Rattle wrote:
] ] A: In the latter scenario, where first you learn, and then
] ] you help the visitor out, you have two places where the
] ] computer has to make intelligent judgments. I am not
] ] saying that is not an interesting or promising approach,
] ] but it does put more strain on the computer. When you
] ] tell it what your interests are, then the computer only
] ] has to be intelligent to use that information to try to
] ] help you out. They are both part of the same goal of
] ] trying to help people out with personal information--it
] ] is just a matter of how you get there. We will be seeing
] ] more of this in the future.

Bah... Strain on the computer is not the issue. The balance is on one side you have technology that is dumb. You have to take time to configure it (and you have to learn how to configure it). It always does exactly what you configure it to do. If it does something wrong you can change the configuration. It always operates "correctly."

On the other side you have technology that learns. You don't have to configure it, but it does have to learn. The biggest problem with these tools is that you don't get to take advantage of it right away, and when you first try it its at its dumbest, and is unlikely to impress you. Few people have the patience to get learning tools up to speed. (Amazon's learning tools work well because they aren't the primary focus of your activity. They are a side effect, so they are useful when they get smart, but when they are dumb they aren't in the way.) Also, you often can't configure learning tools. If it learns something thats wrong its often hard to quickly change its behavior. It has to unlearn.

Memestreams suffers from all of these problems. Our system cannot be controlled by the user. It learns quickly, but it is terribly slow at unlearning. We need to address that later problem. But, the way you make this really work is by finding a middle point between these two approaches where you can get the advantages of technology that learns without being hampered by its weaknesses. Tivo is a good example, but I think its too heavy on the configured side.

Circles are under the control of the user. Totally configurable. With circles the agent is really like an RSS aggregator. With the Top Users the agent is unconfigurable. It is a learning system. We need to bridge the gap between these two ways of operation in the design. Circles work off the bat to get you the content you want to read, and as the agent learns to starts bringing in it's knowledge. You can control how much influence the agent has over what you see. You also need to be able to feed back directly into the agent's decisions... Telling it you don't like things that it has selected...

RE: Interview with Google's Craig Silverstein | CNET News.com