For eleven weeks, the independent group METR looked closely at how Anthropic develops its models. Their finding: on well-specified tasks, progress is roughly twice as fast as before.
Much of that acceleration may come from AI itself helping with research — an uncertain estimate, around 80%. It is not yet a machine inventing the next generation on its own, but assistance already woven into daily R&D.
The researchers saw no clear bottleneck for the year ahead. They remain cautious: the numbers mainly cover well-framed tasks, and other factors also played a role.
Machines are moving faster. It falls to us to stay present, and to choose what those gains are for.