OpenSlopware deleted, forked, and revived – by me on El Reg

The METR study referenced (but not linked) in that paragraph has a breakdown of time spent https://arxiv.org/pdf/2507.09089 in figure 19 on page 33 which contradicts the description in the Register article. Open-source developers participating in the study who were randomly assigned to be allowed to use AI to work on an issue in their repo only spent insignificantly longer testing and debugging, but did spend much longer reviewing AI output, prompting AI, waiting for AI and "idle / overhead." In particular, time spent prompting AI alone was longer than the reduction in active coding.

Lessons for personal productivity include avoiding situations where you're blocked waiting for your tools https://xkcd.com/303/ , not using AI for tasks that you can implement in code as quickly as you can describe them in natural language, and on the meta-level using external randomness to decide between different approaches for the same task and measuring how long it takes you, rather than relying on your gut feeling how easy it was.