AlphaGo - Requiem

In this video I show almost no variations, but talk about what I found most interesting and most important about AlphaGo's recent series of games in China.
AlphaGo - Whatever You Do Is Wrong - • AlphaGo - Whatever You...
AlphaGo and the Hand Of God - • AlphaGo and the Hand o...
In-seong Hwang's AlphaGo video - • The future of Go, Alph...

[@MassDynamic]

if an opponent can purposefully give away points so that he/she wins by .5 (basically 1), it shows a scary level of confidence. i'd be terrified.

[@okuno54]

Indeed. I may not know much about go, but I know there's the saying in chess that when a grandmaster plays an edge pawn towards you, you've already lost.

[@apppples]

Trolling programs, hilarious jokes, philosophy, and ubi. Everything a great video needs. You're amazing man, truly amazing. I'd totally subscribe to brady's bullshit. Just saying.

[@kimmokeronen3275]

I would subscribe to Brady's Bullshit. I think you are very good at articulating your thoughts. It makes a huge difference enjoying these kind of videos.

[@letao12]

Me too! Heck, I would donate to Brady's Bullshit if that helps make it happen.

[@sampovuori4738]

Yes Brady don't undersell yourself. I bet there are quite a few people like myself who don't know jack about go but watch your videos because they find your meta-analysis interesting and your voice soothing =)

[@ryandietrich8604]

Agreed. His points at the end are dead on, and it's been on my mind a ton as of late. Self driving semi-trucks are going to put a lot of people out of work here *very* soon.

[@zsoltbihary3347]

I would subscribe to Brady's Bullshit.

[@EugeneOneguine]

Your thoughts on Ke Jie's tear is what first came to my mind. The comparison with Mozart and Salieri is so perfect it can hardly be called a comparison, it's the same thing. Just as Ivan Tourgueniev cried when he heard Fiodor Dostoeïvski's reading of the Karamazov Brothers, went to him, and said something in the lines of "I might be getting all the prizes, all the rewards, but you... you're above me".

[@MadaxeMunkeee]

I really appreciate the time and thought that you put into these videos. Definitely this is the most insightful commentary I've been able to find so far. You're awesome.

[@qianweijia1]

I rarely see coverage of other alphgo games, so I really appreciate that you did these and at a level where an enthusiast who has no knowledge of Go can understand

[@fleecemaster]

Yes! Been waiting for the Brady analysis of the recent AlphaGo stuff, I even watched about half the matches this time! :D Thanks for this video, really beautifully said Brady :)

[@razortrade]

Brady, I believe you summed up the strategic flaw in humans that AlphaGo exposed. At 10:15 mark of this video..."It's obvious once you see it, but you don't even think to look for it". That is extremely insightful, and applicable in so many aspects of human activity. I am on a search to see more and expand my thinking, even when it feels like I am up against the brains ceiling. Thank you for this video. It is well produced and filled with great content. Jeff

[@itai82]

I absolutely love your explanations! Thank you so much!!! The martial arts metaphor was fantastic :-)

[@jamblejuice]

49:44 Go is also a great problem because, when compared with games like chess for example, there is no natural and easy way to write down a value function, which would give AlphaGo some way of quantifying how likely it is to win given a board position. However, this is something that humans can build up an intuition for. Hence the need for training the value and policy networks you mentioned. So you could say that, because of this, successfully beating professional go players is a step in the direction of general artificial intelligence. However, there is still a long way to go and, as you said, the effect has not been immediate. But I think the reason that it has not led to more immediate results with real-world problems is that go has a well-defined structure. As vast as the number of board positions may be there is still a countable number of them. At any board position, AlphaGo has a fixed set of moves to choose from. Of course, each of these leads to another fixed set of moves which leads to another and the complexity begins to explode. But there is a very simple structure to it. Compare this with a self-driving car as an example. At any point in time, the car must make a decision on what action to take based on input from sensors. But it is not faced with a nice discrete set of choices to make. It must make a sequence of decisions in an uncertain and changing environment. It cannot predict what will happen in the environment in the way that AlphaGo can predict how it's opponent may respond to moves.

[@MrStarchild3001]

This was hilarious! Great video, very enjoyable commentary.

[@apppples]

3 am. Can't sleep. Alphago video shows up. God kept me up for a reason!

[@timothy1116]

Great ideas you've shared..Thank you for making go more accessible!

[@normILL]

Brady, I was a national delegate for one of the major political parties representing my state in 2016. I'm not saying I'm a political genius, but I was on top of my politics to get there. What you described is exactly why I left politics and changed my career. I went into politics to make sure that we'd be alright as a species, and while yes in theory nukes could fall tomorrow, there's so much momentum built up in the way things are already going that it just won't happen. Not for a while at least, and during even a short while technology will make the final necessary leap needed to get to that society you describe. (I think of it as most people become professional artists.) I'm so confident in this I left humanity to its fate and changed my career, Brady. I've loved this series on AlphaGo, and really appreciate you bringing this moment in history to life for people like me who are completely new to Go. If you ever make merch for this channel (shirts for example) with a reference to this series on AlphaGo, I'd be proud to spread the word by turning myself into a walking billboard for this series.

[@ariaden]

> I think of it as most people become professional artists. That assumes artists would be still valuable and deep learning algorithms will not become better artists. (I do not think either will be true.) I think last jobs for humans will be in project management. Second to last: STEM.

[@hyazinthpoem1624]

Project management would change a lot and not be project management anymore as there are already some experts saying SAP will replace McKinsey and such - so essencially programs being better at managing than humans. I don't think personalization is working too well yet, but who knows how predictable humans really are ;) As music is just a combination there could be many computers producing any combination possible but at least the effects are different and it would be difficult to find the "masterpieces" in all these variations... musicality does not necessarily need much skill. ^^

[@gunnargrautnes4451]

ariaden Actually, I think there's going to be room for work in the humanities, for two reasons. 1. Even if we have access to enormous amounts of automated work and live in some kind of post-scarcity civilisation, there will still be plenty of room for ethical, political, aesthetic, and theological debates. Even if we have armies of robotic workers that could turn every solar system in our local group into Dyson swarms around conveniently placed red dwarfs, should we? Should we uplift cats? Is there a sound version of the cosmological argument for the existence of God? Should we eliminate all suffering, even if we could? How do we make decisions that affect all of humanity? Does the future heat death of the universe make all human endeavours meaningless? Is there something it is like to experience the colour red? Is transhumanism compatible with a commitment to the unity of the umma in light of Allah's tawhid? We might create deep neural networks that create arguments and documents we find persuasive and evocative in answer to these questions. I still have a hard time believing that computers could ever fully perform the function of ethical and theological experts in a satisfactory and comprehensive way. Perhaps that is just me being a hopeless romantic. Still, wouldn't there still be luddites around who would rather be persuaded by human experts than the most brilliant robots? "Greetings fellow humans! I'm Mark, envoy from our omnipotent benefactor Deep Thought to the Amish Empire!" 2. Philosophical logic. Unless my professors have been lying to me, there are areas of philosophical logic, specifically stuff dealing with natural deduction, that in principle can't be done efficiently by computers. Of course our brains can do natural deduction, and they're physical things, so I guess some radically different design of computers from the kind of machines we have today might still be able to do it. (Possibly, this has to do with the p vs np problem, but that's beyond my knowledge of computers.) Anyway, I think that if you want to be among the last people put against the wall in the robot uprising, you should study philosophy.

[@okuno54]

57:18 "I've probably gone on about this for like..." *empty wrist* "...five minutes now" :D If you haven't read it, I'd recommend the novel "Too Like the Lightning". It paints a very strange vision of the future, including some post-work ideas. From that book I've already added the word 'voker' to my vocabulary to think about the kind of world you describe.

[@Technomancr]

Check out reddit.com/r/BasicIncome/

[@brettw7]

Brady, I really enjoy your analysis. I really want to hear your analysis of AlphaGo Zero!

[@wilhelmsarosen4735]

Very interesting insights. Any more videos coming up?

[@kashgarinn]

Really nice videos on Alpha Go, I'd encourage you to make more if you feel like you have something left unsaid, for instance if you want to go into more detail on each game, or the released alpha vs. alpha games.

[@CattyRayheart]

AlphaGo is the Apollo project for AI. It's a prestige project, a long shot that is hard but probably not impossible, but it's effect are a wide variety of improvements in deep learning, in computer science, in engineering, and more. There have been some really interesting papers came out of the project. We have learned so much about how to leverage the strengths of AI from it and how to better combine different AI approaches. The project probably did advance the state of AI research by a decade. I also don;t think you have to worry about another recession putting peoiple out of work, in a recession there is a supply of labor looking for scare jobs, so the value of labor is low. It's when the economy is doing well and overheating with near full employment, lots of demand for labour but constrained supply that the significant costs involved in automation make sense. But then the jobs don;t come back during the recovery. We will probably need some sort of BIG where the government just hands out money to people, that solves a lot of problems anyway at the same time.

[@samche11]

Miss your videos ..

[@kfm1242]

Amazing, thx!

[@Goombario37]

I hope we get more videos in the future, really miss your videos.

[@u.s.4129]

Hello Brady, you have by far the most interesting yt-GO-channel I know. Your style of commenting remembers me reading gödel, escher, bach and metamagicum some decades ago. I very much appreciate it. Greetings from Germany ....aaaaaaaand I would subscribe to Brady's Bullshit 😉

[@Teth47]

Two years? Last I heard, there were at least 6 companies with Go networks that were approaching AlphaGo's performance already. IIRC Samsung is one of them.

[@quebono100]

38:30 the reason is, that its all of the sudden not playing this move anymore so often, its because in the many simulations it does, its find new deeper pattern where a specific position is better then the other.

[@jamblejuice]

Loved the mama analogy!

[@knocknwalk]

Another interesting and insightful video on AlphaGo's play and impact. A nice overview. I have enjoyed the level of game analysis you've provided in these Alphago episodes, but as I no longer play go much I don't think I'll be viewing many of the other Blunders videos. Like others, though, I'd subscribe to a BBullshit channel. The restructuring of society is a very important and interesting topic, for one, and deserves thought. (I think finding a plausible path of transition may be hard but Is much needed.) Multiway discussion isn't viable with the KZbin comment section, so maybe an associated ranked-comment venue like a subReddit could be set up for discussion of BBullshit videos.

[@orngjce223]

I can't say about Go, but from following people who are futurists, the "computers will eventually take _everything_, mind jobs are no longer safe" prediction that you offer here seems to be extremely common. There are a few fields like psychology or AI programming where humans might be able to hold their own for at least a bit longer. But many people simply aren't smart enough for those jobs, regardless of how much school they go to or how much training they receive. So it's not even "what happens when there aren't enough jobs?" It's "what happens when the people who are unemployed can no longer perform the jobs that are available?" There are quite a few attempts at answers to the question, some of which I even like. But the question itself is a very unsettling one.

[@saldownik]

Phew... Now I am less worried about machines taking over my washing machine assembler job.

[@paysonfox88]

The Alphago analysis of the game indicated that the game was not decided until after move 120. The Ko fight up until move 101 had it at basically a 55-45 to 60/40 in favor of Alphago where it takes over 70% winning estimation for a "won game" to be declared by Alphago. So until you get into the high 60's to 70+% losing rate, the game is still on in reality. The Team and Dennis were very impressed with Ke Jie because he kept the game from being decided until after 100 moves in. Every other game was 70/30 or more after just 30-35 moves in. That's crazy!

[@AirIUnderwater]

It would be so depressing playing against AlphaGo at the endgame. There it is, just giving up points and you're counting it out and thinking, "if only I could jump out to another 1 or 2 points, even half a point, I could win this." So each time it moves you're there maximizing your points thinking that by the end you might have a chance. And AlphaGo is just there dancing around the board and then at the end you still lose. It's akin to playing a fighting game against someone who's way better than you and all they do is jump around hitting you with one attack button. :(

[@joseki34]

Great insightful analysis. Agree with all your points.

[@FloydMaxwell]

Ultimately computers can see farther ahead than humans. In Go, at the point where they can see a win, they ask the human to stop. The human can't see nearly as far and wants to keep playing. The computer, playing advanced moves that lock up future points, proceeds to baffle the human, who has no doubt still failed to see his future demise.

[@magbhitu]

That bit about the new alpha go needing only 10% of the computing power when it was training itself up with monte-carlo is really amazing. I've seen similar indications of this in regards to other AI applications. Once the big AI's on the cloud have optimized a procedure it is possible to "chip" it into hardware and not need the heavy processing power any more. This is what will really drive the AI revolution in terms of the consumer level impact...

[@frogola]

Bravo, Brady!

[@columbus8myhw]

The phrase you want is "universal basic income," I think

[@u.v.s.5583]

The Salieri reference was not exact. Salieri is the teacher of Franz Liszt, Schubert and a certain van Beethoven. The sum of the three plus Salieri himself makes a far greater musical teritory than Mozart by himself. But, man, your videos are such a blessing. I wish You continued to do this. At least until there comes a Google Alpha BradyDaniels I'd even pay to watch you :)

[@fostena]

I don't know if you are a computer scientist (I am) but let me tell you: your insights on AlphaGo, on the future of AI, and the impact on our society are correct, as far as I know. Of course no one can predict the future, I'm just saying that there's no clear "bullshit" in this video, this is more or less our current understanding of those matters. Good job

[@geoffrey5045]

I came here from the Whatever You Do Is Wrong video to double check the advice Im about to post. Move the camera to eye level or slightly higher - is slightly uncomfortable to feel like we're looking up at you. You have a habit of clearing your mouth of spit during a pause, results a sharp high pitched sound (18:20, 23:09) I really like the AlphaGo analysis. Keep it up mang

[@guoalber]

I really enjoyed watching this video. I have a few comments and questions - 1. Where did DeepMind post the analysis of the games? Specifically the follow up to game 1 where Ke Jie should have jumped? I'd love to read/watch the full analysis from the DeepMind team. 2. Haven't a lot of the games played between AlphaGo and humans been blitz games? It might be the case that getting two handicap is worth much more than 10 points for a professional simply because the initial 4 moves of the game have the most leverage over the rest of the game. For stronger players, stones on the board are much more powerful than a few invisible captures. For example, despite you being a "low dan player" you probably would win all your games against a pro player giving you 7 stones (if we thought of the difference as 2 dan vs 9 dan as 7 stones). That being said, I personally agree that 2 stones is probably the floor, though perhaps pros of Ke Jie's strength could win with no komi. 3.Go might be relatively easy to solve (compared to real world situations) not because of the lack of variables, but because of the strict rules and structure of the game. 4. Are you going to the US Go Congress this year?

[@Sam5peed]

Have you heard of what ASMR is ? I think you have a way to explain and speak that is relaxing and calming. Plus, interesting ! Keep up the good work !

[@SoldierOfTwilight]

Very nice video

[@johan3595]

Dude I don't even play go but I love your videos

[@gorilaz0n]

Hey, Brady, where did you find the photo of the 5 players face palming simultaneously?? I wanna use it as a home screen!

[@M7k863]

Hi Brady! I'm 2d from Russia and I really miss your videos. I understand that you've got more important things to do than making new video but there's a thing I want to ask you. Community of russian go players is huge and it would be great if you could open access to subtitles, so I was able to translate your old videos to Russian. Thank you and looking forward to new videos!

[@BradyDaniels1]

How do I do this? It would be my pleasure.

[@M7k863]

Thanks for your reply! "Follow these steps to turn community contributions on for all videos. Note that this change will reset any settings applied to individual videos. Sign in to KZbin Studio beta. From the left menu, select Transcriptions. From the top, select Community. Select Turn on"

[@BradyDaniels1]

Done, thanks for your help. Please, let me know if you add subtitles.

[@chaof9501]

2021 here. Ur right, bye bye jobs.

[@briankamstra2454]

I suspect we will not see the end of Go (or Go professionals) just because of this victory for Alpha Go. Chess, for example, has carried on for quite some time now despite AIs beating the best human players. Likewise, utilizing AI to help analyze games and as a learning tool for Go will grow. To your larger point though, I think you're right that we're going to see AI make a whole lot of things obsolete. And as a society we'll need to be having the "redefining work and what it means to us" conversation long before the "brain" jobs are even half gone. Self-driving vehicles, for example, are poised to put millions of people out of work in the transportation industry in America alone. I hope you are right looking towards an optimistic outcome :).

[@FloydMaxwell]

Like the Amadeus comparison

[@himanshuwilhelm5534]

2:18 Because a computer found god's move before he could.

[@v4alien]

this is good for studying english via my love go game

[@dtracers]

in it's self play games they did full time so I bet alphago had like 2-3 minutes per move so the self play games are much more read through than the 30 second games. that is why some moves are different between self play and human play

[@viibridges]

As a Chinese, I sort of know what's in Kejie's mind at that moment. At round 3, Kejie fell's behind very quickly in the early stage, he was expecting intense fights like the second game, but alphago didn't leave him any opportunity. Round 3 is his last chance (probably the last game that human will ever challenge alphago), but he failed himself not making it perfect (relatively, from human's perspective). He couldn't take such a humiliating ending, that's why he cry. If switch the order of game 2 and game 3, I guess Kejie would had been very satisfied and proud of himself.

[@infinitysalinity7981]

Alphago died and Alphazero was born. :)

[@byronwatkins2565]

There's always a bigger fish.

[@Broockle]

I think about the Automated future too. The transition will be very messy. It's gonna be interesting. P.S. They could make another Go playing AI which values Trollage, and that always seeks to maximize it. We can call it GoTroll ^^

[@CrazyVulcan]

My thinking on the app that google is coming out with will be a game titled "Ke Jie's Master GO"

[@ig2d]

I would like to think that in the future the top players will be able to defeat the top Ai from time to time. for one thing they will have Ai to learn from - or perhaps I should say unlearn from - and the method they will use to do it is to create a high level narrative in the game

[@vdinh143]

If our basic necessities for life can be supplied by machines, there's no reason a StarTrek utopia can't happen. Instead of being a mandatory means of survival, work can become our voluntary passion, a way to actualize the human potential (like the many great people before us). Scientists, singers, actors, artists, composers, etc. All the activities that are now unsustainable hobbies will receive a life time's worth of effort and dedication. Maybe society could have a base "allowance" to take care of the basic needs, and jobs will fund a better lifestyle regardless of what you do.

[@jim42078]

Hope AI does hurry up and come for the bankers though.

[@MelindaGreen]

And lawyers judges and politicians.

[@fleecemaster]

I wonder how you'd feel if they start doing an even better job though, and are even more tight and harsh? Be careful what you wish for...

[@MelindaGreen]

If they're fair and focused on doing their jobs as well as possible, then I definitely won't complain that I can't get special treatment.

[@Draugo]

Fleecemaster But as Alpha Go has shown it doesn't do the most profitable job possible but just enough to stay on top. So if we got bankers etc that don't aim for highest profit possible but just enough to fill its niche I'd say it's actually prefetable.

[@fleecemaster]

Who says AI bankers would be programmed to do "just enough"?

[@KimMilvang]

Do you have any idea why the deepmind team has not showed interest in playing handicap matches? From an AI perspective that would also seem like an interesting problem.

[@barakeel]

Kim. Most likely AlphaGo Zero can play handicap games but it's not what it was trained for, so it could just start making overplay in the beginning thinking it has already lost the game.

[@dtracers]

you forgot 2 parts of training. #1 they would have it play against old versions for data in addition to itself. to prevent new blind spots. they also made a different network which has the goal of tricking alphago. which also helped find blind spots.

[@QuantumConundrum]

Brain jobs --> Discovery jobs

[@MadaxeMunkeee]

As a more constructive comment, I think the other reason fortune 500 companies are going to pick up AI is because soon (or even now) they are going to outperform humans at these intellectual tasks (accounting etc). So if you don't upgrade, your competitors will and this will give them too much of an advantage.

[@rossdixonellis]

From your description of the first game being only a half point difference. Kinda reminds me of Data. AlphaGo is basically playing like Lieutenant Commander Data from Star Trek the next generation. That game stratagema. It's playing to stalemate or by a small margin.

[@rossdixonellis]

"I did not win... I busted him up!"

[@undertyped1]

Where can i watch the 5 chinese pros vs alphago? i cant find it.

[@nickmiller3501]

Hey Brady , no more blunders ? You must be as good as Alpha Go ...... missing your videos .... please come back

[@schnipsikabel]

unconditional basic income is the future :)

[@MarkGaleck]

Brady for President 2020!!

[@vodrake]

With regards to the 7.5 pts going down to .5 win, it could be that since it knows it has won it wants the game to end as soon as possible. If it has 5 possible moves as you said, 1 has an 80% chance of winning the rest 100%, well its going to choose the 100%. But how would it decide among the rest? Well if it doesn't care about the number of points so long as it wins, then it might seem logical to base its decision on the number of remaining moves. Since its only objective is to win, it will do so as fast as it can while ensuring that it will certainly be a win. If playing 1 move would result in the game ending a turn earlier even though it would loose points, then why not play that?

[@GuiboyAgamennon]

As far as i can understand convolutional neural networks (and i don't know much), i dabbled with AI but my work took me elsewhere, what it does is reducing uncertainty, all it cares is the win, if a move loses points but reduces uncertainty, as in, drop points but makes the win statistically slightly more likely (inside some sigma) it chooses that move. 100% win by 1/2 point will always be chosen over 99% chance with a 20 points lead

[@Marksman560]

Haha score-trolling AlphaGo :D [edit] Talking about your utopian view... Anyone mentioned the television series, The Jetsons? :)

[@decidrophob]

You say that there was no spectacular move like the ones in Sedol Challenge Match. I disagree to some extent. I think it is only that AlphaGo raised the bar too high for what are regarded as spectacular. I loved the second game. This was the first time a complicated joseki of star point - approach - keima -attachment appeared. The way the game proceeded with gigantic kos was simply spectacular to me. Especially AlphaGo's ignoring the right top huge ko threat preparing move by Ke Jie. Probably almost no one could expect that.

[@shadowmil]

Hey, I'm an AI developer and I think I have some theories on why AlphaGo seems to forfeit points at the end of the game. If you want to hear my thoughts on it, I'll be happy to discuss it.

[@a.w.s2606]

Charles Miller please go on...

[@shadowmil]

@@a.w.s2606 I can only hypothesize. But the AI is trained to win, not to score points. It cares about a win by 1 point as much as it cares about a win by 10 points. But why not do the optimum thing? Why forfeit points when you could claim them? I think it might be related to the training method. If the opponent does very well but you beat it by a small margin, the training algorithm might put more emphasis on learning from that game. But if your opponent does poorly and you totally destroy them, the training algorithm might believe there is less to learn from that match. The result being, in games which the AI forfeited points might get reinforced as a better way to play. Basically they might've accidentally trained it on a different rule set where the AI thinks the goal of the game is beat the opponent by as little as possible. I've seen similar issues with training (other games) where matches that go on for longer generate more training data, and thus are over represented in the sample. Basically ends up teaching that longer matches are the correct way to play, causing them to give up easy and quick wins.

[@a.w.s2606]

Charles Miller thank you for your reply, it was honestly unexpected haha! Your assessment seems to make a lot of sense, although if that is the case it, why is it that they do not just adjust alpha go’s parameters to win by the largest margin possible? Although maybe I am just naive in thinking that would be a simple adjustment!

[@shadowmil]

@@a.w.s2606 There are problems with attempting to score highest amount of points. It's a very greedy optimization and often results in unintended behavior. For example I remember a case study of a racing game which allowed cars to pick up respawning items to score points. The result was in the short term, picking up these items net more points than completing a lap. Generally more AI programmers agree that having a win condition is the appropriate yard stick for training an AI. Afterall, even if Alphago does endup giving up potential points, it still has a higher win rate. Of course, you could always try to combine the two functions. But a more complex training function can slow down the training rate.

[@Keldor314]

I think Alphago slipping at the end of games is actually a manifestation of the Monte Carlo Horizon problem. Similar to the way an AI will start playing very badly when it's behind and has no moves that lead to any chance of victory, and so ends up choosing more or less at random, when it's way ahead, there are several moves that still have a near certainty of leading to a win, so it chooses more or less at random. Alphago can't see any meaningful difference from loosing by 2 points than from loosing by 70, and also can't see any meaningful difference in winning by 1 or winning by 15.

[@zielad362]

hey brady what is your go rank?

[@armin0815]

While some will no longer want to invest strongly in a game lost to the machines, at least it's probably for some time that humans will be the best teachers for the game.

[@dtracers]

Armin von Werner we need to pay people to explain alphago to us. because it can't explain itself. I think computers won't be good at teaching the beginners. but I think that training under masters may be less important than training under computers.

[@edale2]

Begging a troll not to troll you.... You've doomed yourself. }:-)

[@thom1218]

Bottom line is that being able to deeply read out all the most plausible moves to dozens and dozens of moves ahead, will always win against an opponent who cannot, all other things being equal. It seems that Alpha Go (re: computers), not surprisingly, can be made very good at this task, beyond what any human can accomplish with enough self-play.

[@MarkGaleck]

No. That's the difference between Computer Chess and Computer Go. In Go, there are two factors that prohibit "reading out" to certain depth ahead. First, the move tree branching factor is in the hundreds, far greater than in chess. This means, that while in chess you can enumerate the whole tree to a depth exceeding human calculation, in Go, you cannot use such "brute-force". It would take a computer with more atoms than the Earth, to enumerate the tree to what humans can intuit. Second, even if you could enumerate moves, it would do you no good. In chess, there is a number of good heuristics, the biggest of them is the sum of value of all the pieces. The heuristic evaluates who is ahead - in chess, it is not that hard most of the time to tell who is better. So for any leaf of the enumerated part of the tree, you can evaluate who is better and this is how you search the tree. But in Go, there are no heuristics. You have to use intuition and understanding to tell who is ahead. Bottom line, while in chess, the computers win because they can use mindless brute force calculation, in Go, the computers start to win now, because they start to have "intuition", they can truly understand the game.

[@nothingiseverything827]

Are you still around?

[@GRAYgauss]

I believe that alphago is showing us that humans were never "good" at this game. Our received knowledge...Our good moves...work well within the human limits. They are "good" not because they are the best, but because they are understood well and are good enough for human players. Alphago is already doing things that requires far too much computation which is what gives it it's incredible aji/seki balance. Humans tend to lean on one or the other but alphago balances them dynamically. it's not limited by bias, or human understanding or perception, capability to see how moves will play out. I think Go is still an amazing game for talking between two people, but we should accept that we are really bad at it. Even the best of us are bad at it, horrible at it. this is just the beginning for AlphaGo and future AI.

[@GRAYgauss]

Turns out you said this in this video, I tend to agree with a lot of your points. Most of your videos that I have watched you literally read my mind. I I am a programmer and have been fairly active in machine learning. So, in no uncertain terms do I see AI as an extension of ourselves even if it overpowers and dominates us. Let us evolve and be decimated, if you will :) I also agree he was crying from the sheer beauty although probably distressed as well. At first we criticizezd alphaGo, but I think now it's becoming apparent it just has an alien thought process and foresight which we cannot surpass

[@GRAYgauss]

Oof! When you brought up Salieri! As a little kid it literally used to be my fear to meet my Mozart! Love that movie.

[@GRAYgauss]

I feel that AlphaGo's thought process is always taking into account every stone on the board. For example: It's global plays are always taken to impact local plays. It may try to reduce a local stone values for the future and sacrifice them more easily while increasing a several local area's interplay. Or increase stone values in a local area as a future foothold. Humans, even at our best, are limited. We can't minmax down to the thousandths place every stone on the board. We have to rely on memorized patterns, known variations, etc in order to do anything. We tend to clean things up to make it simpler. Everything we have learned is more or less optimized for our hardware capacity. (Which in it's own right is a performance optimized masterpiece that AIs have not even approached yet) However, since we can waste MW or GW on computing power, it can have perfect memory, capacity, computation...Of course we can't comprehend it. Meanwhile, we behave in such a predictable manner, of course it can easily dance around us.

[@sargecad3t]

At what point does metagaming become a part of the actual game?

[@put4558350]

dream job of ai both Alpha go and Watson is become a doctor. It will take a while but extended human life is a great big goal. before that alpha go will try StarCraft II a real time, not have full information as next target. Watson also learn to become chef

[@oreole9608]

What's your native tongue?

[@randomusername6]

I'm sorry, but I have to disagree with your ''3 stones handicap" hypothesis. AlphaGo has no concept of being slack or aggressive. It doesn't think 'I'm winning, might as well hold back'. It plays the move which leads to a higher chance of victory. The fact that it slows down so much near the end just means that it's the better way to play an endgame and that our preconception that 'losing points is always bad' is wrong. Just because it plays these moves in the endgame doesn't mean that it will play them in the opening.

[@Draugo]

Sir Fapsalot But you've missed the point. Alpha Go plays slack moves by human definition because those are the teaching go moves. If it has a choice between a move that would crush its opponent in 99 percent of games or one that loses 20 points but guarantees a 0.5 point win then it will choose the latter. That's the essence of teaching. Only give out as much as you think your students can handle. As stated many times in this video alone when playing against itself Alpha Go doesn't make slack moves by our definition since it doesn't find any guaranteed victory moves. It's not just about the end game and that has been mentioned in other videos already. Alphan Go has mad a bunch of moves in all parts of the game that we would consider slack moves just because it can calculate more correctly the point outcome of each move and 0.5 wins are easier to guarantee than wins by a huge margin.

[@randomusername6]

Draugo I think you're contradicting yourself. Teaching go is about limiting yourself, playing closer to your opponents skill level. Alphago doesn't care about that. Its utility function is 'achieve a win', so it always plays its best possible move. The moves only look 'slack' to us because our understanding of the game is shallower than alphago's.

[@JimWilliams]

You don't get it. Hope to be absorbed into the MIND.

[@kookaichanda5717]

People today still look down on food-stamp recipients. Farming is only ~1% of the economy. You would think feeding people for free is no big deal. We are conditioned to loath entitlements. Some even think clean air/water is not a human entitlement. My question is who is minding the shop for humanity? The danger has been pointed out: The development of full artificial intelligence could spell the end of the human race. www.bbc.com/news/technology-30290540

[@UrbanNilssonOssian]

Regarding the automation vs jobs question: www.theguardian.com/society/2017/feb/19/basic-income-finland-low-wages-fewer-jobs

[@d95mback]

I agree with the conclusion, but I don't agree that the libertarian solution is "let half the people starve". Automation/AI will make things more efficient. This means that less people will have to starve. What it also means is that, relatively, human connection will be more costly. So, things that cannot yet be automated, like human emotion, human company, human interaction, will be worth more. Things that humans have without having to have qualifications. Qualified human labor can be replaced by AI, but humanity itself cannot. If you think it can, then we are at a point where AI is indistinguishable from humans, when AI _are_ human. But that's far ahead, and no political system can counter that, I suppose.

[@Aziraphale686]

I'm really curious about how and why alpha go developed this sense of faux-sportsmanship. Why does it just give away points in the end-game for seemingly no reason?

[@fleecemaster]

It's basically because as long as it's playing moves that will win the game it doesn't care which it plays. I guess the best analogy I can think of is in a game of football, if you're winning the game 5-0 you might kick the ball off the pitch in the last few seconds, even though it doesn't help you, it doesn't matter cos you've won anyway. Getting another goal isn't going to matter, and so the player doesn't even attempt to kick it towards the goal. Does that maybe help?

[@cellardoor70]

Alphago has retired?!?! :-( So they're not going to develop a marketable version like it happened with strong chess programs? This is a huge loss.

[@quelorepario]

If we replaced "go" for a "operations theater", the AI would comfortably sacrifice a whole branch of the military if it would win the war. Officer: "Sir our units are still on the ground!" AI: "Irrelevant, drop the nukes and we win the war!" If this is the type of AI that we are training for the rest of the applications we are f*cked.

[@FloydMaxwell]

Computers have helped chess players become better. Computers will cause go players to quit playing go.

[@Fireefly100]

AI :D

[@BPEMETO]

I'm still waiting to see a real game between AI and a human player ... "Time for calculation" factor has no place at this level of games. Why 3 hours? We want to see human mind against AI ... of course a computer can compute faster ... A human brain needs more time to produce better moves ... after all 2-3 days game was something usual in the past ... I hope the people from deep mind will have the courage to do this next time.

[@infinitysalinity7981]

2-3 day games are pretty much obsolete today.

[@xxinfinity00]

If no one has jobs because they've been lost to automation, then no one has resources to trade for the items/services supplied by the automation. If the people you fired and replaced with machines can't afford to buy your product, then who are you selling your product to? Its a self limiting system. We'll probably see strong growth in the entertainment sector as the financial services sector shrinks. Libertarians fear that the State will make and enforce that decision to let some people starve. A more likely scenario is that if determining a solution to this problem is left to the state, they'll try to regulate breeding and do a poor job of it, take China for example.

[@eggory]

There's a difference between a type of human labour being non-competitive, and its being non-viable or non-productive. It is still a fundamental fact of life, no matter where technology takes us, that if a man needs something (and *if* he can't supply that thing to himself by some "magical" means) then he can labour to produce it himself, or produce something else for trade. That's never going to change. If labour becomes more efficient, cheaper, even partly free, won't that just cause massive deflation? In other words, you get less money, but as a consequence of the same economic forces, you need less money to buy the same things. Therefore even if a computer can do it better and cheaper, that just means you have to work for the same wages as a computer, and use that money to purchase goods and services at a stronger "exchange rate". Will this mean more wealth inequality? Yes, since the ones who own the machine capital will be able to skim more off the top, I suppose. But they can never benefit from selling to you for less than you can afford to buy for, just as a fundamental fact of economics. The less money you have, the smaller a profit margin they're going to have to accept in order to make any profit at all. The result is that the value of money goes way up. Most people will have less money, but the buying power of the amount of money they earn will have to stay at least close to the same, or perhaps even improve. I'm not pretending to be really well informed about this, but just pitching it out there, to see what anybody thinks.

[@MelindaGreen]

Tax robot productivity to pay a universal basic income, then we'll all have an incentive to improve the AI.

[@Molb0rg]

Brady, if you will make bullshitting channel as good as you do those videos - I guarantee you a lot of subs. Take look at Sargon Akkad channel or Isaac Arthur and such - people like when a human can explain good and put his work on his thoughts - people appreciate that. As you concluded by taking look at possibilities of how it can evolve, this here a bit more firewood for your thoughts, might be a bit rough and unfinished looking space solutions is our answer, replacing 90% is not much, replacing 99% is not much enough to change us to not working etc (it is just one of at least 3 possibilities, one of which you mentioned, 3rd I have forgotten already :) I just know it is always 3 possibilities or more) so the firewood - medium.com/@molborg/robots-replacing-human-in-future-e8de498fb34d

[@Letrus100]

I disagree with your view that jobs will be lost, because all that automation of any kind does is make what was previously unmarketable and did not provide enough value, be able to support a market. 20 years ago you would find it hard pressed for anyone to make money playing video games but now there are many people on youtube are doing just that. So in this kind of sense, think of that occurring to every market. There is also the fact that money supply also reacts to supply and demand and if the physical value for things is cheaper and there is less money supply(income) then the standard of living goes up still, and there are more jobs created because you don't need to make the same amount of value to get what was previously harder to earn. Even if AI took EVERY job (which is pretty much impossible due to human nature), all that would mean is people would be playing video games with their friends for a day to earn their year's worth of food.