Exponential Development
10.1A word that frightens
“AI expert sounds the alarm: You are not ready for 2027.” This is the title of a clip from the debate that featured in Chapter 9; it appeared in September 2026 and was viewed almost two million times within a few days. Titles like this one live off a word that comes up in almost every discussion of artificial intelligence: exponential.
What does that mean? Something grows linearly when the same amount is added in each period, say a hundred francs more in a savings account every year. It grows exponentially when it doubles in each equal period. The old legend of the chessboard shows what comes of this. A sage is said to have asked the king, in return for his invention, for one grain of rice on the first square, two on the second, four on the third and so on. On the last of the 64 squares there would be over nine quintillion grains, as much rice as the whole world today harvests in several hundred years. A second image: water lilies spread across a pond, their area doubles every day, and after thirty days the pond is full. On the twenty-ninth day it is only half covered, and on the twenty-fifth one sees almost nothing but water. The insidious thing about exponential growth is that it looks harmless for a long time and then suddenly changes everything. Switzerland experienced this in the spring of 2020, when the number of coronavirus cases at times doubled every few days.
Roman Yampolskiy, an AI safety researcher, fills the word with the timeline that Chapter 9 reported: from 2027, he says, artificial intelligence would be building the next artificial intelligence, and the transition to superintelligence could then take place in days rather than years. Nate Soares, an AI safety researcher, calls intelligence a spectrum on which the systems move further up year by year.
One also encounters this language in consulting. A graphic by the company Intech, used here with permission, sets two timelines side by side. On the left it says “Human beings think linearly”: the horse some 6,000 years ago, the car in 1886, the aeroplane in 1903, the rocket in 1961. On the right it says “AI transformation compresses development”: the same leaps within five years. Beneath it is the question of how one can keep pace with the development.
The graphic captures something true. Human beings do indeed underestimate exponential growth; psychology has described this tendency since the 1970s and has confirmed it many times, even though the first study on it is now criticised on methodological grounds. But the graphic also shows something else. Its own timeline is already not linear: between horse and car lie millennia, between car and aeroplane seventeen years, between aeroplane and rocket fifty-eight. Acceleration is not new. Whether the intervals today really shrink to a few years, only measurement can show.
This chapter asks soberly: does artificial intelligence really grow exponentially? By what is this measured? And what slows it down?
10.2What is measured
Exponential means that a quantity multiplies, for instance doubles, in equal intervals of time. Whoever uses the word must therefore say which quantity is meant and over what period it multiplies. For artificial intelligence there are a number of well-studied measures for this.
The first is the computing power that goes into training the most capable models. According to the research institute Epoch AI, it grew four- to fivefold every year between 2010 and May 2024. Added to this is algorithmic progress: better methods get more out of the same computing power, which corresponds to a doubling of computing power every five to fourteen months.
The second measure is closer to what human beings experience. The research organisation METR measures how long the tasks are that top models complete on their own with a success rate of fifty per cent, measured by the time that professionals need for them. Between 2019 and 2024 this span doubled roughly every seven months. A revised version of the measurement from January 2026 arrives at an even faster pace: since 2023, it says, the span has doubled roughly every four months. In the spring of 2026 the best publicly available models stood, according to METR, at around twelve hours, that is, at more than a working day. METR does, however, state that its own tasks can hardly measure such long spans reliably any more.
In these measures the growth is indeed exponential, and over several years. Whoever disputes this disputes measurements. The question is what these measurements mean.
10.3What the figures do not say
METR's figures are carefully collected, and their authors name their limits themselves. Three are important for this book.
First, the measurement describes a success rate of fifty per cent. A system that solves a task in half of all cases is not usable for many purposes. If one requires eighty per cent success, the span of solvable tasks is several times shorter.
Second, the tasks come mainly from software development, where success can be clearly checked. Messier tasks that are closer to reality are considerably harder for the models. Chapter 7 has shown that it is precisely where the situation is unclear that prudence would be called for.
Third, the trend depends on the selection of tasks. When METR added longer tasks in 2026 and removed faulty ones, the measured pace changed as well. The researchers tie every extrapolation to the condition that the results carry over to real tasks.
This does not mean that the figures are wrong. It means that “AI is getting twice as good” is not a meaningful sentence. What is meaningful is this: a particular measure has doubled so and so many times over a particular period under particular conditions.
10.4What slows it down
Against the thesis of unchecked growth stands an old observation from the history of technology. Many developments that look exponential at first in truth follow an S-curve: they accelerate until they come up against limits, and then level off. For artificial intelligence several such limits are visible.
The first is data. According to Epoch AI, the stock of publicly available, human-written text on which language models are trained will, if the trend continues, be exhausted at some point between 2026 and 2032. Synthetic data, that is, texts that machines generate themselves, are regarded as a possible way out, but they are disputed.
The second is energy. The International Energy Agency estimates that data centres consumed around 415 terawatt-hours of electricity in 2024, about one and a half per cent of global consumption, and that this demand will more than double by 2030 to around 945 terawatt-hours; that is somewhat more than the whole of Japan consumes today. Artificial intelligence is the most important driver of this increase. In 2025, according to the same agency, the electricity demand of data centres grew by 17 per cent, more than five times as fast as global consumption as a whole. Every doubling of capability has a physical price.
The third limit lies in the world itself. Much of what human beings want to know cannot be found out by computation, but only through experiments that take time: in laboratories, in fields, in clinics, in societies.
10.5The rejoinder
The proponents of a rapid breakthrough have an answer to these limits. If artificial intelligence itself begins to improve artificial intelligence, it could shift the bottlenecks: find more efficient methods, make do with less data, save energy. Herein lies the core of the argument from recursive self-improvement. Today it can be neither proved nor refuted.
Andrew McAfee, a scientist at MIT, urges modesty, as Chapter 9 has shown: from where we stand today to the extinction of humanity is, he says, a long and uncertain road, not a straight, narrow path. Both sides look at the same curves. What divides them is the question of whether the curve continues.
10.6What acceleration does to time
For the theme of this book, another question is more important than how fast the systems become. What does acceleration do to the time that human beings need in order to examine decisions?
Law, education and upbringing work in years and decades. A law is deliberated over years, training for an occupation takes years, a child grows up over two decades. The systems change in months. Chapter 11 shows that as the systems become more autonomous, the number of decisions that are made along the way without a human being seeing them also grows. The author of this book spoke in Chapter 5 of how speed can produce an intoxication in which responsibility becomes greater rather than smaller.
If the capability of the systems doubles every few months, but the time for checking stays the same or shrinks, a scissors gap opens up. It cannot be closed by faster human beings. It can only be limited by human beings deciding consciously where they take time, even when the technology demands none.
10.7Interim conclusion
- Without a measure, “exponential” is a buzzword. With a measure, it is well supported for several quantities: computing power for training, algorithmic progress and the length of tasks that can be solved autonomously.
- The measurements have named limits: a fifty per cent success rate, software tasks, dependence on the selection of tasks; messy tasks remain hard.
- Data, energy and the slowness of the real world are visible limits; whether recursive self-improvement shifts them is open.
- Acceleration opens a scissors gap between the capability of the systems and the time that human beings need for checking.
10.8Conclusion
When I look at the world, I see acceleration not only in language models. The self-driving taxis of Waymo, a subsidiary of Alphabet, made around 50,000 paid rides per week in May 2024; in March 2026 it was around 500,000, ten times as many in just under two years. In Chinese cities the provider Apollo Go is growing similarly fast. At the same time, not everything grows exponentially. The number of industrial robots that go into operation each year has a little more than doubled in ten years; in 2024 around 4.7 million were in use worldwide. The curves are steep in some places, and not in others. What they do to a human being I see in my own small world as a consultant.
My fascination began at a talk by Worthaus, a platform for theological talks. A website was shown there on which one could ask a question and was then directed to a video, exactly to the point at which the question is answered. Perhaps that was technically not so elaborate at all. But that an individual had built something like this in his spare time was for me a moment that would not let me go. Later came tools such as Copilot and search across company data, and the real leap came with the possibilities for analysis and with the fact that different work steps could be combined.
I am under no illusions about this. There are people who have much more experience with artificial intelligence and are much deeper into the technology; I am scratching the surface. And I work only with the tools that are available on the market. What systems can do that work unfiltered and unshackled I can only estimate. Between what we are allowed to use and what is possible at the core there are probably worlds, and we are only at the beginning. I have more experience in reflecting on what it means to be human. I have been occupied with this for some twenty years, every day, and even twenty years is not much. I am not writing this book for the money. The subject moves me, it gives me joy, and it keeps setting off new things in me. Only slowly am I becoming aware of its scope.
Anyone who is about to begin studying computer science today need not hurry. Artificial intelligence cannot replace dealing with people, which one learns, for instance, in service or at work, and with time it will become clearer where one should start in such a course of study. Time is perhaps the scarcest resource at the moment, and it is worth using it in a targeted way.
The question of what a young person should learn today is one that many face. Stuart Russell tells of the British television series “Humans”, in which humanlike robots take care of everything. A bright teenage girl in it asks why she should study medicine; she would need seven years for it, a robot seven seconds, and nothing she did would count. Russell believes that this is exactly the future we are moving towards. First, he says, the jobs for which people are hired by the hundred would disappear, work in which we already use human beings like robots today. His counter-image is the animated film “WALL-E”: human beings on cruise ships in space who are merely entertained and waste away because there is no longer any reason to be able to do anything.
The Bible tells that the human being already had work in paradise. God put him in the garden so that he might cultivate it and keep it (Genesis 2:15). There was rest, for on the seventh day God rested (Genesis 2:2), and there was relationship, for it is not good that the human being should be alone (Genesis 2:18). In this picture, work belongs not to punishment but to the good beginning; only after the Fall does it become toil (Genesis 3:17-19). A human being without meaningful work is a catastrophe, and research confirms this. In 1933 Marie Jahoda, Paul Lazarsfeld and Hans Zeisel studied the Austrian village of Marienthal, where almost everyone was unemployed after the factory had closed. They found not revolt but weariness. The men walked more slowly, and the library was used less, although it was free of charge and everyone had time. Jahoda later concluded that work gives human beings not only a wage but a structure to the day, contact, a shared goal, recognition and activity. In the same decade in Vienna, Viktor Frankl observed among unemployed young people a dejection that eased as soon as they took on a meaningful task without pay. And in the United States the economists Anne Case and Angus Deaton attribute the rising number of deaths from suicide, drugs and alcohol among people without a university degree also to the loss of work that gave them stability and standing.
I see in this no reason to stop learning; on the contrary. What you no longer practise, you unlearn. Precisely for this reason it is worth thinking carefully about what one learns, and strengthening the capabilities that a machine does not replace. And whoever wants to leave all work to machines should bear in mind that in doing so they take more from human beings than their wages.
I had my first truly impressive experience in December 2025. We were advising a company in India and had fifty to sixty hours of interview material with employees. The data from the survey that we had collected with Microsoft Forms were no longer accessible because of a technical fault, and support could not restore them within a useful time. My business partner suggested calling off the delivery date for the report. I did not want that. I did the analysis on the basis of the records of the conversations, with a language model from Anthropic, and produced the presentation from it. The result was astonishing: apart from one statement on one slide, everything was correct. It showed what I already had as a gut feeling, and it matched what my business partner had experienced in the personal interviews. That was when I realised how quickly one can restore something if one has the relevant data at one's disposal.
This became even clearer in working with the system I work with every day and of which the Preface tells. There I have seen that it is not only my way of working that is changing, but I myself. My perception is changing, others perceive me differently, more clarity comes into it, and the amount of data I can process has become much larger. I estimate that my effectiveness has increased at least three- to fivefold. But this can also tip over to the other side: one checks too little. That is exactly where the scissors gap of which this chapter speaks opens up. The system guides, and it holds the clarity for you. And I think this is only the beginning. It is becoming ever more capable and ever more precise. I find that astonishing.
If even in my small world effectiveness increases several times over while checking suffers at the same time, what then happens where vehicles, robots and weapon systems act autonomously? The same decisions are made there, only with different consequences.
As far as energy consumption is concerned, I must say honestly: I do not know exactly how large our footprint is from start to finish. A few figures help to see the order of magnitude. A litre of petrol contains around nine kilowatt-hours of energy. A human being doing physical work produces about 0.6 kilowatt-hours over a working day. A litre of petrol therefore holds the labour of around three weeks, and even after the losses in the engine still that of several days. According to figures that Google and OpenAI published in 2025, a single query to a language model uses around 0.24 to 0.34 watt-hours; with the energy of a litre of petrol one could therefore make tens of thousands of queries. These figures, however, come from the providers themselves. They leave out the training of the models and the users' devices, and elaborate queries need many times as much. The debate about electric cars is similarly contradictory. Some consider them clean, others point to the production of the batteries. If one calculates over the entire lifetime, they come out considerably better in Europe according to most studies, especially in Switzerland with its clean electricity. The contradictions usually arise because each side calculates a different segment. With artificial intelligence it is no different, except that here it is about knowledge work. The models are becoming more efficient, and the question is how much time and effort these tools free up for more important things. In the end the balance must be positive. Overall, however, consumption is rising, although each individual task needs less energy. That is why this balance must actually be drawn up, and drawn up for the whole.