A research organisation called METR has been quietly doing something nobody else had bothered to do: measuring not what AI can answer, but what AI can actually do. Their findings have a shape I haven't been able to stop thinking about. AI's ability to complete real-world tasks is doubling roughly every seven months. That's not a prediction. It's a measurement — six years of data, validated across multiple benchmarks. And the trajectory it describes changes the conversation about preparation completely.
The benchmark nobody was running
Traditional AI benchmarks measure whether a model can answer a question correctly. Can it pass a bar exam? Can it solve a maths problem? Can it identify an image? These tests matter for some things, but they tell you almost nothing about whether AI can do sustained work — the kind of work that fills an actual job.
METR's insight was to reframe the measurement entirely. Instead of asking "can it answer this?", they asked: "how long would a human expert take to complete this task — and can the AI do it?" They assembled hundreds of multi-step tasks across software engineering and complex reasoning, recorded expert completion times for each, and then tested frontier AI models against them.
The result is a single metric they call the "time horizon" — the task length at which an AI hits 50% success. And when you plot how that metric has moved over the past six years, a very clear curve emerges.
Six years of data. One clear trend: the task horizon is rising, and it doesn't look like slowing down.
What the numbers actually mean
Today, frontier AI models — tested on Claude 3.7 Sonnet — reliably complete tasks that take a human expert up to about one hour. They succeed on nearly 100% of tasks that take humans under four minutes. They succeed on fewer than 10% of tasks that take over four hours.
That sounds like a limitation. But extrapolate the doubling curve — and METR's sensitivity analysis shows it holds even if the measurements are off by a factor of ten — and the picture shifts dramatically. Within one to four years, AI is projected to be capable of completing tasks that currently take human experts an entire month to finish. Autonomously. Without supervision at every step.
That's not speculative. It's a measurement with a trajectory attached. And most organisations are not building at the speed that trajectory demands.
"In under a decade, we will see AI agents that can independently complete a large fraction of software tasks that currently take humans days or weeks." — METR Research Team
Then 270 of the world's most credible people signed a letter
Around the same time as METR's research was circulating, something unusual happened. A statement appeared at wemustactnow.ai — organised by Stanford's Digital Economy Lab and signed by over 270 economists, researchers, and technology leaders. The signatories include ten Nobel Prize winners: Joseph Stiglitz, Paul Krugman, Ben Bernanke. AI pioneers: Yoshua Bengio. Former tech executives: Eric Schmidt. Researchers from MIT, Harvard, OpenAI, Anthropic, Google DeepMind.
The statement is careful in its language but urgent in its message. It doesn't oppose AI. It doesn't call for a slowdown. What it calls for is this: economists, policymakers, and technology leaders must act now to understand the economics of transformative AI and to build the incentives, guardrails, and institutions needed to steer it in a direction that complements humans and benefits society.
The comparison in the background of that statement is the Industrial Revolution. That transformation reshaped economies, workforces, and communities — and it took generations to play out. AI may compress a similar scale of disruption into a decade. The window to build institutional responses isn't open indefinitely.
The lead time problem
Here's what I keep coming back to. If month-long autonomous AI work is one to four years away, the window to prepare is not a comfortable distance in the future. It's right now. The organisations that will manage this transition well — the ones whose people understand what's changing, whose leaders have a plan, whose teams have been equipped rather than blindsided — are building that capacity today.
The organisations that will struggle are the ones treating AI readiness as a future problem. Because by the time month-long autonomous task completion arrives, it won't feel like an arriving wave. It will feel like a wall that appeared overnight.
The METR research is a gift, actually. It gives us something most transitions don't: a measurable trajectory with enough lead time to act on it. The question is whether organisations use that lead time — or spend it watching the doubling curve climb.
