METR says it can rarely find humans who beat top AI at coding tasks.
METR, a nonprofit that tests AI models, published its quarterly capability report on Friday. The length of tasks that models can finish is doubling about every 3.6 months, it found. Models still judge the quality of work worse than human experts do.
Why it mattersMETR says the evidence cannot rule out a drastic speed-up of AI research within a couple of years. Its measurements are among the most cited gauges of AI progress.