Every three months, something incredible happens: no, it’s not NVIDIA’s latest earnings report. No, it’s not the Fed cutting rates again. No, it’s not Taylor Swift announcing another tour.
But it’s the fact that an AI doubles the time it can work on its own — without getting lost and without needing us, humans. Which means that, if you think AI is already amazing at what it does now…then wait for next quarter!
See, In March 2025, the best model in the world could handle one hour on a single task. Today, in April 2026, Claude Opus 4.6 works twelve hours straight on a single task. And here’s the part I really need you to feel: in October of this year, that number will probably be twenty-four hours. By January next year, forty-eight. and so on.
Unless you have been living in a cave, you’ve likely seen the AI chart that is the talk in Silicon Valley and that’s moving the global stock market right now. Not the S&P chart, but s a chart made by a 30-person nonprofit in Berkeley, California, called METR — Model Evaluation and Threat Research. This is the chart that’s deciding your professional future, without asking your permission.
Today, we’re going to understand what it’s telling you about the evolution of AI, so you don’t get caught off guard when the next doubling hits.
WHAT THIS CHART ACTUALLY MEASURES
For years, AI intelligence was measured as human intelligence: in tests.
A little bit like the SATs. You’d run a model through a math test, a law test, a reading test, and see how it performed. Companies would run their models through batteries of standardized exams, assessing how they stacked up against rival models at solving math problems, answering legal questions or summarizing text accurately.
These were useful measurements. But they didn’t work well when it came to A.I. agents — systems designed to work autonomously for minutes or hours at a time, because what you really wanted to know, if you were interested in these systems, was how long they could work before getting stuck. Could they handle a simple task that would take a human a few minutes, or a more complex task that would take someone a few hours?
So Beth Barnes, co-founder and CEO of METR, by asking: “How long can an AI work without getting lost?” came up with METR. METR’s researchers attempted to track this by creating a benchmark of software engineering tasks — like debugging code, setting up servers and training small A.I. models. They hired expert software developers to do the tasks, and then they had A.I. agents attempt the same tasks. When an agent succeeded at a task, they logged the time it had taken the human expert to do the same work. They plotted the results on a single chart — task length on one axis, time on the other — and produced a trend line across years of A.I. progress.
What they found was surprising. At first, the length, in human-hours, of a task an A.I. agent was able to complete reliably was doubling roughly every seven months. More recently, with models like Anthropic’s Claude Opus 4.5 and OpenAI’s GPT-5.2, the line took a sharp upward turn — the task length is now doubling every three to four months.
But numbers in the abstract don’t land. So look at the chart for a second and follow the labels with me. In 2019, the best AI in the world could answer a single question — four seconds of human attention. In 2022, GPT-3.5 could count the words in a passage — about thirty-six seconds of clerical work. In 2023, GPT-4 crossed into finding a fact on the web — six minutes, roughly the work of a curious intern. By late 2024, the o1-preview model was training a classifier — thirty minutes, the work of a junior analyst. In 2025, GPT-5 was training adversarially robust image models — four hours, the afternoon of a competent machine learning engineer. And today, Claude Opus 4.6 is implementing complex protocols from multiple technical specifications — ten hours, what a senior engineer does in a working day.
Notice what just happened. The chart isn’t only showing more time. It’s showing a category change. We went from “answer a question” to “implement a system from specifications.” From reactive to architectural. From intern work to senior engineer work — in seven years. That’s not just an acceleration of speed.
That’s an acceleration of what kind of cognitive work counts as automatable. And if you think your own work is safe because it’s more strategic, more complex, more human — I’d ask you to sit with this curve a little longer.
Now, this is the kind of curve where my training as an economist starts screaming at me. Because there’s a principle in technology forecasting called Amara’s Law, named after Roy Amara, the Stanford-trained futurist. Amara’s Law says: we tend to overestimate the impact of a technology in the short run, and dramatically underestimate it in the long run.
Most people right now are still in the first half of Amara’s Law on AI. The over-hype phase. They saw ChatGPT in 2023, watched it write a mediocre poem, and decided this was just another tech cycle. What they’re missing is the second half. The part where the underestimation kicks in. The part where, three doublings later, the world has quietly reorganized itself around a capability they were dismissing eighteen months ago.
This chart isn’t science fiction. It’s Amara’s Law, plotted in real time.
YOUR BRAIN WAS NOT BUILT TO READ THIS CHART
But hold on. Before you go fire your whole team like Jack Doresy and Mark Zuckerberg have been doing recently — or, more likely, spiral into existential dread about your own job — let’s pause for a second.
Because there’s something that needs to be said. And it’s uncomfortable.
Your brain was literally not built to read this chart.
Daniel Kahneman, the Nobel laureate who basically invented behavioral economics, spent his career mapping a single insight: humans operate on two cognitive systems. System 1 is fast, instinctive, intuitive. System 2 is slow, deliberate, analytical. Almost all of our daily decisions — including how we read information — happen in System 1.
And System 1 has one weakness above all others. It cannot process exponential change.
We, as a species, evolved on the savanna. We hunted antelope, counted fruit, measured distance with our eyes. All linear. One more antelope, one less fruit. When System 1 looks at an exponential graph, it instinctively flattens it into a steeply rising straight line — because that’s the only shape it knows.
Think about it: If you told someone in 1995 that within fifteen years, almost every adult on Earth would carry a device with the entire knowledge of humanity in their pocket, free, instantly accessible — they’d have called you delusional. Yet that’s exactly what happened. Not because it was unpredictable. But because System 1 cannot project compound curves.
You can’t. But it’s fine. I don’t either. Nobody does.
So, what to do about it?
THE UNCOMFORTABLE EXERCISE
Do this with me. Pause for a moment — please just don’t, if you are driving do it later. But do it.
Open your calendar from last week. List five tasks that took more than an hour each. A planning meeting, a report, a market analysis, a deck, a long email response, a competitor research piece. Five real tasks.
Now answer me — not for me, for yourself:
Of those five, how many could Claude Opus 4.6, working twelve hours straight, do today? Not 100%. But 70%, 80% of the work.
Probably more than you’d like to admit. Right?
Now the painful question. In six months — when that same model, or its successor, is working twenty-four to forty-eight hours autonomously — how many of those tasks still make sense for you to do yourself?
That’s the question that will define the next decade of your career.
And it’s exactly what I wrote about in “Between You and AI.” The second pillar of my framework — Behavioral Transformation — opens with a skill I call Augmentation. The principle is simple to say and brutal to practice: automate the routine, elevate the human.
But to do that, you need to know what counts as “routine” today. And what was human yesterday and is becoming routine tomorrow.
So what do we do with this?
The answer — and this is the main message of this episode — is not to panic, but to install a discipline.
I call it the quarterly delegation audit. It’s simple. Three months is roughly the speed of the next doubling. So every three months, you sit down with yourself — thirty minutes, a coffee, a notebook — and you answer three questions. Three.
One: what was I doing three months ago that AI now does better than me?
Two: what do I still do today that will probably become an AI task three months from now?
Three: given that, where should I be investing my human hours right now to be on the right side of this curve?
Thirty minutes. Every three months. It’s the cheapest investment you can make in your own career.
And listen, I’ll be direct: if you don’t do this, someone else will do it for you. It’ll be your CEO, your CFO, your board, the market. And a decision made about you, by someone else, is rarely a good one for you.
Sp Here’s the question I’ll leave you with:
How much of your calendar next week is already below the Claude Opus 4.6 line?
Sit with that. And when the next model drops — three, four months from now, with twenty-four autonomous hours — sit with it again.
Want to continue exploring the impact of artificial intelligence, leadership, and the future of work? Discover my podcast Between You and AI | Leadership, Human Skills & the Future of Work.
Check out my book “Between You and AI,” published by Wiley — that’s where I unpack the whole framework of nine human skills for navigating this wave. And you can find me at andreaiorio.com for everything else: keynotes, content, conversations.
