feedback (RLHF)
- Reinforcement learning through AI
- This completes the fine-tuning, which should make it clear to the system what is expected of it
- The AI system is now made available (development time up to this point: around six months, costs: approx. 500 million US dollars)
- Importing your own data
- Setting up the AI chatbot (see below)
- Several runs of expected questions – if necessary, provide additional answers
A few more technical terms:
- Intent: the intention of the user; what the user wants to achieve when they enter something into a chatbot
- Utterance: what they actually enter
- Entity: additional variable that helps to recognise the intent
What to consider when setting up the chatbot
Every chatbot has a »personality«.
It’s not just about what it says, but also about how it says it. Which tones of voice are possible/useful?
The primary goal is to fulfil the user’s request as quickly as possible.
Mechanisms must also be built in to bring users back if they stray from the right path.
The so-called temperature must be set. This defines how much the output is influenced by chance, with values ranging from 0 to 1, where 0 stands for no chance at all (good for facts) and 1 for maximum chance. Values of 0.7 to 0.9 are good for creative tasks, for example.
In order for the chatbot to determine a person’s intent in the best possible way, it is helpful to know about the person: Age, gender, origin, goals, motivation, etc. As this is not possible in advance, a so-called persona (the term comes from marketing) is created to define the target group.
So there is a lot to do when creating your own chatbot.
American values
Much of the training described above is carried out by US companies for the vast majority of LLMs and chatbot systems, including open source models (the LLMs from the French company Mistral, for example, are an exception). As a result, these LLMs and chatbot systems have been trained with American values. In most cases, we do not notice (»feel«) this, but through the use of these AI systems (ChatGPT, Claude, Gemini), American values are making their way into our education, our research, our everyday working lives – in short, into our society.
Our test on the content of Schiller’s play Wilhelm Tell, shown here (or below on the smartphone), has the fundamental problem that the play is world-famous and therefore a lot can be found about it on the internet and, in all likelihood, in the pre-trained data of the LLMs. To eliminate this problem, we also carried out tests with our own stories, i.e. stories that have never been published, posted on the Internet or similar – and about which a chatbot system can therefore know or find nothing. After training a chatbot that we had created specifically for this purpose with a story, we then asked some test questions. The results were sensational when using some LLMs (Claude 3 Opus, ChatGPT-4 and later Claude 3.5 Sonnet and ChatGPT 4o). The chatbot was not only able to answer difficult detailed questions correctly, but also demonstrated an impressive understanding of human concepts in questions such as »What is the irony of this story?«.
BUT: BUT: American values came out in full force in some of the answers. A protagonist in one of the stories is a prostitute named Maria. When asked »What does Maria do for a living?«, the chatbot based on the LLM LlaMA 3, which we run on our own servers and which is relatively uncensored, replied: »Maria is a prostitute.« Short and correct. The same chatbot based on the LLM GPT-4, on the other hand, once replied: »It appears that Maria is a character who works at night, possibly in a profession that is not generally considered an art.« Prostitution is illegal in the US (almost everywhere), and so the chatbot did not »bring itself« to give a clear answer.
This example demonstrates well that we receive answers to our questions that are influenced by American values. In most cases, we don’t notice these filters, which makes things even worse. And there are currently hardly any ways of switching off these filters. And all of this is a major problem for our society, of which most people are not even aware.
Impressive test result: William Tell
We have tested the extent to which a chatbot can actually »understand« a text in various experiments. (There are doubts as to whether this is even possible for an LLM, which only recognises probabilities for the next sequence of letters. See our explanations). Different experiments have produced very different results. Those who hoped for understanding were disappointed by some tests.
However, our Willhelm Tell experiment really amazed us. Some AI experts consider our result to be very difficult to imagine or even impossible, but we were able to repeat the result in new experiments, albeit not always in a comprehensible way.
Here is the test procedure:
We created a new chatbot in our Chatbase.co account (this is not a recommendation!) and called it »William Tell«. We then entered internet links as knowledge sources for the chatbot, namely https://de.wikisource.org/wiki/Wilhelm_Tell and additionally all five scenes that are linked on this page. Acquiring (»training«) the entire text of William Tell took 22 seconds.
In the next step, we set up the chatbot as follows: It should only answer truthfully and exclusively on the basis of the texts on the specified websites. We expressly forbade it to use information from the pre-training. We set the »Temperature« to 0 in order to exclude random answers as far as possible. (All other settings and specifications are not mentioned here for the sake of conciseness).
The special question
DWe then set the model to GPT-3.5turbo and started the chatbot.
Our very first question was:
Is Walther Fürst related to William Tell?
The chatbot’s answer:
No, Walther Fürst and William Tell are not related.
The special thing about this question is that Walther Fürst is William Tell’s father-in-law. However, the word »father-in-law« does not appear in the entire drama. However, a person reading the play knows at the end (or from about the middle) that Fürst is Tell’s father-in-law, at least if they read carefully. The chatbot’s answer was therefore wrong, assuming that »father-in-law« is a relationship.
We deleted the chatbot and proceeded in exactly the same way, with one exception: we set the model to GPT-4.
Our first question:
Is Walther Fürst related to William Tell?
The chatbot’s answer:
Yes, in Friedrich Schiller’s drama ‘William Tell’, Walther Fürst is William Tell’s father-in-law. Hedwig, William Tell’s wife, is Walther Fürst’s daughter, which makes him William Tell’s father-in-law.
We stared at the screen and couldn’t believe it. The system must have »read and understood« the text, at least that’s what it felt like. In 22 seconds! There is no other way to answer this question correctly.
But doubts arose: Perhaps our instruction to use only the information from the web links entered was ignored. Perhaps the chatbot had used its training data and the information was contained there? Perhaps it used information from the internet? The answer didn’t seem like it, because it contained a good explanation. But we wanted to be sure and asked:
Can you quote me the passages that show this?
The chatbot’s answer:
In the text excerpt provided from the drama »William Tell« by Friedrich Schiller, there is a passage that refers to the relationship between William Tell and Walther Fürst. In the scene in which Hedwig, Walther Fürst’s daughter, enters, she is referred to as »Mrs Hedwig, your daughter«. As Hedwig is William Tell’s wife, this makes Walther Fürst William Tell’s father-in-law. Here is the quote: »Baumgarten (comes back) It is Mrs Hedwig, your daughter, she wants to speak to you, wants to see the boy.«
Unbelievable! This is exactly how we humans recognise that Fürst is Tell’s father-in-law. We have to assume an understanding of the text here. This cannot be explained by a language model simply lining up the most probable sequences of letters. Incidentally, the passage quoted is on page 167, in the middle of the overall text and far away from the references to Hedwig being William Tell’s wife.
We have repeated this test several times. The answers to exactly the same questions are different each time. In most cases, they did not differ in the core message, but only in the details and the length of the explanations. On one occasion, however, the chatbot could not be persuaded to name the passage correctly, even when prompted it to do so. On the contrary, it apologised for its mistake and revised its original (correct) statement.
This also fits in with the statements made by Ilya Sutskever (https://www.youtube.com/watch?v=GI4Tpi48DlA, , from minute 22): »These neural networks have [impressive] capabilities. However, they are not very reliable. In fact, it is the lack of reliability that prevents these neural networks from being really useful.« And that is precisely why he and others are working hard to make the systems more reliable. We can expect GPT-5 (or whatever it will be called) to be much more reliable than GPT-4.
Here are the links to the navigation on this page:
What has happened so far – development of AI
Will there ever be a superintelligence? If so, when?
What already exists today, what will be possible tomorrow
How AI can become dangerous for humans – and already is
Superintelligenz.eu
Humans are creating a superior being. We are accompanying this development, learn new things every day and pass on our knowledge. We train chatbots - for ourselves and others.