That update blog is funny. The only data they can get at reports slowdowns, but they struggle to believe it because developers self-report amazing speedups.

You'd get the same sort of results if you were studying the benefits of substance abuse.

"It is difficult to study the downsides of opiates because none of our participants were willing to go a day without opiates. For this reason, opiates must be really good and we're just missing something."

My bad, I messed up by being lazy while switching from decreases in time taken (that they report) to increased in throughput. (Yes, it's not just flipping the sign, but as I said, I was being lazy!) The broad point still holds, their initial findings have been reversed, and they expect selection effects masked an higher speedup.

The language is confusing, but the chart helps: https://metr.org/assets/images/uplift-2026-post/uplift_timel...