BOINC is optimized for tasks that are “embarassingly parallel.” This seems to work well for things like protein folding and searching for aliens. It has also been applied to render farms. The real question is whether the same technology can be expanded to training ML models. Can ML models be made embarrassingly parallel?