Facebook's AI Research Labs
fastcompany.com
fastcompany.com
I think feel like it's important to somehow measure the level of progress being made by the current explosion of deep learning processes. I'm personally not that impressed by translation applications or Google search innovations - the translations I see still seem barely functional, noticeably better than purely literal translation but not very much more useful than purely literal translation.
Alphago was definite progress. Are there that many problems that could be approached in a similar way?
Clearly, making ads work is important to a company's bottom line. But it seems like there are going to be hard limits to just extrapolating patterns - I know youtube's recommendation engine has gotten worse for me over time and it seems like the smartest entity in the world can only figure out so much future buying from past online surfing and past purchases combined. And even more, there's only so much ads in particular are going change this.
So they try one method, small improvement, try another, small improvement, try a third and suddenly they have a breakthrough in some field.
Trouble is that right now you don't know what technique will work in what field. That means that it could be in Go, or learning what to eat, or driving cars, or cancer diagnosis etc.
So wherever the next breakthroughs happen to be is probably where one of the "battles for AI" will be waged.
The one exception to this rule is an AI that can create other AI. If that happens, the war for AI is over. Whether it has been lost or won depends on your POV.
But it's there also the question of how much progress is possible in improving the answer to a given question. Some problems might present a situation where "you can't squeeze blood from a stone".
There is also the factor that those situations where humans still do better than computers are situations where we know progress is theoretically possible at some level whereas in situations where computers already outperform human, it is harder to have a theoretical reason to think further progress is possible.
Humans have made immense progress in the last 10K years, surpassing all other animals in capabilities, etc.
And individual geniuses have at times pushed a given field forward beyond what was otherwise expected (For example Kary Mullis or Craig Venter in biology).
That would lead one to imagine entity with a broad intelligence that surpasses humans could do amazing things.
All that said "artificial intelligence" still isn't going to be a super box that you can point at a context-free stream of data and get magic predictions from.
We have the ability to talk of seeing things "from the viewpoint of an X" and possibly even imagine that viewpoint. But no one other animal or entity (that we know of) has the ability to talk about or imagine taking the viewpoint of another species. And such capacities seem to have been instrumental in the rise of humans as apex predators everywhere, etc.
We can't even say for certain that what your geniuses of choice has more of, is the same thing we have more of than animals.
It's easy to imagine supergeniuses, but as soon as you try to describe rigorously what they should be capable of, you're immediately on shakier ground.
It's a bit like imagining future computer programs. We're doing a lot of things with our computers today that people couldn't have imagined 40 years ago, or outright rejected (my father's thesis advisor asserted that computers would get a lot more powerful, but would never become powerful enough to compute with images). But there are also a lot of things they thought would be easy, which turned out hard, or even impossible.
In computer science, we have some ideas of the "hard limits", since the field started with Turing and computational complexity theory has progressed steadily. For human-like intelligence, or "general learning", we know very little (except that every hard limit we know from computer science applies to brains too!).
"We can't even say for certain that what your geniuses of choice has more of, is the same thing we have more of than animals."
Agreed. It could even be that thing that makes these "geniuses" is simply circumstances.
Regardless, the various human beings who accomplish an incredible amount in certain ways show that accomplish such advances is possible for material object. This is in contrast to looking at a apparently semi-random stream of data and expecting to predict N times better than ordinary statistics. It's merely "some reason to think things are possible" versus "no a-priori reason to think it is possible".
Also, with computing, even a lot of complexity theory is hard interpret. For example, a fair portion of NP-complete problems (such as SAT) are actually reasonably easy most of the time, with the theoretical worst exponential case performance being fairly rare.
No, the real battle for the best AI is being fought on the various global stock exchanges. The vast majority of trades are AI now. In this way AI virtually already controls the price of all globally traded goods.
Don't forget meetings. Lots and lots of meetings.
Yes.
You haven't been paying attention, that has been the name of the game since Locke.
"If you want a picture of the future, imagine The Facebook burrowing everywhere -- forever."
To my mind, the former is largely (though not exclusively) based on logical reasoning (as in formal logic) and the latter is largely (though not exclusively) based on statistical reasoning.
I hope one of these articles will take some time to bring us up to date on the recent developments in contrast to the other.