The "future of search" is very intimately related to the "future of AI", and quite frankly, Quora has nothing to do with either, except that maybe it will be a

The "future of search" is very intimately related to the "future of AI", and quite frankly, Quora has nothing to do with either, except that maybe it will be a decent source of answers to specific questions about startups... As for the future of search and AI, the way I see things, there are eventually going to be two main AI camps that emerge (NB, I'm not saying that this is the state of the AI field right now, but I think it will segregate more along these lines in the next few years): 1) Big data AI: this is the Google style of AI, which usually assumes that the best way to do AI tasks is to use fairly naive algorithms and toss truckloads of data at them. 2) Deep model AI: this group is more about models, and believes that the best approach is to extract as much structure as possible from a limited data set. More data helps, to be sure, but the emphasis is more on getting as much as possible out of the data rather than getting more data. The way things are going now, Google is going to win at big-data AI, no ifs, ands, or buts - they've got more data than just about anyone. So the way I see it, the only way someone is actually going to unseat Google is if it turns out that deep-model is a workable approach (thus far it hasn't had many serious successes), and Google doesn't focus on it early enough (because if they do, then their massive data availability will make sure they win anyways - that said, the folks like Peter Norvig driving the attention at Google are very explicitly in favor of data-based approaches, so I don't see Google leading the pack on model-centric research). IMO there are some good reasons to think that deep-model approaches are viable: despite Google's massive billion-book data sets, humans are still better at doing, for instance, translation, despite the fact that a good translator might have only received the equivalent of maybe .1% of the input that Google leans on for its translation approach. The question is, will anyone actually figure out how to do it well? That's up in the air, but the fact that evolution figured out how to do it means that it's probably not terribly difficult, we just haven't looked in the right places yet. My personal opinion is that the main pinch point in the typical big-data approach is that it's limited by statistical correctness, whereas human intelligence is not (we happily assume patterns exist in data even when we don't have enough data to make a proper statistical inference, and then we filter out incorrect assumptions later, also by using statistically incorrect heuristics and patterns - in other words, we're usually wrong, but sometimes we get lucky, and as long as we can eventually recognize that we're wrong, we do just fine). I think deep-model hopefuls would be wise to look more seriously at explicitly statistically unsound approaches if they want a shot at beating out big-data...