"Yes AlphaGo only won by 0.5, but it was not at all a close game. It's an artifact of its training objective."
"it prefers to win by 0.5 with 99.9999999% chance instead of 10.0 with 99.99% chance."
I really wouldn't be shocked in those final moves it was optimizing for 5 or 6 "9"s of success probability. Which will still true to programming, it also obfuscates how good it really it.
That said, there is some benefit to not completely embarrassing your human player who previously was known as the best, but has volunteered to be beaten repeatedly on international TV.
Going into this Ke Jie knew he was going to lose, and still agreed.
Speaking as an inexperienced go player, I would not be shocked if the new Alpha Go were 500-1000 Elo points better than Ke Jie. It's true skill being masked by its ultra-conservative style of play.
The original AlphaGo only lost a single game to Lee Sedol that was essentially just a bug, not because it was "weaker". DeepMind has said this completely retrained AlphaGo is significantly stronger than the original.
 Partial support for the possibly 1000 pt Elo advantage of Alpha Go. http://en.chessbase.com/post/alphago-vs-lee-sedol-history-in...
I'm not sure that's right, everything was working correctly, it just didn't read out a low-probability move very deep.
From my reading, it essentially got AlphaGo into a state where it was no longer reading the board correctly. The algorithmic bug was play a decent but extremely improbable move and AlphaGo won't know how to respond.
Or by a similar argument, I think most people would say that if Lee Sedol hadn't played that move and the game continued 'normally' he would've lost like the other games. The rarity of the move is why he won, not the "strength".
Essentially they trained the app on too specific data. Their main fix was to retrain the next version from "scratch" instead of from moves humans are likely to make.
A) Conservative play (by humans) tends to maintain the margin, not shrink it
B) Show me just one human player who would not waver under the psychologic pressure of only leading by 0.5 points
It's a half-point win which doesn't seem like a lot, but AlphaGo is intentionally taking lower-risk paths to lock down the win in exchange for some win margin. If you watch the actual game and some commentary, it really feels like AlphaGo is playing at the next level. Some of its moves are so inhuman and subtle.
This is a very exciting time for Go; a lot of traditional wisdom has been shaken up in the last couple of years! I heard a story about a Korean pro study group that was doing nothing but studying AlphaGo's games for new insights for some time. I'm looking forward to seeing the future of play as extremely strong Go AI becomes more widely accessible.
The best way to test strength, assuming Ke Jie continues to lose, would be to start giving handicap stones until the winrate stabilizes. I'd guess that AlphaGo is no more than two stones stronger, but maybe that's just more bias towards humans..
Weightlifting isn't the best example because it's a big group competition without the individual 1-on-1 aspect to it, and chance plays less of a role. Yeah, you might feel out of sorts on competition day because you didn't peak in your training properly, lift 20 kg left than you're capable of, and subsequently not place, but it's not supremely embarrassing. You're still a very strong person, which you demonstrated. Rousey, though, just got demolished.
The natural occurrence of a human being capable of defeating a bot like this would be very rare, and would take decades to train. In the case of this bot, you just set up a cluster and deploy the same software, to produce as many instances of this bot as you want in less than a day. In this sense, AI has a huge advantage.
Finally, by the time a human being can beat this version of the bot, there is going to be a much stronger version.
But Ke Jie probably will take that as an insult.
While it might not be very gratifying to lose to a computer, players often enjoy the opportunity to play stronger players and learn something new.
Then, the way to be a professional 9 dan involves a lot of losing and frustration. I don't think he is very thin-skinned in this respect.
I only took one A.I. class in school and had a few that involved games as projects and remember several times the lecturers/professors (10 years ago) mentioning that Go was a next level target after chess due to extremely high branching factor.
(EDIT: Specifically last year I was witness to a lot of threads regarding the quality of the AI in the video game Civilization 6 that asserted that poor AI in video games was no longer excusable since computers now beat experts at Go...)
The poor AI in Civ is repeatedly based around smart systems stacked on top of each other.
So imagine Software Development, you decide you want to add 10 AI systems to run aspects of the AI. You build all 10 AI systems, each one doing a thing.
Then you add them all together, and they all clash. That's Civ AI.
It's almost deterministic in that, all the AI systems fight each other in a consistent way, so you could have just ignored writing all your complicated AI, and made the AI just follow a single decision tree, and it would be the same results. Except a single decision tree would follow logic. Random systems competing with each other, has no real logic.
Then when AI problems come about, the programmers tweak the systems to fix a single problem, which can maybe work, but will break many other things in the process.
So each Civ game, to fix the AI, you start by shutting off all the AI systems one by one. The goal being when I ask the game to do X, X occurs.
Once you get all the stuff removed so you can actually make the game do something, you then create actions that make things happen.
That can make the game has significantly insanely good AI. The best example of this is a mod in Civ 5 called Vox Popeli which has the most profound Civ AI anyone has come up with, and takes ideas from all the Civ modders.
To get the AI to a point where the Human thinks it's good is not particularly hard if you focus on simple, easy to understand concepts, and not large systems that are supposed to cover every possibility organically.
Add in the fact that each player needs to build their own deck (choosing from 20,000+ cards) and that new cards are printed regularly, significantly shifting the expected metagame, and building an AI that can routinely win in a larger card pool format like Modern or Vintage seems absurdly difficult.
The list of things that humans do better than computers is getting very short very fast.
Particularly, it's usually a multiplayer game at the outset, and has many varieties other than hold 'em (which just happens to be the currently—and this is a fairly recent development—popular form.)
It's true that poker is not completely solved, but neither is chess. It is still not known whether white has a forced win. It is possible that some day this will be known, and if the answer is yes, then a human playing white may be able to learn the strategy and hence beat any machine. But that will be only marginally more interesting than the fact that a human player can force a draw in tic-tac-toe. The difference between "fallen" and "solved" is just not that interesting. Checkers is solved, but no one cares.
I see no reason to believe that the techniques used to beat the best human players in one variant of poker cannot be extended to beat the best human players in any variant of poker, or, for that matter, to any game with randomness and hidden state.
I expect StarCraft will be the next professionally played game to experience destruction at the hands of AI. Then it will be clear that "fog of war" doesn't matter much in the decision tree from a probabilistic perspective.
The irony being that faced with a "fog of war" maybe you do want to underestimate your opponent because then, if you end up surviving, you can claim brilliant insight and leadership which bolsters the short-term position, even though in fact you simply were lucky. So in the open-ended problem of tactical maneuvering in business or war, AI might not have an advantage any time soon.
The difference is quite drastic.
There don't seem to be good Stratego computer AIs (which actually surprised me a little). But I can easily believe that's at least in part because not a huge amount of effort has been put into it.
This means that the history of the game is pretty much irrelevant. Usually only the last two moves are important because you can't recapture a stone and get to the very same position as two moves before. It's called the ko rule (another unfortunate word clash with English.)
There is also triple ko (three linked kos), which requires to remember more moves back, like for repeated positions in chess. A go game with a triple ko has no result: there is no draw so the game must be played again if in a tournament. It's very rare.
True under Japanese rules. Chinese rules at least theoretically use a superko rule. But many comments online say in practice the game is voided under Chinese rules too.
Modern Go AI also has "intuitive" components just like we do (large neural networks trained to immediately provide a heat map of what should be the next best move). That NN is not limited by board size.
BTW, even when playing by "intuition" alone (no tree search), AlphaGo is still incredibly strong. It would still beat most amateur players that way.
AlphaGo should be able to be retrained using different variables/game rules.
Something organized in good spirits like an AI Olympics would be amazing to watch.
I feel like that would lead to a new generation of players trained on new strategies from a blank slate. Could be interesting.
Disclosure: I only know the basics about playing Go :)
More will be published by end of this week
Vathys changes that."
Prior to 2015, no programs could compete evenly with any among hundreds of Go professionals. Two years later, even the human World Champion admits that his competency is far below that of a computer program.
"I am quite convinced by this loss that AlphaGo is really strong. From AlphaGo there are lots of things that are worthwhile learning and exploring." -- Ke Jie
The time to start working on and funding AI Safety research is now.
Note: I am not saying that AGI is imminent. The point is that we do NOT know when it will emerge and AI Safety research is very difficult and will likely take a long time to complete.
Let it go, man, cause it's gone. A new era has dawned.
In either case Alpha Go Ke Ji is said to be 3 stones stronger than Alpha Go Lee Sedol which the team claims comes nearly completely from algorithm improvements.
j/k someone will have to program the poets
Keep in mind this is all a translation from Polish.
But computers have much lower idle consumption than humans. So if we assume alphago idles at ~45w that would consume around 1000 kcal for 21 hours leaving 1500 kcal for the 3 hours it was given to think which would be more like 580 watts.
That sounds pretty doable.
and in Chinese:
"Regarding the go match between Ke Jie and AlphaGo, no website, without exception, may carry a live stream. If one has been announced in advance, please immediately withdraw it. Please convey the gist of this to sports channels. Again, we stress: this match may not be broadcast live in any form and without exception, including text commentary, photography, video streams, self-media accounts and so on. No website (including sports and technology channels) or desktop or mobile apps may issue news alerts or push notifications about the course or result of the match."
Sounds pretty harsh... will be interesting to see how the rest of the event goes.
Sites like Sina and Sogou are able to report, and those aren't small sites. They got millions of views every day.
That site you quoted is anti-PRC news site. Not trying to discredit, but it isn't a total ban. At least the result isn't and probably won't.
That's the lowest bar in the world. It's a damn board game. The stream was six hours of two men sitting across a table from each other placing black and white stones. Optionally, some of the streams had commentators talking about how good the moves were as they were played. That such a thing could be subjected to censorship is absurd.
As for the result not being censored, well, it's literally a single bit of information. That was only censored because it's not practical to do so.
According to the tieba page, the ban was first noticed as CCTV5 deleted their Weibo (a microblogging service) post on the livestream.
Welcome to the 4th part of confidence doctrine -- cultural confidence.
I haven't been back in a while but my guess is things have sort of "settled in" more. The Chinese Firewall is absurdly effective - last I heard it was difficult to even get a VPN through.
Is this true? Could you talk about what the most common search engines in actual use are there?
Obviously I am asking about this as a total outsider. I've never even been to China.
Edit: my apologies, bad wording. Kudos to xiaoma for catching my mistake, thank you 小马 (my guess at your name in 中文）。Today, all Google properties I'm aware of are blocked. The blocking was implemented gradually across Google properties over the past 7 years. The search engine has been blocked for the past 7 years, since the beginning of the block.
Besides blocking entire sections of the net outright (like Google address blocks), poisoning controversial domains, etc, even if it can't directly inspect the traffic due to good encryption (say in the instance of OpenVPN or IPSEC), it will slowly degrade and eventually null-route your traffic over the course of minutes, depending on its judgement of the likelihood (based on packet structure and history) that your activity isn't "normal" usage.
Currently the only functional ways of getting around the GFW is VPN through stunnel (TCP OpenVPN traffic re-wrapped in TLS, thus pretending to be https traffic, and incurring triple TCP performance penalties), similar convoluted protocols like Shadowsocks, obfsproxy, and other China specific tools.
Last time I checked the only Google service available now seems to be the Google Translate.
One project monitors how many popluar sites are blocked in China:
Tor is also blocked so it takes a considerable amount of effort to make it work. VPNs and other measures (socket proxy, etc.) work but they're also sinfully slow, and it breaks all the time.
The most common search engine there is Baidu: https://baidu.com. Bing has some market share as well.
I see claims everywhere that "China is censoring coverage of the match", but anyone can just click on those links and see that it is untrue.