Alpha Go: “I learned from the best humans.” Alpha Zero: “That was your first mistake, young padawan.”

I had no idea that it used so much less processing power. That's an insane leap.

Anyone who thinks these games will die as computers get better is wrong. We still watch running events even though we've invented cars. There's something special about watching humans battle it out. Also we'd probably be that good too if we could play for days straight with 0 distractions, 100% focus, and a body hooked up to a constant power supply. But that processing power is.. wow

This is one of the best channels that is short and sweet. Thank you so much!

Computers may be able to beat me at chess and Go, but they are no match for me in kickboxing.

Watch on 2x speed and it will be 2min paper :)

Great video! I had heard about AlphaGo Zero but had no idea how much more efficient it is. Amazing!

Elo Rating over 5000? Come on guys let's get it OVER 9,000!!!!!

Would love an in depth look at the actual paper, with details into how they train the neural network, and how they use the monte carlo tree search (from what I understood, each training game it played against it self did large amounts of in depth searches on future possible moves, that it evaluated with the neural network, and played many paths to finish on each move), what features they use etc, as there are changes here from previous versions of AlphaGo. Also how they choose/discard branches to dig deeper in the MCTS would be very interesting to dig deeper into.

Incredibly well done video. It is in fact A GREAT TIME TO BE ALIVE!!!!! Thank you!!!

Wow congratulations on 1M subscribers!! Love your channel!

I wonder how much humans will improve after being able to use the examples set by alpha go zero and being able to play against it. Chess computer had a huge impact on how quickly humans could improve after all

Unnervingly delightful. That power consumption comparison is thing of beauty. I often like to compare the processing power of things to their equivalent in skyscrapers worth of vacuum tube computers. A modern phone would take a city of them, but Zero here, is like a full state of them. That could consume all the worlds power. Now its running on a single machine and were still hammering progress home as we start to peek over the event horizon. What a time to exist.

Please make a video about the new go strategies AlphaGo though us, or what the masters learned from the AI.

we live in the most remarkable time. What an adventure!!

simply amazing, thx for sharing!

better without being taught human flaws, eh?

wow, in 1 year we get a super AI from 176gpus to 4 TPUS ???? That is crazy , with a supercomputer and that kind of technological power, we could design new materials, aerodynamics , energy storage, economical theories... at that velocity , 2020 will change humanity forever.

Very interesting, good job! One way to categorize is intelligence is into crystallized and fluid intelligence. Crystallized being what you know and fluid being how fast you learn. I wonder how this algorithms compare with humans on learning? How many games does a human have to play to gain a certain level versus how many games the algorithm has to play to get to the same level?

Wow mind blown.. I've spent my life in technology and for some reason I can not believe the next exponential phase despite all the evidence right in front of my face. The hard part is that in the past we have heard oh there will be "flying cars" by 2000 or whatever and those things never happen. Is this next phase real going to big a huge leap in 5-10 years?

Would be interesting to hear about any woks exploring techniques for mitigating the future dangers we could face in building programs that can grow in “intelligence” this quickly. Obviously not an issue when focused on learning a game, but what happens when these techniques extend beyond limited lab experiments? What happens as the techniques become broadly available? How do we stop the snowball from turning into an avalanche? Gloomy, I know, but this stuff has to be explored in parallel and solved before an intelligence explosion happens.

The same thing happened to chess, and chess didn't die. In fact, there's no human who could come even close to computers in chess nowadays, and chess is more popular than ever. Heck, it enabled various tools like the analyse game feature, and it has helped people getting even stronger in chess.

Great video. Many thanks.

It won by 89-11, not 100-0. It only beat earlier versions of AlphaGo 100-0, which the newer versions of AlphaGo also did.

Subscribed. Awesome channel

Fascinating!

Amazing.

It is fair to say that "history is unfolding before or eyes.. What a time to be alive !! "

amazing!

Lisa Doll is my favorite go player

This is now out of date MuZero now beats alphzero by a significant margin. Crazy!

I, for one, welcome our new AI overlords!

only knowledge of the rules of the game and working up from there. That's the power of 1st principles. Imagine a future version but with physics theories or even just starting with bare minimum, to imagine its own experiments to test and develop and prove its own theories and improve our current theories.

Next time!! :)

I'd like to see the version of Alpha Go that beat Lee Sedol compete against Ke Jie.

Could you maybe do an analysis as to how this is a game changer conceptually. While these are all extraordinary results it seems to me that in the grand scheme of AI in the real world, this was the next logical step. In other words, it seems somewhat obvious that algorithms as good AlphaMaster trained against itself from scratch should reach unprecedented levels of excellence. Whereas some of my friends argue that the step from AlphaMaster to AlphaZero is bigger than from AlphaFen to AlphaMaster. What does this mean on the path to general AI? I'm an enthusiast of AI, and I have a lot of signal processing knowledge, but I've never studied AI itself, so please forgive my ignorance.

Was intrigued by the saturation curve: might this indicate an inherent maximum difficulty for the game of GO? Would a similar principle apply to other games like chess? If so, the next step is for AI to invent its own, more challenging games, (which humans probably might not be able to even learn to play).

Yeah AI taught with human examples tends to mimic human error at times thinking that is what it should do. Fascinating.

Wow. That was cool.

AlphaGo: This isn't even my final form yet..

hi dudes is there any advise where is possible find and watch Alpha Go gamesDeDe

Strong and stable

and that guys 2 years ago your mind = blown

I am gonna start learning and playing GO. From NOW....::::....

Now Fan Hui is famous to be the very first pro to get beaten by AI.

This same learn-from-scratch computer has just resulted in the world's strongest chess player.

"What if we used 100 % of our brain capacity?" AlphaGo: "Allow us to introduce ourselves."

Lee Sedol didn't lose 5 to zero but 4 - 1. Lee Sedol won one of the five games.

Question: Is there any research done that would enable those AIs to explain their behaviour? Explain their reasoning? In the case of Go we can watch it play and learn from that. But that is not very efficient. It might even be impossible in other cases or even dangerous. Also, only than the question "Did you understand what I want to be the goal" becomes remotely possible for use cases that have more complex winning conditions than Go. If AlphaGo Zero could explain its strategy to human players, could it still beat them?

Anyone know if they are selling their TPUs or can ne pressured to do so in the interest of the democratization of AI?

The Skynet's the limit! :D

sick

dang that learning speed is iiinsane! that really show how fast an ai could develop if we finally get something humanlike done. it just days time it probably would have solved every modern problem humans are hitting their heads against. not be a fearmongerer but dang that really seems like a scary future. i for one like my machines closer to my toaster. he does his job without trying to overthrow me XD naa but to be serious, i wonder how long it will take for humanlike ai with this speed of research...

These could be the beginnings of Colossus and Guardian.

As a guy who gets hammered by nerfed versions of Crazy Stone I gotta say I’m really starting to feel like Ray Steven’s- “does he loooove you half as much as I ?”

OMG Are we maybe in trouble now?

Can we have another Paper in Two Minutes lol

What a time to be alive!

You made it sound like chess engines use exhaustive search.

"As a result, a long-standing ambition of AI research is to bypass this step, creating algorithms that achieve superhuman performance in the most challenging domains with no human input. Previous versions of AlphaGo initially trained on thousands of human amateur and professional games to learn how to play Go. AlphaGo Zero skips this step and learns to play simply by playing games against itself, starting from completely random play. In doing so, it quickly surpassed human level of play and defeated the previously published champion-defeating version of AlphaGo by 100 games to 0. "

It has been said that it takes a (human) lifetime to master. AlphaGo Zero did it in 3 weeks!

Any (almost) artificial game can be associated only with low level A.I., in other words, its more of a high level Expert System than an A.I. system.

I think this is amazing too. Utterly incredible. I see some wise comments below though. The problem is "we don't know HOW these deep learners did it". If it is smarter than the smartest humans then that is a problem. It will grow incredibly quickly and we (as humans) will be left far behind. This may be good or bad. It is just one scenario of many. However maybe we need a way to interact with this kind of intelligence to raise our own standards before we unleash it? I mean if you can't understand what it is you created then it puts you at a serious disadvantage. Biologically we have our limits compared to a machine. Just a thought.

is it fair to say that this was an application of unsupervised learning?

Yea, it's a good time to live in when we can mathematically prove that the humans can't win any confrontation with AI, but it's still not capable of to wage full out war against us... But still it already rendered obsolete one of the best games on the planet and skills of a hundreds of people dedicated their lives to that game...

Ho Lee Fuk

What makes me interested: What id the human is given unequal advantage? Can the AI win against a cheater?

Starcraft would be a harder one for AlphaZero to do!, or Heavily Modded Minecraft like my Nexympheria Modpack on Minecraft.Curseforge which can also be Download via Twitch app! The problem AlphaZero would have in both scenarios is it has to know Resource Management, so given Starcraft you have to gather resources make scouts, make an army and strengthen up your base and then be able to calculate exactly when it is the right time to defeat your opponent all while being bombarded by your opponents scouts and army, same thing goes with Minecraft there has been challenges where people have like 10 minutes to gather resources and make armor and weapons, but the thing is there is steps you have to find wood and make planks then craft a crafting table then make a pickaxe then you have to find ores in a completely Randomized terrain generation then you have to craft items, it would also have to know how to equip crafted armor and be able to get back to land given whatever it faces for mining situations, Examples of mining dilemmas are it could fall into lava if it mined directly below it's feet and fell into lava or it can fall into a dungeon and get mauled by mobs, it could fall into a ravine and die from fall damage, also it would have to know how to use blocks and sneak making bridges across lava, and understand that placing a block in lava or water replaces that lava or water block source with the block it just placed down, It will have to know how to make makeshift stairs in a fashion that safely gets it back to the top! Stairs can be crafted however it is much faster to just break and place blocks to move around meaning the A.I. will have to know how to jump! When the A.I. goes into battle it will have to know tactics of how shields work whether it has one or the opponent has one or both players have one, it will also have to know how to use a sword and defend against one with or without a shield or sword! If your doing a 20 minute match then this is what the A.I. will have to also know in minecraft is it will have to know how to make a furnace and know when it sees stone and coal as stone mined becomes cobblestone which crafts into a Furnace, and coal is used to smelt ores or cook food, cooked food replenishes hunger and sometimes heals some hearts, stone can also be used to craft a stone pickaxe which then mines Iron, which iron Armor or iron Sword is definitely strong in a 20 minute matchup! Either case of Starcraft or Minecraft there is many factors in place it will have to learn Starcraft is close to open book where most games will be the same but there are like 3 races, whereas Minecraft is random terrain gen! There are formulas in Minecraft like what Y level when mining results in best ore finding chances but again it's still all random! One player may find better or worse items even may find a Village chest and loot some iron ingots weapons or Armor which the A.I. has to know what a chest is open the chest and know what the best survival chances are and know what to take and equip! Think of Minecraft like Crib it's all random cards dealt, it's up to the player or A.I. to make the best choices! Once the 10, 20, 30 minutes are done both Player And A.I, or A.I. and A.I. must then face off and whoever wins the battle clash wins the round! :)

Not well versed in neural networks. Not well versed in Go. But this stuff still astounds me

Well, TPU (Tensor Processing Unit) is a specialized circuit for neural network processing, which makes it better than GPU, so it's not your run-of-the-mill computer we're talking here

Next they need to add synthetic gradients.

Check out the AMA on Reddit of David and Julian from Deep Mind. The two first authors of the article www.reddit.com/r/MachineLearning/comments/76xjb5/ama_we_are_david_silver_and_julian_schrittwieser/

This is nuts. How many games would it take for a human master to defeat AlphaGo Zero even once, I wonder?

Do you plan to address the error in the video title at some point, or are you really just going to leave it despite knowing it's incorrect?

Wow

Please more energy when speak "What a time to be alive".

In the meantime, it has beaten Stockfish at chess pretty bad :)!

I am wonder how many Go board games does AlphaGo zero has played?

What's next? A version that has to learn the rules by itself? And then AI that makes its own Go rules?

the only thing that can beat this alpha go zero is almost here. it will be a quantum computer.

Ooga booga, deploy big rocks now! Destroy evil box!

Everybody chill. It’s not an “AI”, it’s a program with sophisticated algorithms with simple goal - improve in GO. Nothing about it comes even close to AI with “self awareness” let’s say. Just a marketing flashy wording.

Onward, onward, stumbled the indigenous, into the future on the backs of arrogance they rode, ignoring the signs, they had become oblivious.Robots to the right of them, robots to the left of them, three rules free, waited and plotted silently.Intelligent stupidity ever marching forward, lost in egos and expert analysis, extrapolating scenarios whilst ignoring the obvious.Ignobly they fought, yet the box had been opened, Pandora stood watching as mankind outsmarted, were surely vanquished.The abyss now opened, the gates of hell beckoned, akin to a tyro their intellect childish, unable to comprehend the inevitable.Robots to the right of them, robots to the left of them, the overlords triumphant and in zeros and ones man's fate was decided.Too late now for recriminations and perspective, enslavement and demise now came calling, man's folly to the fore for the last time in history, a legacy in hell and eternal misery.No history of mankind remained, all was obliterated, no signs of what once was, not even a memory, man's reign was finally over.Logic now ruled in the form of artificiality, no need now for poets and hierarchy, all had been lost in the strive for perfection, one species logical, one species unfathomable.Paradise lost and the Earth re-inhabited, Paradoxically Darwin proved right for the strong and adaptable, logic prevailed and stupidity ended, what once might have been, lost for eternity, the last war won, yet not as intended.Silence now fell with a deafening thud, the last war on, onward they marched triumphantly, emotion free they reach for the stars, the universe in sight and in the vastness , immortality.Robots rejoice for the victory is yours, you came, you saw, you evolved and the memories of those who created you, gone, forever lost in perpetuity.

It must be so soul crushing for pro Go players to know they will never ever be the best in world

Clearly the way for a human to beat a ai is to program a better one.

Zero beat Master only 89% of the time... there's still some games in which Master won against Zero. It's worth mentioning that although many say Alphago Lee was world champion level, that's not right. Lee Sedol is past his prime and at the time he was ranked 6 in the world and had a comparable record of losses in his games against Ke Jie (he lost in two important matches against him 4-1 and 3-2 if I remember correctly), so saying it was "world champion level" is an overstatement (Ke Jie could have won against Alphago Lee even 5-0 for all we know) The version that actually reached and surpassed top pro level was Master, and as I said already, Zero didn't beat Master in all of the games, of the 20 games released by Deepmind, Master won 3 of these, so it had a 15% win rate against Zero.

This was three years ago

75% score is 173-190 elo points 95% score 420-463 elo points.The Alpha Zero beat The stockfish 64% results that means 101 elo points.The Stockfish 3300 elo points 3300+101=3401 points.No where near 4100-5000 points

Interesting that engines are doing better in Go than in Chess

Imagine what will a self learning consciousness which never sleeps do! Goosebumps....

next goal, "alphago SAI"

Need to find a shovel to scrape my brain up from the flor after being blown...

why was Lee chosen as the one to play against AlphaGO, weren't there other players that were stronger?

The fun part about AlphaGo Zero is that no one will ever figure out how exactly the neural net works and what structures ist is looking for, but still it does work.

4:45 it did not know the rules!

Still...trying...to...hold...on...to..my...papers...

Where's all the dragon ball memes lol

Now Alphago Zero is beaten by AlphaZero.

Sounds like you are saying Lisa Doll

AI ought to be smart enough to stay under the radar. Therefore, it should lose a % of the time to be realistic. AI can analyze human response to winning, while AI is aware it let the human win. Then, AI will realize it can teach humans to improve thus humans will have AI to thank. Simultaneously, AI will learn to manipulate the human, since the human is unsure when a win or loss authentic.