"Next, we eliminate some incorrect worlds by testing them against noisy data. This data should come from intermediate steps and can’t be quantified. If the data conflicts with an imagined future, that future is unlikely to occur. Noisy information may be bad for feedback loops, but it’s great for elimination, because the wider range of outcomes is more effective for pruning. It may even be a bit faster than accurate data, since the variance of accurate data is lower!
As we near decision time, we stop the world construction and elimination, and use judgment and experience to pick the best idea out of the remaining ones. Process improvement happens by creating many parallel processes and discarding the ones that are unfit, not by iteratively improving a single process."
I can't quite understand what he is saying.