The people closest to artificial intelligence are putting dates on it. In September 2024, Sam Altman wrote that superintelligence could arrive "in a few thousand days." A month later, Dario Amodei wrote that powerful AI "could come as early as 2026," and added in the same sentence that it could also take much longer. The four largest cloud companies plan to spend roughly $725 billion on capital projects in 2026, up from about $410 billion the year before.
Underneath the dates is a claim about the shape of a curve: that it keeps going up. History suggests a different default. Most technology curves bend.
This article is not an argument that AI is overrated. The systems are useful today, and we build software on them, so we have a stake in the answer. It is an attempt to lay out a question that gets less attention than it deserves: can the current approach reach artificial general intelligence (AGI) without another breakthrough, and can anyone put a breakthrough on a schedule?
I do not know. Neither does anyone quoted here. What follows is the evidence on both sides.
A new technology usually starts slowly, improves very fast for a while, then flattens. Further progress comes from a different idea that starts a new curve. From the inside, the steep middle of an S-curve looks identical to an exponential that will never stop.
Computer chips. Processor clock speeds rose more than a hundredfold between 1990 and 2004, from about 33 MHz to 3.8 GHz. Then Intel cancelled its planned 4 GHz chip. More than twenty years later, the fastest desktop processors peak at around 6 GHz. The limit was physics: smaller, faster transistors stopped getting proportionally cooler. Computing kept improving, but through a different idea, many cores and graphics processors working in parallel.
Air travel. Fifty-five years separate the Wright brothers' first flight from the Boeing 707, which carried passengers at about 600 miles per hour. Anyone extending that line in 1958 would have predicted routine supersonic travel by now. Concorde did fly at twice the speed of sound from 1976 to 2003, and then it was retired. A Boeing 787 today cruises at about the same speed as the 707. This limit was economics, not physics: fuel, noise, and ticket prices.
Self-driving cars. In April 2019, Elon Musk said Tesla would have a million robotaxis on the road the following year. Driverless ride services do exist now in several US cities. They arrived years later, from a different company, in a narrower form. The first 95% of the driving problem was solved quickly. The last few percent of reliability took most of the time.
Three different ceilings: physics, economics, and reliability. In each case the ceiling was not the end of progress. It was the point where the cheap gains from one idea ran out.
In 1960, Herbert Simon, who later won a Nobel Prize, wrote that "machines will be capable, within twenty years, of doing any work that a man can do." In 1967, Marvin Minsky of MIT wrote that "within a generation" the problems of creating artificial intelligence would be "substantially solved."
These were the leading researchers of their day, with real results behind them. What followed was a collapse in research funding in the 1970s, a commercial boom in "expert systems" in the 1980s, and a second collapse when those systems proved too brittle.
The underlying ideas were not wrong. Neural networks, the basis of today's AI, date to the 1950s. They waited decades for enough data and enough computing power. The approach first worked at scale in 2012, and the design behind today's language models was published in 2017.
So the breakthroughs did come. Nobody dated them in advance, and the gap between Minsky's prediction and a system that resembled it was more than fifty years.
The training data is finite. Today's models learned from most of the text humans have published online. In December 2024, Ilya Sutskever, a co-founder of OpenAI, told a research conference that this method "as we know it will unquestionably end" because "we have but one internet." The research group Epoch AI estimates the usable stock of public text at around 300 trillion tokens, and projects that training sets will reach that size between 2026 and 2032.
Each step costs a multiple, not an increment. Published scaling research finds that a model's error falls as a power law in computing power and data. In plain terms, each equal improvement requires multiplying the inputs. That works while the money multiplies too. The International Energy Agency estimates that data centers used about 415 terawatt-hours of electricity in 2024, roughly 1.5% of the world's supply, and projects 945 terawatt-hours by 2030. A curve that needs its inputs to keep multiplying can run out of power lines and capital before it runs out of ideas. That is how Concorde ended.
Autonomy depends on the last decimal places. A system that is right 95% of the time and one that is right 99.9% of the time are different products. If a job takes twenty steps and each step succeeds 95% of the time, the whole job succeeds about 36% of the time. At 99% per step, it succeeds 82% of the time. Unsupervised work lives in those decimals, which is the lesson of the self-driving car.
What is measured in the lab is not always what happens at work. In 2025 the research group METR ran a controlled trial with 16 experienced software developers. With AI tools, they took 19% longer to finish their tasks, while believing they had been about 20% faster. The study was small, used early-2025 tools, and its authors warn against generalizing from it. The narrow point stands: perceived gains and measured gains can differ.
Adoption also has its own clock. The economist Paul David showed that electricity took about forty years to raise factory productivity, because factories had to be redesigned around it first.
A fair reading has to start with this: the best public measurement of the curve has not bent.
METR tracks the length of the tasks AI systems can complete. From 2019 to 2025 that length doubled about every seven months. Its January 2026 update found the pace since 2023 was faster, doubling roughly every 131 days. If a ceiling exists, it does not show in that data yet.
There are also answers to each limit above:
- Data. Newer systems improve by reasoning for longer at the moment of use, by training on generated data, and by learning from the results of their own actions. Sutskever's own conclusion was that the method will change, not that progress will stop.
- Cost. The price of running a model at a fixed level of quality has fallen steeply each year, and $725 billion of annual investment is a force of its own. The earlier AI collapses were partly funding droughts.
- Self-improvement. No aircraft ever helped design its successor. The leading labs say their systems now write a large share of the code used to build the next ones. If that loop strengthens, comparisons to earlier technologies stop holding.
There is one caution. Every one of these arguments is also the kind of argument that sounds strongest in the steep part of an S-curve.
There is no agreed definition, and the dates depend on it.
In 2023 the research group AI Impacts surveyed 2,778 AI researchers. Asked when AI would outperform humans at every task, their median answer for a 50% chance was 2047. A second group, asked when every occupation could be fully automated, answered 2116. The same field, in the same year, gave answers 69 years apart depending on how the question was worded.
The 2047 figure was also 13 years earlier than the answer to the identical question one year before. Expert forecasts in this field move a long way in twelve months.
A prediction that cannot be checked cannot be wrong. Before arguing about a date, it is worth asking what observable event would count as arrival.
These are the things worth watching, whichever way you lean:
- Task length outside software. Does the doubling trend hold as tasks stretch to days and weeks, and in work that is not programming?
- Cost per step. Does each new generation need a larger multiple of power and money than the last?
- Error rates on long, unsupervised work. Do they fall as fast as headline capability rises?
- Measured productivity. Do gains show up in controlled studies and national statistics, not only in surveys of how people feel?
- A change of method. If the leading labs move to a new design, that is evidence the old curve was flattening. It may also be the next breakthrough.
If you run a treatment facility, you do not have to pick a side to make a good decision. Both outcomes lead to the same rule: judge a tool on what it does today with a clinician reviewing the output, not on what a roadmap says it will do alone in three years.
If the curve keeps climbing, you lose nothing by having bought what worked. If it flattens, you did not bet your program on the part that never arrived.
Every major technology so far has hit a ceiling, and most found a successor. The open questions are how high this ceiling sits and whether the next breakthrough will keep to anyone's calendar. The people saying three years may be right. Simon and Minsky were among the best in their field too.
- Sam Altman, "The Intelligence Age" (September 2024), as reported by TechRadar
- Dario Amodei, "Machines of Loving Grace" (October 2024)
- Yahoo Finance, hyperscaler capital spending for 2026
- Quote Investigator on Herbert Simon and Marvin Minsky
- Maginative, Ilya Sutskever at NeurIPS 2024
- Epoch AI, "Will we run out of data?"
- Kaplan et al., "Scaling Laws for Neural Language Models" (2020)
- International Energy Agency, "Energy and AI" (2025)
- METR, developer productivity trial (2025) and Time Horizon 1.1 (2026)
- Paul David, "The Dynamo and the Computer" (1990)
- AI Impacts, 2023 Expert Survey on Progress in AI
β
- #artificial intelligence
- #agi
- #ai timelines
- #technology history
- #ai adoption