BEN SEELY在AlphaGO取胜后发布的FB
Wowwwwwww, a computer program beat a top Go (Baduk) player in a game for the first time in history. (See my last post about this match).Since I've won world championships in a game (Othello) in which the AI has been better than the best humans for 20 years, I figured I would talk about this.
What AlphaGo has done is an extremely significant achievement in artificial intelligence. And I bet this suggests a lot of fast progress in AI in many applications, in the near future. I frankly think this is a far more important milestone than Deep Blue vs. Kasparov.
Or, to put it in scary-ese, "the age of man is coming to an end".
Or maybe it's just beginning smile 表情符. AI could end up being extremely empowering, too.
AlphaGo's abilities to make good decisions when faced with a huge number of possible options is extremely promising for applications in so many "human" areas, and hard problems to solve.
AlphaGo also seems to have something very akin to "intuition", and this also seems exciting.
I had actually figured that the human (Lee Sedol) would lose, even though most people expected Sedol to win every game, because the rate of progress AlphaGo had made suggested huge potential in its learning curve.
Also, the Go players were not humble enough. Since Go is a complicated game, they should have known that the theoretical ceiling for perfect-ish play is far higher than what humans have so far achieved. It's not like top humans playing each other get a draw in every game, so obviously there was potential to play far better. The greater this potential, the easier it is to make rapid improvements in the AI.
Also, when a huge institution like Google throws a ton of resources at a problem, they can often achieve remarkable things in a short period of time. Also, I figured Google anticipated they would win, and that dictated their timing in setting the match. It's better PR to win than to lose.
I'm not totally sure this was the result I was rooting for, though. It's always been fun to believe that what we humans can do is somehow special, and this is being steadily disproven over and over and over.
It won't be long before literally nothing humans do is uniquely special. Although we will be able to claim the special distinction of being the first species on Earth to ever create from scratch something smarter than us smile .
Othello programs have been better than humans for 20 years, and the top Chess programs have been better than humans for around 12-15 years, so I feel I can make some predictions from that experience, about what Go players, and all humans, are going to be experiencing in the future as AI gets stronger and stronger.
1. AI gets more aggressive as it gets smarter.
I figured AlphaGo would become more aggressive than it was before, and I was right. Top AI, when it is superior to humans, is typically far more aggressive than the humans are.
The only reason humans ever have to be cautious is if we can't figure out whether or not the risk will be fatal. The nature of greater intelligence, though, is the ability to figure out when taking a risk *will* pay off. Top AI will steadily seem more and more precisely aggressive, doing things we are too scared to even consider, but the AI will figure out that it works out.
2. AI also doesn't have fear. We humans only have one life, so we have a perpetual and extreme bias towards fear and caution.
We humans are far too cautious in 99.99% of situations. But we have to be cautious and not take all those opportunities we could be taking, because on every 10,000th opportunity, it would kill us if we tried to take it.
That is the whole basis for fear, e.g. "Better safe than sorry". And if you really dig deep down into it, almost everything we do and rationalize as being due to other motivations, is really at its root motivated by fear. We just don't like being reminded all the time of just how much we are ruled by fear and the limits of our intelligence, so we come up with other explanations for what we do- our rationalizations.
AI, in the abstract world in which it lives, cannot die. This is actually the biggest underlying reason for why it can learn so quickly; it has no fear, and no ego.
3. We are going to be ritually humiliated, over and over and over again, with one experience after another of learning just how stupid all of us humans are.
(Although at least the benefit to this humiliation will be that AI will be able to do things for us, and teach us things, which we never could have done ourselves. It's worth it, even if our egos have to take a hit!)
This was also why I figured humans were overestimating Lee Sedol's chances of beating Alpha Go. As an Othello player, I'm used to losing to AI. Go players have not had those thousands of experiences in seeing that AI can be smarter than our puny human brains. Top Go players' brains aren't much more talented than top Chess or Othello players (I don't think), it just took longer to figure out how to adapt AI to the game of Go.
Most humans are arrogant about what humans can do. We're so used to being the best, we don't fully realize how frail and stupid we really are.
(As a side note, try to think about everything you know, which you did not learn from other human being, or which wasn't a simple step beyond what you had learned from other human beings. In the absence of all the knowledge other humans possess- which is our real wellspring of genius- we are phenomenally ignorant.
Without the knowledge of other humans, we could not be smarter than the typical feral human raised by wild animals- which is incredibly stupid. I try to remember this every time I'm trying to learn something new, that my natural intelligence isn't any greater than that of a feral human, so I shouldn't expect myself to start out beyond the feral ignorant state.)
Incidentally, I’ve wondered for a while whether part of why a lot of Othello players seem so humble despite being so intelligent and accomplished is because they have such frequent experiences in having AI point out how stupid they can be. Humiliation can be... humbling.
So hey, if humanity becomes less often obnoxiously arrogant due to the humbling experiences of interacting with strong AI, that could be nice!
4. We are going to learn that there are exponentially more things that are wonderful and really do work, than we ever thought possible.
Prior to strong Othello AI, humans played a very narrow range of openings in their games. These openings were decent enough.
But once strong AI came along, all of a sudden we learned that there were *far* more playable openings, than we had thought there were. They had just seemed implausible and scary-bad, up until the AI proved that they were playable.
That is what AI is going to eventually bring us in our lives. We are going to be *shocked* by the things that work, which we had never considered, or had dismissed too quickly.
In some sense, we are going to be entering the "Age of the Counterintuitive". We have figured out most of what our human intuition can lead us to understanding, so now nearly all the new solutions are going to be in the counterintuitive areas.
5. AI will make Go players (and all humans) smarter and better at everything we do. Human Othello players have improved significantly since strong AI came along, and so have top Chess players. We can learn more, and faster, than ever before.
It will not typically take 10 years or whatever, to gain what was formerly considered world-class expertise in a given field. It will typically take 1-2 years.
We learn so much from the best humans, but we can learn even more from AI once it surpasses the top humans. Humanity will become better, concurrent with AI improving.
And as some humans get better, even the humans without access to top AI will get better, since so much of human behavior is built on cultural modeling. "Monkey see, monkey do." All humans will absorb better knowledge and behaviors, quite naturally, just like how children nowadays grow up naturally being knowledgeable and savvy in ways few of our ancestors ever managed.
.....
Those are some of my main thoughts on the future of humanity's relationship with AI. What are your thoughts? I know some of you are actual AI experts! Gunnar Andersson Stéphane Nicolet Vlad Petric Edmund Yiu :)
预测 阿尔法狗全胜 :o:o:o:o:o:o
页:
[1]