OpenAI Built Gaming Bots That Can Work As a Team With Inhuman Precision (qz.com) 97
OpenAI said on Monday that its newest AI bots can hold their own as a team of five against human gamers at Dota 2, a multiplayer game popular in e-sports for its complexity and necessity for teamwork. The AI research lab is looking to take the bots to Dota 2 championship matches in August to compete against the pros. From a report: Dota 2 is a challenging game for AI to master simply because of the amount of decisions that the players have to juggle. While chess can end in fewer than 40 moves, and Go fewer than 150, OpenAI's Dota 2 bots make 20,000 moves over the course of a 45 minute game. While OpenAI showed last year that the bots could go one on one against a human professional in a curated snippet of the game, the company wasn't entirely sure that they could scale up to five against five.
But the research team doesn't credit this breakthrough to a new technique or a lightbulb moment, rather a simple idea. "As long as the AI can explore, it will learn, given enough time," Greg Brockman, OpenAI's chief technology officer, told Quartz. The bots learn from self-play, meaning two bots playing each other and learning from each side's successes and failures. By using a huge stack of 256 graphics processing units (GPUs) with 128,000 processing cores, the researchers were able to speed up the AI's gameplay so that they learned from the equivalent of 180 years of gameplay for every day it trained.
But the research team doesn't credit this breakthrough to a new technique or a lightbulb moment, rather a simple idea. "As long as the AI can explore, it will learn, given enough time," Greg Brockman, OpenAI's chief technology officer, told Quartz. The bots learn from self-play, meaning two bots playing each other and learning from each side's successes and failures. By using a huge stack of 256 graphics processing units (GPUs) with 128,000 processing cores, the researchers were able to speed up the AI's gameplay so that they learned from the equivalent of 180 years of gameplay for every day it trained.
Re:Training is not AI (Score:5, Insightful)
Re: (Score:2)
Slaves or Masters or Mind Children or Friends etc? (Score:2)
See Hans Moravec's informed speculations like his book "Mind Children": https://www.goodreads.com/book... [goodreads.com]
Or going beyond that to the nature of consciousness and reality:
http://www.frc.ri.cmu.edu/~hpm... [cmu.edu]
And see also Vernor Vinge's various writings on a "Singularity".
That said, hedging our bets by making the world a happier and healthier and more resilient place for everyone right now before a singularity is probably not a bad idea given our trajectory out of any singularity may have a lot to do with out pat
Re: (Score:1)
Re: (Score:1)
Just remember... whenever a computer can do it, it's not A.I. any more.
So when we have a computer with a robotic body that passes for human and is better than humans in every way... folks will say, "Oh that's not A.I."
The amygdala isn't intelligent.
The cerebellum isn't intelligent.
No component of the brain is really intelligent.
I actually think it's plausible that intelligence might emerge from multi player collectives where each "player" isn't intelligent individually.
Re: (Score:2)
Wrong. The intellectual abilities that an infant can exert allow him to give a meaning to his experiences without any required training at all. We train our children because we want them to learn our ways to communicate and to behave in society, not because they cannot learn on their own language and behaviour. No computer program can do this.
So you meant that if you just drop a new born baby in a jungle, the baby will learn how to walk, eat, and talk by itself without any help/train (from humans or animals)? Intelligent abilities of a baby are simply capabilities and speed of learning due to its physique which is NOT ready to do anything by its own, regardless how intelligent it is.
Even though training and learning are 2 separated things, the learning is directly related to training for a baby which is pretty much like a computer. If you are ta
Re: (Score:1)
An intelligence is self emerging, it knows what to do without training. Having to train something means it is not intelligence, so by extension not an artificial intelligence.
From TFA: The bots learn from self-play ...
No human told the bot what to do. It "trained" by playing against itself.
Re: (Score:1)
They must have some kind of guidance, though - even something as simple as "killing good, dying bad, these are the buttons you can push."
The rest though, yeah - self-training. I'm still not sure that it's true "intelligence" though... But what *is* intelligence, exactly?
I dunno.
Re: (Score:2)
I'm still not sure that it's true "intelligence" though...
Of course not, except in a very narrow sense. AI is a field of research to develop machine intelligence. We are making progress, but it will be a long journey.
But what *is* intelligence, exactly?
Intelligence is the ability to formulate an effective initial response to a novel situation.
The wording here is important. It is an "ability" not a mechanism. A system that consistently behaves intelligently is intelligent. The internal mechanism is irrelevant. It is an ability to "formulate" a plan, not a physical ability to act on the plan. I
Re: (Score:1)
Even highly intelligent humans often use random trial and error when confronted with totally new situations.
They also do stupid things like seeing patterns that don't exist or extending from prior situations they think are similar but which are not.
The human brain is delightfully buggy, subject to framing errors, physical defects, and can't even detect when it's broken most the time.
Re: (Score:2)
From TFA: The bots learn from self-play ...
The equivalent of 180 years of self-play must have left them deaf as an adder.
(OK, I admit, I'm really a bot that learned how to post jokes through self-play)
Re: (Score:3)
Re: (Score:3)
Re: (Score:2)
The latest distinction between machine learning and AI that I saw ...
Where did you see that? ML is a proper subset of AI. Period.
Anything that is ML is also AI. But there are subfields of AI, such as min-max and alpha-beta pruning, that are not ML.
Re: (Score:2)
Re: Training is not AI (Score:2)
The end of the line for RTS/MOBA (Score:1)
If the only thing that matters is APM, there's no point in playing.
Re: (Score:1)
Weren't you going to retire off your trolls? Shit talk them and annoy them to drive traffic to your content? I would think this pleases you everything is going according to plan,
You're stupid to realize that creimer spent a year trolling your sorry ass.
Re: (Score:1)
Re: (Score:3)
That makes sense, since Quake [wikia.com] is indeed an Inhuman [wikia.com].
One step closer to doomsday (Score:3)
Re: (Score:3)
Why does this conjure up images of hordes of inhumanely fast robots swarming cities and taking out citizens and soldiers with ease... How long until there forms an upperclass completely immune to revolution or the conscience of its human military?
I think that's quite a ways off. But to Godwin the question, how many hardcore Nazis did it take to run Nazi Germany? Yesterday it might have taken 50% or 10% of the population to support you. Tomorrow it could be 1% or 0.1% because they keep tabs on everyone else.
Re: (Score:2)
"With how easy it is to bribe, cajole, wheedle, or coerce people into doing things,"
You can't judge everyone by the French.
Re: (Score:2)
Yesterday I visited Wikipedia to scope out the mere-exposure effect.
Re: (Score:2)
The Poles were instrumental in winning WWII, from code-breaking to piloting fighters in the Battle of Britain.
Re: (Score:2)
Why does this conjure up images of hordes of inhumanely fast robots swarming cities and taking out citizens and soldiers with ease
Because science fiction about AI that was somewhat useful in some cases but didn't quite solve as many problems as was hoped and yet didn't go out of control doesn't sell nearly as well.
How long until there forms an upperclass completely immune to revolution or the conscience of its human military?
As soon as someone figures out an absolutely foolproof way to identify "upperclass" to an AI. Military history is a millennia old arms race. New weapons are inevitably met with counters to those weapons. If someone makes an autonomous weapon that only targets certain things, someone else will figure out what it's been tar
Re: (Score:1)
Well, please take consolation in the fact that as they are doing this, folks who are being shot will proclaim "Those aren't A.I." as they die.
A.I. is really a potential extinction level threat and people don't take it seriously enough.
At a minimum any A.I. research should have analog power consumption indicators, remote observation, and a physical power connection that can be broken easily (or even one where active steps must be taken to maintain it).
We don't do that in many cases.
Re: (Score:2)
Pouring all your fears into one basket?
When I grew up, the rusty old H-bomb featured as the new-car-smell PELT and we still don't take it seriously enough.
Seems no matter what it is, the Death Race 2000 new car smell eventually wears off.
Re: (Score:1)
That's the main reason I don't want nuclear power. Humans are way to casual with it within a decade or so. Nothing bad happened so they start cutting corners .5%. And nothing bad happens... so they iterate.
And no, I have many baskets. :-)
A.I. is just one of them.
Inhuman or "inhuman"? (Score:2)
You mean that kind of inhuman [wikia.com]?
This is amazing because... (Score:3)
a new word! (Score:2)
Will Bots Create the Meta Game? (Score:3)
When humans start looking to bots to figure out the meta game that'll take out half the fun
Since we're comparing AI to human capability (Score:1)
Re: (Score:2)
Interesting proposition. The reverse could also be approximated: how many people would the power used by this system represent?
Re: (Score:2)
How about instead of that we allow humans to get bigger brains not confined by their energy consumption what would happen then ?
It's worst than that (Score:1)
And the AI's will never forget even the smallest detail and always use its knowledge to 100% efficiency.
If it was confined to games it would already be scary enough. Now imagine the same thing being applied to AI connected to real-life machines.
Re: (Score:1)
Dear past dweller,
If you can get your hands on it's training data you'll instantly know all it's future moves.
Best regards,
John Conner
Re: (Score:1)
Even in the future, people still haven't learned the difference between its and it's.
Complex? (Score:2)
Dota 2? Complex? Starcraft laughs at this, it must be so hard to micro one unit against a few enemies.