A rule of thumb: spend 80% of time in fast execution, and 20% in synthesis.

During the fast twitch time, you're accumulating model error.

Then in the slow twitch time you're abducting new learning out of that error to have a more predictive model.

Now that it's factored into the model, the next phase of fast execution gives you more leverage than you had before.

This is similar to what our brains do during sleep!