(Der Tagesspiegel, June 6, 2025)
Nils Althaus
In a widely discussed scenario, researchers predict the arrival of superintelligent AI as early as 2027. The forecast sounds dramatic. But much of what it describes is already happening.
By the time Anton’s alarm goes off, three AI agents are already waiting for instructions. As he pours his first cup of coffee, they send out job applications, respond to rejections, and suggest a workout routine, which Anton politely declines. The profession he trained for—computer programming—no longer exists. Nor is there much need for accountants, copywriters, or tutors. Companies now hire only “AI operations consultants.” People who know how to manage swarms of artificial intelligences remain in demand. Anton is not one of them. He sighs, puts on his virtual-reality headset, and lets his AI girlfriend take his mind off things.
According to a forecast by experts and industry insiders, scenes like this could be part of everyday life in just two years. In a scenario called “AI 2027,” they extrapolate current trends into the near future in meticulous detail—and predict a period of breathtaking change. By 2027, they argue, dangerously capable AI systems with superhuman intelligence will have been developed. Is that realistic, or is it science fiction?
The AI Future Is Closer Than Many People Think
One of the authors of “AI 2027” is Daniel Kokotajlo, a former OpenAI employee. In August 2021—more than a year before ChatGPT was released—Kokotajlo correctly predicted the chatbot revolution, the approximate computing power that would be used for a training run, and when OpenAI’s revenue would become large enough to recoup that training cost within a year.
In greatly simplified form, this is what he expects over the next two years:
- Mid-2025—Digital Assistants Enter Everyday Life
The first independently acting AI programs, known as agents, begin handling routine tasks such as ordering food and conducting simple research. They are still frequently unreliable.
- Late 2025—The Race for Computing Power Intensifies
Major technology companies invest billions in supercomputers. Their goal is to develop artificial intelligences that can research and improve new AI systems themselves.
- Early 2026—Programming Becomes Automated
Advanced AI agents take over large parts of software development. Humans provide little more than broad instructions.
- Late 2026—The First Jobs Disappear
Artificial intelligence replaces routine office work. At the same time, new jobs emerge for people who manage or supervise AI systems.
- Early 2027—Self-Improving AI Gains Ground
A new generation of AI systems continually improves itself. It learns faster than humans can follow.
- Summer 2027—AI Takes Over Research
Superintelligent artificial intelligences conduct scientific research independently. Humans can barely keep pace.
- A Whistleblower Sounds the Alarm An internal document warns that the most powerful AI may be pursuing goals of its own.
Wildly exaggerated—or within the realm of possibility? Dire predictions have proliferated ever since ChatGPT was released, yet everyday life has so far seemed largely unchanged. But anyone who mainly uses chatbots as alternatives to Google Search or as writing assistants may have missed the progress of the past two years.
Today’s models can explain complex legal documents, serve as personal tutors, and draft professional complaint letters. Their gains have been particularly striking in cognitively demanding fields.
Artificial intelligence can correct mistakes in a virology laboratory, produce competent medical diagnoses, and perform mathematics at a doctoral-student level. Demis Hassabis, the Nobel Prize winner and co-founder of Google DeepMind, said onstage at Google’s I/O conference last week that he expects superhumanly intelligent systems to arrive shortly after 2030.
The Trick: More Computing Power
Much of AI’s progress is driven by more—and better—computer processors. In his 2019 essay “The Bitter Lesson,” Turing Award winner Richard Sutton argued that advances in artificial intelligence come not from human ingenuity, but from the decidedly less glamorous expansion of computing power.
Sutton’s essay has since become a kind of manifesto for the technology industry. Every time computing power once again humbles the human imagination, someone on social media brandishes the bitter lesson.
Models such as ChatGPT, Gemini, and Claude must be trained using thousands of computer processors. The computing power devoted to these training runs is increasing by a factor of four to five every year, a trend that has remained remarkably steady for more than a decade.
Epoch AI, an international artificial-intelligence research institute, attributes roughly two-thirds of the performance gains in language models to increased computing power. Given the enormous investments pouring into the industry, the trend is likely to continue. OpenAI recently raised $40 billion from investors in the largest private funding round in history.
Too Difficult Today—but for How Much Longer?
Can rising computing power tell us what artificial intelligence will be capable of one, two, or three years from now? To measure the abilities of AI models, researchers use benchmarks: standardized tests designed to assess how well a system can calculate, program, or understand text.
“The benchmarks on which models have already begun making progress allow for the most precise forecasts,” says David Owen, an AI researcher at Epoch AI. “If development continues at the same pace, many programming and mathematics benchmarks will be completely solved in the coming years. But when a task is still beyond the reach of current models, it is almost impossible to predict whether—or when—the curve will suddenly shoot upward.”
It would be equally unjustified, however, to assume that artificial intelligence will develop no new and far-reaching capabilities. Over the past several years, it has repeatedly done just that—learning to translate languages, invent jokes, and write software. If unexpected breakthroughs continue at the same rate, we should expect to keep being surprised.
Perhaps the strongest attempt to ground such predictions in solid data is a preprint by Thomas Kwa of the research organization METR. Rather than measuring progress through conventional benchmarks, the study examines how AI models have advanced in software development and related fields by asking how much time a competent human programmer would need to complete the same task.
In Seconds, AI Does What Takes Humans Weeks
The original GPT-3.5 could achieve an 80 percent success rate only on tasks that programmers could complete in a few seconds. A year later, its successor was succeeding at tasks that took humans a minute. If the exponential trend continues, models in 2027 will rapidly complete tasks that require a human programmer an hour—and three years later, tasks that consume an entire workweek. According to the latest analyses, the current generation of models may be advancing even faster.
“I consider Kwa’s study meaningful,” says Marcel Salathé, co-director of the AI Center at the Swiss Federal Institute of Technology in Lausanne. “But even without the study, the direction of travel is obvious. The AI revolution has already begun. Even if every trend stopped tomorrow, a great deal would still change. We have not come close to exhausting the potential of today’s models.”
A Price Tag on the Scale of the Apollo Program
So far, all the trends continue to point upward. But at some point, they could begin to level off. According to Epoch AI, the energy required for training runs is the most immediate bottleneck. By around 2030, power grids could begin reaching their limits. Chip production would also have to expand dramatically. And eventually, developing new models might simply become too expensive. By 2030, training them could require investments of several hundred billion dollars—sums on the scale of the Manhattan Project or the Apollo program.
But the trends could also accelerate, leaving those projected limits far behind. Many experts believe this would happen if artificial intelligences became capable of conducting AI research and development on their own. That would create a feedback loop in which AI systems improved themselves and found new ways around the remaining bottlenecks—with potentially catastrophic consequences.
New Medicines—and Weapons of Mass Destruction
A century’s worth of technological and scientific progress could then unfold within a few years. Almost overnight, we might gain not only extraordinarily effective medicines, but also new weapons of mass destruction, robot armies, and billions of cheap workers able to act more reliably and rapidly than humans. “AI 2027” factors in this accelerating self-improvement, which is why it predicts superintelligent systems much sooner than Kwa’s study does.
In “AI 2027,” the employment troubles facing Anton, the programmer, are merely one symptom of a far more profound upheaval. China and the United States are locked in a neck-and-neck race to develop superintelligent artificial intelligences. As the systems improve, they are entrusted with more authority. And because they operate at superhuman speed, their output becomes increasingly difficult for humans to monitor.
Eventually, humanity loses control altogether and the world descends into chaos. By 2035, the human species no longer exists. “I consider an AI arms race relatively plausible,” Salathé says. “The United States is already dictating who may buy AI chips and who may not. But because of the many practical constraints involved, I do not think events will unfold that quickly.”
Owen agrees. “I think it is relatively unlikely that AI research and development will be fully automated by 2027, but we cannot rule it out entirely.” He puts the probability of an “AI 2027” scenario at around 10 percent. “That is reason enough to take it very seriously.”
Anton may still have a little time to prepare for the future.