The last bit about people ignoring pre-feed-forward neural net techniques, or discounting them as not AI, reminds me of the AI effect. When a field gets so hype-driven it can be hard to separate...
The last bit about people ignoring pre-feed-forward neural net techniques, or discounting them as not AI, reminds me of the AI effect. When a field gets so hype-driven it can be hard to separate fact from dogma, and since even experts don’t have good operating definitions of what general intelligence even is (or how to measure it), it can be easy for people to assume the field of AI is appropriately aligned toward objective progress, and that since recent subjective progress has focused on particular techniques a lay person may assume that the experts know what they are doing, so the techniques become synonymous with the field itself. I think this is ultimately problematic, but I don’t think I have any particular insight to help with the problem. I tend, in my idiolect, to prefer the term machine learning to refer to what many call AI, but since that is possibly also a loaded term, it becomes necessary to unpack that as well. After all, I don’t think anyone will pick up and use “the art of applying universal function approximators to domain-specific problems”, even though that description is likely much more apt than “AI” (at least for current practitioners circa 2020).
I should also note that I don’t even necessarily think that general artificial intelligence is necessarily the only goal worth pursuing. I think humans continue and will continue to benefit from...
I should also note that I don’t even necessarily think that general artificial intelligence is necessarily the only goal worth pursuing. I think humans continue and will continue to benefit from task-specific helper software that needn’t be generally intelligent at all. Technologies such as search engines, photo editors, spell-checkers, or other softwares that provide useful functionality are definitely still worthwhile, regardless of the underlying implementation that empowers that functionality. That is, if you can make a good spell checker with deep neural nets, great! If you can make a 10x faster, but 0.99x as accurate spell checker with different technology, I think it’s probably worth looking at that.
The last bit about people ignoring pre-feed-forward neural net techniques, or discounting them as not AI, reminds me of the AI effect. When a field gets so hype-driven it can be hard to separate fact from dogma, and since even experts don’t have good operating definitions of what general intelligence even is (or how to measure it), it can be easy for people to assume the field of AI is appropriately aligned toward objective progress, and that since recent subjective progress has focused on particular techniques a lay person may assume that the experts know what they are doing, so the techniques become synonymous with the field itself. I think this is ultimately problematic, but I don’t think I have any particular insight to help with the problem. I tend, in my idiolect, to prefer the term machine learning to refer to what many call AI, but since that is possibly also a loaded term, it becomes necessary to unpack that as well. After all, I don’t think anyone will pick up and use “the art of applying universal function approximators to domain-specific problems”, even though that description is likely much more apt than “AI” (at least for current practitioners circa 2020).
I should also note that I don’t even necessarily think that general artificial intelligence is necessarily the only goal worth pursuing. I think humans continue and will continue to benefit from task-specific helper software that needn’t be generally intelligent at all. Technologies such as search engines, photo editors, spell-checkers, or other softwares that provide useful functionality are definitely still worthwhile, regardless of the underlying implementation that empowers that functionality. That is, if you can make a good spell checker with deep neural nets, great! If you can make a 10x faster, but 0.99x as accurate spell checker with different technology, I think it’s probably worth looking at that.