I could go down to the bottom and say, oh this is a one on a scale of one to five all the way up to five. You could do this with a key short cut as well. To provide feedback, it does two things. It makes the system smarter about what I like, and at the same time it uses this information to inform the next person coming through Better about whether they would like that particular recipe or not. Now recipes are a convenient example, but we actually tried this with a whole bunch of different news groups on different topics. And, what we found was, remarkably good. this kind of personalized recommendation. Did a much better job at predicting whether somebody would like an article, than simply looking at the average of what everybody said. In some groups this one happened to be about humor. That was a small gain. It was a very close correlation with people's tastes between our predication and people's tastes. But even the average was pretty good because as people who read usenet news know, rec humor had very little that was actually funny, so most of the ratings in there were just 1s and most people agreed those weren't funny. As you moved up to recipes, what you find is there was very little overall agreement. Almost no correlation between the average in individuals, but much higher with the personalized predictions. Because recipes are much more about taste.