3 ms·
i'm new here and have no idea how long comments are allowed to be. there's a lot to say, but i'll try to be succinct. (and probably fail.) 1) the quickest way
by eikokuma 10y ago
i'm new here and have no idea how long comments are allowed to be. there's a lot to say, but i'll try to be succinct. (and probably fail.)
1) the quickest way to make bad AI is to train it on bad data. and the internet is the king of bad data. if the New York Times is putting out bad recipes and if other humans are regularly buying into fake news articles, blaming the google algorithm seems like shooting the messenger for echoing the low standards of society.
2) many (though not all) of these questions are injecting presuppositions. when you ask google "how long does it take to caramelize onions," you're implying that such a thing can be quantified. you'll get answers which try to tell you how long it takes because semantically that's what fits. but in reality, a lot of things are done when they look done. or done when they smell done. or done when they're thick enough to form peaks.
so i really think this is a human problem. as much as we want to error-proof all our software, in the world of general AI, we will need people to interact responsibly.
- jclos 10y ago1) It really depends on your AI and its goal. If the goal of the AI was to summarize human opinions, then the current approach is fine. What isn't working is that it's doing one thing (here is the answer of some people on the internet) and presenting it as something else (here is the answer). Plus, there are ways to put some bounds of confidence on an algorithmic prediction, and I don't see why they are not displayed along with the answer that was generated. 2) It relates to point 1, but the real solution would be that Google admits "I don't know/it depends" when the confidence is under some threshold.
- jclos 10y ago1) It really depends on your AI and its goal. If the goal of the AI was to summarize human opinions, then the current approach is fine. What isn't working is that it's doing one thing (here is the answer of some people on the internet) and presenting it as something else (here is the answer). Plus, there are ways to put some bounds of confidence on an algorithmic prediction, and I don't see why they are not displayed along with the answer that was generated. 2) It relates to point 1, but the real solution would be that Google admits "I don't know/it depends" when the confidence is under some threshold.
- wutbrodo 10y ago> i'm new here and have no idea how long comments are allowed to be. there's a lot to say, but i'll try to be succinct. (and probably fail.) FWIW you're fine, this wasn't a particularly long comment at all.