The Future of Artificial Intelligence II.
Part Two — Full automation, explosive growth, superintelligence
The question is not whether artificial intelligence will reshape the labour market, but rather how profound—and how rapid—that transformation will be. Every new technology eliminates some occupations and creates new ones, and the newly created roles are almost always much better than those that disappear. However, it is also true that the economic integration of new technologies is happening in increasingly shorter timeframes, and this time the transformation will likely be faster and more widespread than ever before.
Diverse approaches, similar conclusions
The first significant, comprehensive study on the labour market effects of machine intelligence was conducted by researchers at the University of Oxford more than a decade ago, when large language models were not yet in existence.1 Since the emergence of ChatGPT, analyses have projected rather different visions of the future, depending on the temperament and background of their authors.
Goldman Sachs’ March 2023 report on the economic growth impacts of generative artificial intelligence identified 13 out of 39 general workplace activities listed in the O*NET employment database (such as information gathering, communication, coordination) as being exposed to automation by machine intelligence, and assumed that AI is capable of performing tasks rated at level four on O*NET’s seven-point difficulty scale.2 If so, this would mean the automation of roughly a quarter of current workplace activities—allowing employees to spend more time on tasks that machines cannot perform. However, in 7% of jobs, artificial intelligence could fully replace human labour.3 Altogether, over a decade or more, this could result in a 16% increase in productivity—a much greater change than, for example, the spread of personal computers brought about. The investment bank’s analysts also examined what would happen if machines performed at level six instead of level four: in that case, the share of replaced work time would increase only slightly—from 25% to 27%—but the productivity gain would be significantly greater, reaching 33%. (Level four tasks include things like filling out tax returns for a small business or creating a work schedule; level six involves tasks such as organizing the programme for a large conference or gathering data for a complex scientific report.)
McKinsey also based its analysis on the O*NET database but took a different approach: it examined approximately 2,100 workplace activities to determine which of 18 different skills (such as perception, navigation, coordination, reasoning, creativity) are required to what extent for each activity.4 By comparing this with the timeline of when machines are expected to reach the necessary level in each skill, it becomes possible to estimate when machines will be able to replace humans in specific tasks. However, technical feasibility is only the first step: automation also requires integration into existing work environments and processes, it must be economically viable, and it needs to be accepted by users.5 The transition will be fastest in countries that are best prepared and where wages are high—since replacing human labour is most cost-effective there. According to the consulting firm, in the most advanced economies, artificial intelligence could replace up to half of the time currently spent on work activities as early as the end of this decade—and around 85 percent by the mid-2030s—contributing an annual 3.5 to 4 percentage point increase in productivity.6
The GATE model—developed by the research institute Epoch AI—adopts an entirely different approach and can be summarized as follows.7 Investment in the development of artificial intelligence enables an increase in hardware quantity, as well as improvements in both hardware and software efficiency. This leads to a growth in computational capacity, which is used partly to train new models and partly to operate existing ones: the new models enable the automation of an increasing number of tasks, while the existing models perform economically valuable work. This work generates revenue, which partly drives consumption and partly feeds back into AI development, further reinforcing the cycle.
GATE takes into account nearly forty parameters, which can be adjusted via an interactive interface.8 With the default settings, the model predicts automation of 69% of activities by 2030 and 100% by 2034. With “conservative” settings (which involve changing the values of only three parameters), the projected level of automation is 51% in 2030, 81% in 2035, and 100% by 2040.
It is noteworthy how similarly McKinsey (in blue-grey) and Epoch (in green), despite working with completely different methodologies, see the development of automation:
However, these are necessarily simplifying models that do not take into account a range of constraints: never-ending training runs, exponentially increasing electricity demand, potential data scarcity—not to mention raw material supply or regulatory hurdles. It is worth thinking of this as the pace that the most advanced large economies, such as the United States, can at most achieve over the next five years (and which city-states like Dubai, Hong Kong, or Singapore may even surpass), but it is certain that automation cannot continue unabated beyond that. Two factors will slow it down: first, the physical limits of increasing computing capacity (energy and raw materials) will gradually become more and more decisive; second, in order to carry out physical tasks, robots will need to be manufactured.9
The trillion-dollar business
How will automation develop in the coming years, all things considered? Within a few years, the time will come when machine intelligence must move from being a promise of the future to becoming a business of the present, which means that leading artificial intelligence companies will have to finance their infrastructure investments from their own revenues—amounting to hundreds of billions of dollars annually by the end of the decade. What products or services will generate these enormous revenues?10
The revenues from large language models currently mostly come from consumer subscriptions for personal use: employees pay for them themselves to make their lives easier and their work more productive. These typically cost around twenty dollars per month, which, presuming 50 million subscriptions, would amount to 12 billion dollars in annual revenue.11 Research assistants capable of independent data collection represent a separate category, with a current subscription fee of around 200 dollars. It is not unrealistic to assume that about one-tenth of users will opt for this highest-capability category, which would increase revenues by 12 billion dollars a year.
At first glance, corporate subscriptions differ only in the sales channel, and while they now represent a negligible share compared to individual ones, their significance is steadily growing. The main conceptual difference is that they make the use of artificial intelligence an expectation placed on the employee. Their price depends on the size of the company; assuming an average price of one hundred dollars, 10 million subscriptions would correspond to another 12 billion dollars in annual revenue.12
The real transformation will be brought about by autonomous agents capable of independently performing a significant portion of the tasks associated with certain jobs. While they typically won’t be able to fully replace an entire position, they will handle a large part of the related tasks. This means fewer employees will be needed in those roles, with the remaining workers focusing only on tasks the agents cannot perform. If all goes well, for those who remain, this will effectively amount to a promotion: they will (typically) be relieved of the most unpleasant tasks, and their salaries will increase due to improved productivity. The subscription fees for autonomous agents will be benchmarked against the wages of the employees they can replace, and the costs will vary greatly depending on the type of job. In customer service, this could be around one thousand dollars per month, while in fields such as administration, accounting, finance, and law, it could reach two thousand. Replacing half a million jobs at those rates would total roughly eight billion dollars.13 The other major—and much more profitable—area will be software development and research & development, where subscription fees could reach up to ten thousand dollars per month. With 200,000 subscriptions, this amounts to 24 billion dollars in annual revenue.
Finally, as with all online services, free users will likely have to put up with advertisements sooner or later: based on data from Facebook and Google, advertising revenues could easily reach 20–30 billion dollars within a few years.
Let us see in summary:
It does not seem unrealistic at all to achieve all this by 2027–28.14 We’re talking about replacing seven hundred thousand workers, while the number of employed people in the United States is approximately 160 million—that’s barely four per thousand (though it corresponds to a much higher automation rate, since both individual and corporate subscriptions replace many current work tasks without affecting the number of employees). It’s equally certain that not everyone will succeed: for now, Anthropic, Google, Meta, OpenAI, and xAI each have room to thrive, but once they have to compete for the same limited market, that coexistence will turn into a fight for survival.15
After this, increasingly widespread automation, covering more and more job roles, will start to gain momentum, eventually affecting almost all office work. This doesn’t necessarily mean layoffs, but it does mean that companies that can afford to—those that are well-prepared and not constrained by regulations—will not fill vacant positions, will gradually eliminate entry-level roles, and will not renew contracts with subcontractors. The situation of early-career professionals will change the most, as the availability of traditional office jobs continues to shrink.16 Physical labour will gain value, as replacing these jobs with robots is far more capital-intensive, not to mention the location-dependent nature of such work.
There is considerable variation in the forecasts regarding the extent to which automation will generate economic growth. As we have seen, Goldman Sachs projects a one-time expansion of 16–33% over one to two decades, while McKinsey estimates that productivity could double by 2040. (These translate to an annual growth surplus of 1.5–3.0% and 3.5-4.0%, respectively.) Greater growth than this is difficult to imagine: as automation becomes more widespread within a sector, prices will decrease, limiting the growth of those sectors’ share in total production.17 To put these figures into context: even Goldman Sachs’ more conservative estimate exceeds the productivity growth resulting from digitalization (the spread of computers, the internet, and mobile phones), while the less cautious estimate surpasses the growth experienced during the most dynamic decade in human history, the 1960s. And if McKinsey’s projection proves accurate, it would mark the most transformative shift in leading economies in two hundred years.18
Maintaining control over our future
The societal changes brought about by the emergence of artificial intelligence capable of competing with humans carry fundamental risks. The most immediate challenges include distinguishing machine-generated content from reality, transferring control to autonomous devices, and the adaptation of the labour market to rapid and profound changes. In the longer term, however, the real question is: how much control over our future are we willing to relinquish?19
This is what Stuart Russell calls the WALL·E problem: as we gradually hand over more and more tasks involved in the management of our civilization—for the first time in history, not only the storage of information but also its processing—to machines, we may unknowingly reach a point beyond which this process becomes irreversible, and humanity, voluntarily, gradually and irreversibly relinquishes the ability to shape its own destiny.20
The proposition that artificial intelligence poses an existential threat to humanity takes this argument one (significant) step further.21 The best formulation of this problem to date comes from Swedish philosopher Nick Bostrom.22 The theory of artificial superintelligence rests on the assumption that machine intelligence can engage in recursive self-improvement, enabling it to surpass human cognitive capabilities over time. According to Bostrom, this could grant it six “superpowers”:23
Being more intelligent than any of its opponents, it can outwit them—devising superior strategies that account for every factor and possibility, and thinking countless moves ahead.
Through deception, bribery, or blackmail, it can manipulate people into advancing its objectives, thus mobilizing social resources.
It can seize control of computer networks, hack communication channels and financial systems.
It is capable of developing new technologies—for instance, surveillance systems, advanced weapons, or autonomous robots.
Using these capabilities, it can obtain virtually unlimited financial resources, enabling it to purchase services and materials.
Through continuous recursive self-improvement, it can enhance all of these abilities even further.
It is important to clarify that the concern is not that artificial superintelligence—should it ever emerge—will inevitably turn against humanity, but rather that even a very small chance of such an outcome poses a risk serious enough to warrant attention. According to Bostrom, this threat arises partly from the extreme difficulty of translating human values and objectives into a form that a machine can interpret without the risk of catastrophic error, and partly from the likelihood that a superintelligent system will not only strive to achieve its goals, but also to neutralize efforts aimed at preventing them. In order to be able to execute its objectives, it will seek to ensure its own survival and maximize its capabilities and resources—regardless of the interests of humans.24
In order to give credence to fears concerning the threat posed by superintelligence, we must accept two assumptions: that the creation (or the mere existence) of intelligence surpassing that of humans is at all possible, and that the process leading to it could accelerate suddenly, escaping human control. We know nothing about the first; neither is the second particularly convincing. As we have seen, the performance of large models primarily depends on the amount of computation carried out during their training, which in turn is limited by hardware capabilities. The pace of development of the physical infrastructure is constrained, which—barring magical breakthroughs—rules out sudden leaps. One such magical breakthrough would be a software intelligence explosion.
The advancement of algorithms enables increasingly efficient use of resources: the computational power needed for a language model to reach a given performance level halves roughly every six months. This implies that artificial intelligence gets better even if the hardware remains unchanged. A software intelligence explosion is a hypothetical scenario in which this improvement accelerates from its current pace to many times faster. However, the data shows no sign of this; in fact, the rate of change has remained fairly stable since 2017.25
The hypothesis suggests that when artificial intelligence becomes reliably capable of autonomously executing programming tasks that take several hours—which METR’s research estimates will happen around 2027–28—these agents will be deployed in massive numbers to accelerate the development of machine intelligence.26 The more advanced models thus created will in turn result in the emergence of even better agents, which will also join the research process, and so on, repeatedly. It will become possible to fully automate the entire software development process from start to finish, eliminating the need for human intervention. With humans no longer being the bottleneck, the process will accelerate explosively.27
However, it is far from obvious that multiplying the number of researchers would significantly accelerate progress. The correlation is certainly not linear—doubling the number of programmers hardly doubles the pace of progress—there is no empirical data on the precise nature of this relationship. It is possible that the effect would be barely noticeable, for example because the development of software (beyond a certain threshold) depends less on the number of programmers and more on the abilities of the best of them, or because the pace of algorithmic advancement is limited by computational capacity.28
The crucial point, however, is that any computational capacity used by artificial programmers would have to come from the same pool available to their human colleagues. Whether the programmer is human or artificial, running the same experiment requires the same amount of computational power—except that, in the AI’s case, additional capacity is also needed to run the agent itself (even if that extra cost is small). Consequently, to justify allocating capacity to the machine instead of the human, the AI would need to outperform at least some humans—not necessarily all, but certainly the least capable ones. Thus, artificial programmers will not flood the workforce by the millions as new labour, as proponents of a software intelligence explosion suggest, but will merely replace the lowest-performing programmers at any given time—to a far more limited effect. There is no reason to expect automation in software development to follow a different pattern than in other fields; it will unfold in exactly the same way. Its impact will be significant, but far from explosive.29
Today, both the United States and China regard the advancement of machine intelligence and the strengthening of their relative positions within this field as strategic priorities, which guarantees that progress continues at the fastest possible pace. The earlier large-scale automation arrives, the sooner its economic rewards can be realised: if it indeed yields a four percent growth surplus, as McKinsey forecasts, even a few years’ advantage or delay can translate into a substantial divergence in outcomes.
Exciting years are ahead of us.
This is also emphasized in a study by the International Monetary Fund, which distinguishes between jobs exposed to the rise of machine intelligence where the use of AI leads to increased productivity while preserving the role, and those where full substitution results in the elimination of the occupation. In the case of occupations in the first group, competitiveness can only be maintained through the use of artificial intelligence, as any actor that makes broad use of it gains a significant advantage over competitors who are either unable or unwilling to do so. Cazzaniga, Mauro et al. (2024): Gen-AI: Artificial Intelligence and the Future of Work
Chui, Michael et al. (2023): The Economic Potential of Generative AI: The Next Productivity Frontier. The study builds on an earlier one, with the methodology partly available there. Manyika, James et al. (2017): A Future That Works: Automation, Employment, and Productivity
It typically takes three to five years to develop practically useful and profitable applications and processes based on new technology. Therefore, it is not surprising that the economic impact of large language models has so far been relatively modest.
The Economic Potential of Generative AI, Exhibits 9 and 15. Relative to earlier forecasts, generative artificial intelligence has accelerated automation by roughly a decade.
It is reasonable to assume that the automation of most cognitive tasks will precede that of physical ones, as the mass deployment of universal robots requires the development of significant manufacturing capacity. Of the tasks listed in the O*NET database, 34% do not require physical presence. It is important to emphasize that we are talking about tasks, not jobs: only 13% of occupations can have all of their five most important tasks performed remotely. Barnett, Matthew (2025): The Economic Consequences of Automating Remote Work
The starting point of this calculation is the analysis of FutureSearch. FutureSearch (2025): OpenAI’s Path to Sufficient Revenue for an AI Takeoff in 2027
This subscription shows many similarities to streaming services in terms of its (perceived) usefulness, the size of the offering, and the mode of access. Netflix has around 300 million subscribers, one-sixth of that does not seem at all unrealistic.
While it is true that individual and corporate subscriptions are formally substitutable, these figures are not excessive even so.
The automation enabled by artificial intelligence has a global price, meaning that it essentially costs the same everywhere in the world (apart from differences in sales tax rates), regardless of the location of the client. Consequently, automation will begin in countries with the highest wages, and it can be expected that prices will gradually decrease over time so that autonomous agents can become a competitive alternative even in countries with lower wage levels.
As a matter of fact, this fits into the growth trajectory of recent years: OpenAI, the company with by far the highest revenues, has seen its income double every six months in recent years, and it is expected to easily meet its 2025 target of $12.7 billion. Anthropic, currently in second place, is growing at a comparable rate, trailing by roughly three-quarters of a year. Snodin, Ben–Owen, David–Frymire, Luke (2025): The Combined Revenues of Leading AI Companies Grew by Over 9x in 2023–2024; Burnham, Greg (2025): OpenAI is Projecting Unprecedented Revenue Growth. The growing focus on revenue figures indicates that artificial intelligence companies are increasingly seen not just as promises of the future, but as real economic players today. However, this also increases the risks: if the pace of revenue growth falls short of expectations, investors may turn away from the companies in question, or even from the entire sector.
A similar rivalry will unfold in the Chinese market among Alibaba, Baidu, ByteDance, DeepSeek, Tencent, Zhipu AI, and others. How effectively France’s Mistral—the only leading AI company outside the United States and China—can compete remains an open question.
This process has already begun: in the job roles most exposed to the use of artificial intelligence, the demand for entry-level workers has been noticeably declining for years. Rak, Gwendolyn (2025): How Badly Is AI Cutting Early-Career Employment?
A good example of this is the history of computing: even though today’s computers are billions of times more capable than their counterparts from the 1960s, their price is only a fraction of those earlier machines, which is why the share of IT in the GDP does not reflect this advancement in capability. Kurzweil, Ray (2024): The Singularity Is Nearer: When We Merge with AI, Chapter 5. This phenomenon, known as the Baumol effect, remains valid as long as there are sectors that cannot be mechanized (such as housing).
It is important to remember that we are talking about growth surplus, which appears on top of the baseline productivity increase of around one percent. This corresponds to an overall growth rate of 3–5% in the United States, and even more in emerging economies.
These risks were highlighted in a widely discussed open letter initiated by the Future of Life Institute in March 2023, which was signed upon publication by 1,800 prominent researchers and business leaders in the field of artificial intelligence. In retrospect, it is clear how naïve—or, in the case of some signatories, hypocritical—the letter’s call for a six-month pause in training the most advanced models was, and how unnecessary such a measure would have been. Nevertheless, the underlying ethical concerns it raised remain entirely legitimate. Future of Life Institute (2023): Pause Giant AI Experiments
This concern was expressed in a single sentence in another open letter from May 2023, also signed by many: “[m]itigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Center for AI Safety (2023): Statement on AI Risk
Nick Bostrom (2014): Superintelligence: Paths, Dangers, Strategies. In the more than ten years since the book’s publication, the absence of any genuinely original additions to this discourse speaks volumes—on one hand, to Bostrom’s brilliance, and on the other, to the topic’s theoretical nature and the resulting lack of empirical feedback. It is also noteworthy that Superintelligence originated as a single chapter—one not even written by Bostrom himself—in an earlier anthology on global catastrophes. Yudkowsky, Eliezer (2008): Artificial Intelligence as a Positive and Negative Factor in Global Risk. In Bostrom, Nick–Ćirković, Milan M. (eds.) Global Catastrophic Risks. The author, Eliezer Yudkowsky, has since gained prominence as a leading figure of AI doomsters.
Superintelligence, Table 8
Superintelligence, Chapter 7. As Stuart Russell puts it: one of the fundamental life lessons of artificial intelligence is that “you can’t fetch the coffee if you’re dead.” Russell, Stuart (2017): Three Principles for Creating Safer AI, 4:35-6:10
Another relevant benchmark—also from METR—is RE-Bench, which specifically measures the performance of artificial intelligence agents in AI research and development tasks. Wijk, Hjalmar et al. (2024): RE-Bench: Evaluating Frontier AI R&D Capabilities of Language Model Agents Against Human Experts
See, for example, Josephson, Henry (2025): How Fast Can Algorithms Advance Capabilities?
Preparing for the arrival of artificial superintelligence has a considerable following. This community is the source of fascinating scenarios like AI 2027, which envisions the rate of algorithmic advancement increasing fivefold within two years, and then reaching 250 times the current rate within just one additional year. Kokotajlo, Daniel et al. (2025): AI 2027