 An unbiased estimator can still be a bad model. Let's consider this linear regression estimator. The parameter is theta hat. Let's say it's true value is theta star. But on average, if we were to train this model many, many times, we would get the average value of the theta hats to be somewhere here. This distance over here is the bias. And that is represented exactly in this equation. If this distance between these two points is 0, then we have an unbiased estimator. Now, let's say that the first time that we train this model, we get a value that's well over here. The second time that we train the model, we get a value that's over here, and so on. And so if we totally average all the values of theta over here, we might actually end up with a theta that's meeting the same true value of theta star. But each of these individual models will have theta value that really doesn't fit the true value very well. And thus, we also want these points to be close to each other.