HAPPY AI NEW YEAR

The Bubble Doesn't Care If You're Tired

Back in March, METR published a paper that has been lodged in my head ever since. The finding was simple and unsettling: the length of tasks AI can reliably complete is doubling roughly every seven months.

Seven months. Not seven years.

Here is the arithmetic that will not let go of me. Moore’s Law gave us about forty doublings over sixty years. Transistor density advanced every eighteen months or so. That slow compounding rebuilt civilization, but it also gave us something precious along the way: time. Time to adapt. Time to build institutions. Time to argue about what it all meant.

If AI capability doubles every seven months instead of eighteen, you reach forty doublings in roughly twenty-three years, not sixty. A transformation that once unfolded over a lifetime compresses into less than a generation. And this time, nothing is waiting for us to catch up.

ChatGPT launched on November 30, 2022. If you start counting from there, we crossed the seventh doubling about four weeks ago. We are now somewhere in the eighth.

I do not experience that acceleration as an abstract curve. I feel it in the daily work of keeping up. A workflow that took an hour of prompting in the spring took ten minutes last week. That is the optimistic version. The darker one is watching friends quietly fall off the curve while pretending they have not.

“Staying on the bubble” sounds like a toast. In practice, it is a treadmill. The speed keeps increasing. The bubble does not care if you are tired.

So here is to 2026. It will soon separate observers from builders.

https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/