New research indicates a consistent and rapid pace of progress in AI systems, which are expected to reach general human-level capabilities by 2029 – a revolution that could change the labor market and society as a whole.
More than fifty years ago, Gordon Moore coined “Moore’s Law,” which illustrated and predicted the rate of development of chips. Last week, an attempt was made to produce a new law – called “Meter’s Law,” after Huguenot – that describes the speed of development of artificial intelligence. To summarize briefly, it is Doubles its capabilities Every seven months, and according to calculations, you could reach the level of "general artificial intelligence" (which successfully competes with human experts) by 2029.
And now for the small but important details.
METR is a California-based nonprofit that has been trying in recent months to find a pattern that explains the rate of increase in AI capabilities. Why is this important? Because such a pattern can help us understand what AI capabilities will be like in a year, two, or five years. This is not just curiosity. Businesses can choose to invest billions of dollars in AI today, with the informed calculation that they will get the return in the coming years.
Matar decided to focus on the AI's ability to program on its own. It can do so today, of course, but with varying levels of success, which depend mainly on the difficulty of the task. Matar translated the "difficulty" into another parameter: the amount of time it would take a human programmer to successfully complete the task. They set up a series of 169 tasks in programming, cybersecurity, and machine learning, and examined how long it took humans to complete them.
Then they tested how artificial intelligences coped with the exact same tasks.
The researchers tested thirteen artificial intelligences, each considered a leader in its time. They saw a clear pattern in the pace of development. The simplest models – such as GPT-2 from 2019 – could only successfully complete tasks lasting three seconds at most. The next generation, called GPT-3, was released in mid-2020, and was able to handle tasks lasting eight seconds. By early 2022, there was already an AI that could complete tasks lasting thirty seconds, and so on. The success stories continue in this way until the very beginning of 2025, when Claude 3.7-sonnet is able to complete programming tasks lasting an entire hour, all by itself.
What's the law here? Simple: In each generation, artificial intelligence can handle replication tasks that are twice as long as those performed by the previous generation. Or, to put it more simply, it doubles its capabilities every generation. And how long does each generation last? Just seven months.
This means that if today the most advanced artificial intelligence can perform programming tasks that take an hour, then by the end of 2025, it will already be performing tasks that take two hours. And in three generations – that is, 21 months, or around 2027 – it will do eight hours of work, with the click of a single button.
In other words, it could replace the entire day's work of an average programmer.
But here too it is worth considering the fine print.
First of all, these calculations are only valid for a 50% success rate on a task. In other words, the most advanced artificial intelligence today can perform an hour-long task, but it will only do it successfully half the time. If we assume that this will be the case in the coming years, then the human expert will still be of enormous importance: he will have to supervise the artificial intelligence, while it does a whole day's work for him in a few minutes. He will have to check the products to make sure that it does not make a mistake at any stage. He will use the artificial intelligence as a tool, but he will be ultimately responsible for the results.
And what about a more autonomous AI? Say, one that succeeds eighty percent of the time? The researchers also tested it against this strict criterion, and concluded that the same rule applies there: Every seven months, AI doubles its capabilities. The most advanced AI today is capable of performing programming tasks that take a maximum of 15 minutes, with an eighty percent success rate. How long before it can reliably perform tasks that would take a human programmer an entire day? Five generations in total, or three years. That is, in 2028. Maybe in 2029, if we are less generous.
Even when AI reaches this level, we will still need humans to supervise it. The people who will be working on AI at that time will not necessarily be programmers, but probably more like “product managers.” They will understand what the final product should look like, and then they will activate AI in a variety of processes: for developing the website, designing its access, ensuring quality, establishing cyber protection, and so on. It will perform the tasks with a high level of success, and managers will mainly have to go through it to detect minor errors – and demand that it fix them itself.
And if artificial intelligence continues to advance in capabilities, then a few years later it will be able to perform the same tasks that would require a human developer to work an entire day, with a 100 percent success rate. Period.
It should be noted that "Mater's Law" was only tested on tasks in the field of programming and machine learning, and therefore these predictions are only relevant for tasks in this field. It is not yet clear whether "Mater's Law" is also valid for tasks in other fields. Nevertheless, it is certainly possible that a similar law also exists in the fields of medical, legal, insurance, economic consulting, and so on.
If the "meter law" is indeed valid for many different subjects, it means that in the next three years, the world of work will change completely. The brilliant programmer will be replaced by the multidisciplinary "product manager" who will activate a series of artificial programmers, artificial graphic designers, artificial cybersecurity and a host of other artificial intelligences to bring a new product to a high level. We will also see the same turn in many other areas of white-collar professions, including science, engineering, law and medicine.
Want to be future-ready? Learn to work with artificial intelligence. Learn to manage it. It will give you superpowers today, and they will only grow with time.
This is the point where it is worth shedding light on a point regarding "general artificial intelligence" - that is, one that is comparable in its abilities to humans at the highest level. Matar's researchers believe that such general artificial intelligence could perform an entire month's work on its own. This is about the time it takes to train a new recruit in a company, so when artificial intelligence reaches this level, it will be able to enter any company and bring value to it in a short time. The researchers believe that by the end of 2029 we will reach artificial intelligence at this level, with a margin of error of about two years. That is, 2031 at the latest.
This will be the year, in fact, when artificial intelligence will be able to join an existing organization, quickly learn how to work in it, and shortly thereafter – even manage it itself.
And what will happen then, when artificial intelligences will be able to manage organizations, and even establish their own organizations and manage humans?
So we will finally have to think seriously not only about the capabilities it will have, but also about what we want it to do – and what not.
The choice, hopefully, will remain in our hands.
