I do not use AI for where it is. I use it for where it will be in nine months. That sounds grand. It is actually the only rational strategy, once you look at the numbers.
In March 2025, METR published a study called Measuring AI Ability to Complete Long Tasks. They constructed a benchmark of one hundred and seventy tasks, gathered over eight hundred human baselines, and measured how long a task the best available model could reliably complete.
The headline finding: the time horizon of tasks AI can complete doubles roughly every seven months. In the last year, the doubling time has shortened to four months.
The length of tasks models can complete is well predicted by an exponential trend, with a doubling time of around seven months.
Extrapolating the trend, METR projected that within a decade AI agents will independently complete software tasks that currently take a human days or weeks.
A few months is nothing. I have been in enough large organisations to know that procurement alone takes longer than that. The companies I work with are scoping AI platforms that will go live two years from now. If I design those platforms for the model that exists today, they will be obsolete before the first user logs in.
So I spend part of every week deliberately working at the edge of what Claude can do, not the middle. I feed it tasks that are too long. I ask for architectures that are too ambitious. Half the time it fumbles. The other half it shows me something I could not have done myself, which tells me what will be trivial in six months.
My workflow in 2027 will not look like my workflow today. The only way to be ready is to keep meeting the model one step ahead of where it is.
The other risk is subtler. If I keep meeting the model where it was, I will keep shipping 2024 products while everyone else ships 2027 ones. The gap does not stay the same. It compounds.