New DeepMind AI Beats AlphaGo 100-0 | Two Minute Papers #201

preview_player
Показать описание
The AlphaGo Zero paper "Mastering the Game of Go without Human Knowledge" is available here:

Our Patreon page with the details:

We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Andrew Melnychuk, Brian Gilman, Christoph Jadanowski, Dave Rushton-Smith, Dennis Abts, Eric Haddad, Esa Turkulainen, Evan Breznyik, Kaben Gabriel Nanlohy, Malek Cellier, Michael Albrecht, Michael Jensen, Michael Orenstein, Steef, Steve Messina, Sunil Kim, Torsten Reil.

Two Minute Papers Merch:

Photo credits:
Watson - AP Photo/Jeopardy Productions, Inc.

Go board image credits (all CC BY 2.0):

Go ratings:

Károly Zsolnai-Fehér's links:
Рекомендации по теме
Комментарии
Автор

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

magicmulder
Автор

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

HummingbirdCyborg
Автор

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

dinkir
Автор

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

lollollol
Автор

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

UnboxingSve
Автор

Elo Rating over 5000? Come on guys let's get it OVER

liono
Автор

Who came here after watching the AlphaGo documentary?

rockyard
Автор

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

timonix
Автор

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

SiavashFahimi
Автор

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

akanippy
Автор

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?

lance
Автор

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.

Goryus
Автор

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.

iliantrifonov
Автор

better without being taught human flaws, eh?

trapOrdoom
Автор

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

klasop
Автор

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.

dextrian
Автор

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.

CristanMeijer
Автор

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.

AtlasReburdened
Автор

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

The-Dom
Автор

This is now out of date MuZero now beats alphzero by a significant margin.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

benjaminblack