3 ms·
Eliezer Yudowsky on Trump, History (2016)
- suff 7y agoEliezer, you know more than most, that a model must not neglect important features, otherwise the model will not be useful. You also surely know that the only benefit of a model is its ability to predict the future. Let's say, hypothetically, 'a model' failed to account for, say, diplomats laundering foreign aid to the children of U.S. politicians, and also failed to predict 2016 with high certainty, you might correctly conclude that said model will fail, with zero expectation of long-term utility. But if, in the short-term, it helps people sleep at night, convincing themselves they are morally superior, then they are also free to be incorrect. At what point though, do you decide to stop and reassess your model?