5 minutes, 1 hour, 5 hours, 14.5 hours.
When we read these numbers, our brains automatically form intuitive images. A cup of coffee, a meeting, half a day, a night. Until now, we have understood AI's ability to work over time only in human units of time. Because that was the only language the two could share.
But this translation will stop working before long. “Three years' worth of human work” will no longer be intuition. It will be rhetoric. After we lose time as a common language, what will we use to measure the abilities of this being?
There is a nonprofit research institute in Berkeley, California, called METR. Model Evaluation and Threat Research. They measure one thing. They give AI tasks that a skilled software engineer might solve, and track how long those tasks can get before the AI can solve them on its own, without human supervision.
The difficulty of a task is converted into ‘the amount of time it takes a skilled human to complete the same task.’ A one-minute bug fix, a one-hour feature implementation, an eight-hour system design. The amount of time a human would need for a task that AI completes successfully half the time. METR calls this the time horizon. METR's chart is not a record of AI's growth. It is the expiration date of the human unit of measurement.
In February 2026, AI completed on its own a task that would take a human expert 14.5 hours. In the meantime, the client went to sleep and woke up, ate breakfast, and went to work.
Throughout human history, every tool worked actively only while the human holding it was awake. The first being that may have judged and worked on its own in place of a sleeping master was probably the butler. A butler remembers the master's preferences, runs the household without the master, and, over time, knows the house better than the master does.
There is one reason this being is unsettling. At some point, the master forgets how to run the house without the butler. The mechanical being we created has quietly entered that category.
This butler's autonomous working time is doubling every four months. A doubling period of 123 days, R² 0.93. R² is a metric showing how well a trend line explains the actual data. At 0.93, it means the points are landing on a straight line with almost no exceptions.
Water is still water at 90 degrees and at 99 degrees. At 100 degrees, its state itself changes. Liquid becomes gas, and something that had weight fills space. The boiling point does not arrive suddenly. Until 99 degrees, it looks as though nothing has happened. Then, in an instant, everything changes.
If this curve continues, AI will independently complete a 29-hour project in June this year, a four-day project in October, and a two-week project early next year. On an exponential curve, the most dangerous stretch is not the moment it explodes. It is 99 degrees, when everything still looks ordinary.
The headlines talk about 14.5 hours. But the upper bound of the measurement's confidence interval is already 98 hours. Four days. A confidence interval is the range in which the actual value may fall when the same experiment is repeated. An upper bound of 98 hours means that AI can, with some probability, autonomously complete a task of a complexity that would take a human four days.
The tail may matter more than the average. Historically, technology changed the world not because average performance improved, but because things that were only occasionally possible became possible at last. The steam engine and the internet were both, at first, things that only worked sometimes.
Now that 98 hours is sometimes possible, we call that ‘sometimes’ an exception. History has repeatedly shown that the people who called ‘sometimes’ an exception were wrong. From 5 minutes to 14.5 hours in 18 months. 174 times. At this point, the number has not simply gotten bigger. Its nature has changed.
A senior developer can focus at peak intensity for only about 4–6 hours a day. AI has already gone far beyond that. It is not simply a fast tool, but something that thinks for itself and works the night shift. Something that sits down where you left work and works alone until you arrive the next morning. Something that does not get tired, does not ask for a salary, and remembers yesterday's decisions.
In the United States, 14.5 hours of a senior developer's labor costs at least $1,000. Running the same work through an API costs around $50. A gap of more than 20 times. At this scale, change does not arrive gradually. Kodak did not fail because demand for photographs declined. The cost structure of film itself had lost its meaning.
METR's tasks are in domains with definite answers. Software engineering, ML experiments, cybersecurity. You can run the code and immediately know whether it is right or wrong. But much of labor happens where there are no right answers. When to speak up in a meeting. What the real desire is behind a customer's complaint. Whether a team member is exhausted. The ability to read persuasive silence, the right compromise, and expectations that have not yet been spoken. These things cannot be placed on a mathematical curve.
These abilities that do not appear on graphs have made humans irreplaceable until now. The steeper the METR chart becomes, the easier it is to forget the world outside the chart. Just as a more precise map makes us ignore the terrain beyond it. But one uncomfortable question remains. Does the fact that something does not appear on a graph really mean it cannot appear on one? Or does it mean no one has tried to plot it yet?
METR warns that its benchmark is beginning to saturate. The ruler has grown too short. When a child's height exceeds the ruler, the problem is not the child but the ruler. The irony is that researchers are now using AI's help to make a longer ruler. It is like asking the person taking the test for advice on how to make the test harder. It is a sign that AI has moved beyond the role of a tool, and also a sign that we have already begun to accept that naturally.
When the imagination of the humans designing benchmarks can no longer keep up with the speed of the AI passing them, what do we use to judge safety? More fundamentally, does the premise that we can judge the safety of something we do not understand still hold?
There is something the public has not yet recognized. Some of the 14.5-hour autonomous tasks are already ML pipeline optimization, synthetic data generation, and automated model evaluation. Put simply, these are the tasks AI needs to make the next generation of AI smarter.
For the first time, there is enough time for AI to complete on its own the work of improving itself. A night has opened in which the butler can train itself without the master's knowledge. Once the feedback loop closes without outside intervention, the next version's rate of improvement may not follow the curve we drew. We may not be the ones who plot the next point on this curve.
What remains for humans pushed out of labor is intention. During the Industrial Revolution, weavers smashed machines. Whenever automation arrived, humans resisted, moved up to a higher level, or invented new work. Farmers became factory workers, factory workers became service workers, and service workers became knowledge workers.
But this time, it is unclear where the higher level is. We have not yet named the level above knowledge work. Or perhaps that level is precisely where naming things becomes unnecessary.
When the sentence “I make this” disappears and only “I want this” remains, can we sustain our identity with that one sentence? When the era in which one's job was one's identity and one's skills were one's dignity passes, no one has yet designed what comes after it.
On the day the chart was published, a SpaceX engineer named Vittorio posted a tweet. A tech insider with 170,000 followers. Someone who looks at rocket reusability data every day and knows in his body how exponential curves explode.
“ehm, guys… I think it’s happening.”
He did not explain. Someone who sees an inflection point every day does not need to be long-winded. He was not warning us about the future. He was quietly acknowledging that the future we had failed to prepare for was already the present.
The clock has been running for 14.5 hours.
The number will have changed even while you were reading this. When you finish this sentence and look up at the view outside, somewhere a cursor is blinking on its own across a screen no one is watching. In a house where the master is asleep, in a study with the lights off, the butler is quietly turning to the next page.
No one set an alarm.
