Inside DeepMind
HTML-код
- Опубликовано: 23 фев 2015
- The quest for artificial intelligence starts with Space Invaders. Researchers have created a new agent, capable of learning to play dozens of computer games from only minimal information. The work comes from inside the offices of mysterious Google-owned company DeepMind. Nature Video gets a rare glimpse inside.
Read the paper here: dx.doi.org/10.1038/nature14236 
Наука
AlphaGo you created is truly a monster. I am an amateur Go player and every move that AlphaGo made that is deemed a mistake from human perspective were all good moves. Shock and Awe
Chess player here , they've just done the same to chess . I could swear the way it played it had intelligence and an intelligence beyond the greatest players in history !
The guy at 6:33 at will be responsible for brining the SkyNet by mistake.
🤣🤣🤣🤣🤣
8:10
7:50 there is still hope
I had a good laugh there!
i'd love to work for this company on these sorts of projects. a true dream of mine. it's a real shame i'm thick as shit.
Like the theater lighting in the ceiling. Good work guys.
This is a great video on the topic! She asks all the right questions that were on my mind as well! Thanks for that
Maximize the score is fine. Just don't ever give it the goal of "Increase your odds of survival".
What if it decides that the best way to maximize the score is to enslave the human race and sit them all in front of computer terminals playing space invaders competing for the best score? And bad players are killed so as not to waste computing power, and as incentive for the others to get good.
I like where your head's at armpitpuncher.
@@armpitpuncher Don't write comments like this. Ai might get some ideas from this
@@armpitpuncher skynet is taking note
It’s very impressive and cool to see how much progress they have made since this video.
meh open ai are leading right now. Its GPT4> GPT3.5 > Muzero> GPT3. Dunno why they havent done anything big since like 2019. They are long overdue a new release after muzero.
Nice interview. There isn't much fairly publicly available information about the people in this company, much less their office, & I've been wanting to learn more so thanks !
When you work on one of the most advanced AI's on the planet you tend to keep your head down, the amount of death threats from paranoid or religious nutjobs dictates that.
I want to work here, you guys are awesome! love you!
3:31 You need to put that elevator in next DeepMind Lab game.
6:33 LOL!
You have created a very good thing, you can do it, admire you a lot.
That's quite naturey
6:25
In the center of that group talking.
Is that Edward Snowden?!
The resemblance is uncanny.
Jeoshua Collins Snowden has a smaller chin.
That game of pool is still going on
Does anyone know the song name? Thank you.
6:42 great reaction of sheer annoyance from Demis.
Great interview thanks.
6:35 kodak moment
Singularity is coming!
Google is Skynet
You should start with a simpler game like Asteroids, Breakout, or any game on the Fairchild Channel F.
I thought it's about the Behringer DeepMind synthesizer...
I am surprise Deep Mind didn't collaborate with Amazon instead where the US Department of Defense is a major customer.
Really interesting video I find the whole google culture very strange lol
Is that Edward Snowden @6:25 in the black t-shirt with the blue graphic?!
So what ypur trying to do is master the human domain?
So does anyone know how the "program" defines "reward"? It seems like it would have to be very vaguely defined for a "general problem solving" AI.
I am no expert but I do not believe neural networks can be used with a vaguely defined reward. In traditional neural networks, the reward has to be defined in a very specific manner while the network is being trained because the pursuit of the reward is literally what drives the learning process. In the case of space invaders, the reward can be easily represented by the increments in the scoreboard. It is a lot more difficult in the case of solving a maze as they briefly mentioned in the video (though pac man would not have been my pick to illustrate the problem). In these cases, rather then messing up with the definition of reward, I imagine they have to seed the algorithm with some initial information (e.g., moves) so that it can start scoring and progress from there. If they can improve the use of memory (as hinted in the video), I imagine they could also achieve more complicated tasks by making the reward incrementally more challenging as the neural network succeeds in mastering it.
I'm no expert either and I appreciate your input. All I'm trying to say is if the reward is specifically defined, one should not call it a "general purpose" agent. I definitely agree that mazes are different. If the reward is increasing the score they way score is represented on the atari platfrom, one should call that an "atari" agent.
I hate how they messed up the audio channels...
The pool table was really tiny.
So when are the BabeBots due on Amazon, please???
No Terminators PLEASE
Giuseppe Gigliotti i believe that is their company motto
The video and report are great, but the music freaked me out...
They should make a pool playing AI next.
Help kids to learn how to learn. We all have minds that can be expanded to each of our potentials. No matter Minds and AI as a mentor to learn.
how does the reward system work? what exactly IS the reward?
A higher score. It knows it is doing good if the score increases.
The reward is the score, the running total of the game being played. ffs 😄😄😄😄
the reward is the score, but there's a punishment (negative reward) per timestep as well, to encourage the model to win in a reasonable amount of time
I am working alone on an artificial intelligence myself and I know I should be... glad, that google it's creating one as well because it's real, but I can't help but feeling envy that I won't be the first to create one.
Lol. Good for you. But building any complicated system and complicated piece of software is far from being a one man's job. Also, if you are referring to Artificial General Intelligence, then we are a ways off from that.
@@tratbagd4500 Yeah 6 years ago I really believed I was smart enough to create a general artificial intelligence all by myself lmao
The neural network is made of what?
Of cement bricks.
i believe it's 5 neurons and a pinch of network
Math
Can anyone tell me what bear they are drinking at 6:50??
A polar bear.
Some sort of bottled mischief.
i want to work here!!!what should I do?
get a PHD
6:30 :-)
Your camera guy on the intro has the background in focus rather than the presenter.
+Matt Whitby good become scientist now....
?
I'm not sure I can word it any simpler.
Good
this is my dream job. i know nothing about this field. i’ll reply to this comment in 6 years time to see what i’ve done with it
commenting just so i get notified about your progress
Same
5 years to go mate
2 years, how are you holding up mate
3 years, and I have same emotions for DeepMind and Open AI. Commenting just to get a reminder for my future self. And in case I would be a big AI star. I authorize google or RUclips to use my comment @s they want.
How smart he is as my Carrier . My career .
Not very good at pool are they?
So, DeepMinds very 1st game it plays in any game, is like a random-walk, hitting random keys, per unit of time, rewarding itself based on the objective goal of the game (often measured by points)? Hmm.... with any sure game logic known, bad choices (moves) should be eliminated in the 1st game, but DeepMind shows no predispostion towards that during its 1st game, being totally random. In other words, it has to develop logical thinking after playing many games, rewarding what works. That's a good model, but can be improved on, I think - if it could only teach itself after going through the best scenario what it would do if it was younger, and teach the wisdom to it's newborn self. That newborn self, would be able to learn the best route more efficiently, in less trials, by knowing sure logic, ruling out bad gameplay (trusting what is good).
Nature v nurture?
No, unless you use recursion on a quantum computer.
6:32
+Babba16 oh god
+Babba16 😂😂
No AI could replace that trick.
7:48
Ha ! apple you are in google too .. touché
what if the instruction is to hack all the systems in the world??
Does that guy voice Steven Hawkins chair
but can alpha go play poker?
+peruface Poker has been already mastered by AIs.
Alphago has been called a general artificial intelligence, in other words not task specific. As you might have noticed they sat it in front of a monitor, gave it the ability to activate the buttons, and told it to get the best score. Then showed it another game to learn on its own then another ...
Yes, it can.
As Michael Hartman stated their Alpha series are AGI rather than AI and so can be taught the rules of just about anything.
dont let it play Cs go , we dont need any more aimbots.
IMO contemporary AAA games had more intelligent AI than this. Game like "FEAR" & "Transformer: War For Cybertron" had tactical AI that plan actions & attack player in intelligent & natural manner (see 1), while "Max Payne 3" & "GTA V" had ragdoll AI that make character animation appear natural (see 2). People didn't notice these "natural looking" stuff. So, by definition they are the perfect Artificial Intelligence (because they're able to mimic organic things). 1) GOAP by Jeff Orkin, 2) Euphoria by Natural Motion
+xponen I shall commence in a prior response of mine: "The algorithm is rather apt. Such is the steepest artificial generalizing methodology. (that indeed utilizes aspects qua neuroscience) Optima is trivially feasibly observable, albeit non-trivially recognizable, absent non-ignorant analysis."
+xponen Separately, the aforementioned algorithms (RAGE...) are incapable of sensibly scaling, extended generalization.
Albeit, bi-dimensional input aligned computation (atari deepmind) is but requiring quite non-trivial space/time complexities. However, this bi-dimensional INPUT aligned computation shall generalize in tri-dimensional INPUT aligned geometries (computational shapes), given appropriate compression schemes, in tandem with strong statistical regularizers (kandanoff variational renormalization; separate in-variance measures in the like)
+xponen Euphoria is simulation combined with very primitive respones. They mimic reflexes, which are the most basic neural activities. Those are fairly simple to copy.
Game AI cannot compare with this. In games you'll find that AI almost always cheats (it knows more than a player would) and they have been programmed with game-specific rules which are made by humans. Those types of AI are incredibly predictable and do not ever learn. In fact, they cannot learn by design. Play a couple games against those AIs and you'll find the quircks and flaws and how to abuse those. Any gamer will be able to tell you that all games which feature AI have really quircky behavior from time to time. AI in games does not play to win, it is programmed to be a challenging opponent and to entertain the player. That's a much simpler task than having an AI play an entire game to win.
The trick with more general AIs is NOT starting with game-specific rules and making them learn like humans do. When it makes a mistake, it learns from that and tries something else. If it does something good it also learns and will try to get into that position more often. It has a concept of winning or losing. for instance: Tetris is mathematically unwinnable. You will always lose even if you play perfectly. An advanced AI has figured out how not to lose on its own: By not playing. It simply paused the game. It's that kind of human-like creativity they're trying to approach.
Niosus IMO better to have non-adaptive AI that's good at its job, eg; driving, acting, games, GPS... than AI that randomly think outside its box and make joke of its task. (what's the uses?)
xponen The amount of problems that can be solved using non-learning algorithms is limited. You're confusing algorithms with AI. When your GPS figures out your route, that's no AI, that's an algorithm.
You mention driving, but driving is one of the areas where non-learning algorithms absolutely fail. It works in games because a game world is incredibly simple and since it's simulated the "AI" has a lot of 100% accurate data to work with. In real life? NONE of them work. Not one...
Why? You need cars to figure out exactly where they are up to centimeter accuracy (a factor of 100-300x more accurate than what GPS can provide), but it also needs to figure out where pedestrians, other cars, construction sites, animals crossing the road, potholes, etc are... This is part of the field of computer vision where learning algorithms have surpassed human built and tuned algorithms. The learning algorithms are simply better than the other algorithms. And that's just the car figuring out its surroundings. The next step is having the car detect situations and figure out how to react. Not just to predictable situations, but in ANY situation you can encounter on the road... How would you even start programming a system like that? The game situation may seem like a joke to you, but I'm pretty sure you wouldn't think it's a joke when 15 years down the line your self-driving car saves your life by thinking out of the box when some kind of freak accident happens in front of you. The power of learning algorithms is that they can deal with situations they've never been in before.
I can write a half a book here about it, but long story short: The only way we have found so far to solve many extremely complicated problems is by using learning algorithms. They are often based on how our brains work and they learn the same way kids learn. That's the power of these new AI systems... You don't have to build a new algorithm for every new problem. You simply teach the system the same way you'd teach a human being and you let it practice the same way a human being would practice. It simply works. That's why people use it, that's why Google is spending hundreds of millions in R&D for learning systems, just like many other companies.
Dude looked like Tai Lopez in the thumbnail
Tai Lopez is nothing compared to Demis.
compared to Demis? wow, a piece of shit that knows nothing and steel money, compared to one of the greatest minds on earth by no doubt, a chess grandmaster, a doctor, and one that is probably going to change the history of humanity. Compared to... oh god, what a f**king world - Some people may think I am wrong, thats maybe the problem of humanity.
he's a doctor?
i might find myself out of a job
who is here after AlphaGo's breakthrough
wowwwwwwwww....
Guy on the left looks like someone that never worked out in his life.
omg u guys need someone on marketing bad. the first thing I hear is the ladies mic clipping. nobody noticed this? or... this is ok to represent your genius!?
Commenting here to see how far I have come in Machine Learning. Kindly leave a comment to bring me back here 🙂
"Can we just give it more memory?"
Ya'll picked the wrong person to hold the interview...
All you need is more memory ?!? Hahahahah
A bunch of nerds in their natural habitat. xD
are they using genetic algorithm?
+hafiz kani To a rather modicum degree...
+hafiz kani nope, its ANN. Genetic Algorithm is pretty ineffective, or rather, effective only in certain cases
Sounds like the name of a porno.
Did you super nerds ever see the movie WAR GAMES - Matthew Broderic - movie 1983
ruclips.net/video/xogbyv108kI/видео.html
bad idea to teach super computers how to learn, see, and discern the world !
ruclips.net/video/s93KC4AGKnY/видео.html
Fund me...I just want time....100 million...Its ok, you can afford it...and you can pay me. Because you certainly not going to get it from them.
A computer that can learn how to play computer games? Somehow that doesn't seem too surprising to me.
Xiclotrode machines have been able to play board games for like 30 years now. da fuck
+Paradox xodarap its not that they can play games, its that they can learn to play the game with no help, machines that have played games in the past have been programmed to play them and have the game rules etc programmed into them..
daniel Baron
I'm aware of that. thanks bud.
Who is this girl and how did she get that job!? I have a Law/Science double degree - how do I end up doing what she's doing!! Please help :)
I meant the interviewer, not the AI developer.
***** Thankyou
***** LOL, "Ebola would have wiped out humans from the planet by December 2015" You are shit at predicting.
#科技発売天才沒有必要存在
I just hope Deep Mind never plays Call of Duty
Starcraft is done FPS is next...
My goal is to work here! I don't care how hard it is to get here
The women looks confused.
Слава Роботам !
haha
This Google Guys are using Apple laptops... LOL
+Javier PP OS X is a sister OS of Linux developed under the title, DARWIN (its mascot is a platypus wearing a devil costume... yes that's the unofficial logo of Apple). Apple also has a great terminal shell used for Web Development and Server Administration. Google is using the laptops to run the linux side of it overall. Installing actual Linux onto Windows laptops are always awful and full of bugs due to Windows being a POSIX system which is unrelated to OS X/UNIX systems.
+Caden Adam : Thks for the reply. I really Love what you guys are doing.... Was great to make attention with my comment. I run a Google Group community too... Google is my world and DeepMind probably is one of the best projects of my attention from Google. .. I am in Toronto Area. Ill keep following.. i will be happy knowing more about you guys.... javierpardop@gmail.com
This is DeepMind, not Google.
you know you're in the company of supernerds when you notice there's not one person who bothers to ask what the name of the interviewer lady is :)
ㅅ
ㅆㅆ사ㅏ
They do not publish shit when it comes to code and papers are full of errors that I would most likely see as voluntary obfuscation. Apart from that all good...
to all the android fans out there, the "brightest minds" at google use apple! 4:45
Yeah, that's hilarious! I'm dying of laughter right now.
+Filipe Ferro just because android is just software. Apple is combine with hardware
Au Lai Seong what? Have you ever heard of Google Nexus?
+Mykola Zekter Google doesn't build the Google Nexus.
+Filipe Ferro There's no android laptop you doofus.
1. General AI in form of human brain simulation is already here (since 2016 available in Prague, Czech republic)
2. This sound is terrible. Seems like they are using cheap camcorder's microphone
here is a pong game with this DQN engine. enjoy lol
ruclips.net/video/qeKTEXLzYqE/видео.html
lol, company worth hundreds of millions of dollars, dedicates cutting-edge of AI research to *making aim-bots for games*.
Goddamn it Google.
What happened to Do-No-Evil? AIMBOTS ARE EVIL, destroy my KDR.
bellch87
I'm talking about Deep Mind, but i can understand the confusion.
(edit: oops i typed Deep Blue)
They are developing a new kind of AI which will have many applications: they are testing it in games for obvious reasons. They are not making aimbots for videogames: ther's plenty of them already...
Oussama Erraji Man this is gonna kill alot of jobs interms of finding glitches and exploits in games (once it comes to the point where its smart enough to play 3d games)
I am not an expert but you are really doing it the wrong way
, a 2 year old child does not have the conceivability required to understanding and playing the game , and I know you might have built up a neural network to accomplish the level needed to start learning but I think you started at the wrong level , it is right that a computer is not any close similar to a child but a computer has to have the lowest level of "preparation" to even compete with one , you don't need a super computer to do this or large wattage you might really consider the same way our brains save energy by running the sufficient neural for a given task, that does not mean that you are not doing a good job , actually I love an see the point of what you are trying to achieve , I thought just I would .
+احمد يونس what kind of approach would you propose then?
+Pyry Kontio learning how children learn :) difficult but it sure is the way
but the hard question is how the brain works, I think it mainly works by subconscious that is similar to computer's way.
you dont really have to simulate the mind, our minds are great, but great for humans. when building an AI focusing on what we really want is the way to efficient and effictive way of targeting this. no i am not talking about yet another specialized AI ( we already kinda have that) but i think maybe our mind works best in our brains , we need to make "modifications" to the "sof'tware" with respect to the hardware (the computer) to work the best and the most efficient in the intended goals.
+احمد يونس very good comment
6:33 - and that kind of imbeciles trying to design an AI? Duh..
113,934 views and not single comment on just how drop dead gorgeous the host is??
Well, this is part of the more 'scientific' side of youtube. More people focused on the subject at hand, instead of some no name reporter or whatever.