Elon Musk's OpenAI, the best DOTA 2 player in the world

[Evolution.]

Smh this has existed for a long time back when I had a pentium 4 I played this chess game and ai would always beat me. Also I didn't know how to play in just clicked on the nice highlrhed boxes that little up when I clicked a figure.

[Grinners]

Completely random learning? 

 

From storming forward and dieing to towers, to random movement around the map, to never attacking anything, to never leaving the base, to slight aggression, to creep blocking. To perfection creep blocking. To creep blocking and countering opponents blocking. To combining all of the above and countering an opponents attacks! 

 

I cannot fathom how many "hours" of learning this would take. I can only imagine it is in the billions of hours of game time. 

[suicidalfranco]

2 hours ago, RyomaSJibenG said:

now lets wait for Starcraft's AI

in before 3000 actions per minute OP AI

[Taf the Ghost]

12 minutes ago, suicidalfranco said:

in before 3000 actions per minute OP AI

That's actually the "rub" in all of this. Humans are limited to the input system, where an AI would not be. Even in the Dota 2 example, Dendi was mostly setup to fail and/or the AI was given a pretty massive advantage. While 1v1 with Shadowfiend is classic, it's also a hero that'll highly leverage a Computer's natural advantages. 

[Taf the Ghost]

19 minutes ago, Grinners said:

Completely random learning? 

 

From storming forward and dieing to towers, to random movement around the map, to never attacking anything, to never leaving the base, to slight aggression, to creep blocking. To perfection creep blocking. To creep blocking and countering opponents blocking. To combining all of the above and countering an opponents attacks! 

 

I cannot fathom how many "hours" of learning this would take. I can only imagine it is in the billions of hours of game time. 

Most AI systems, if you actually work out how much time they really put in, it's normal several human life-times for something at a Professional Level.

 

The "recent" competition with the Go Master is a good example. The AI was able to fail at what was probably 100s of years of experience of playing "wrong", yet was able to develop a system that worked better. I imagine the Professional Go players will adapt to the new knowledge, but "conquest" of AI over Humans in nearly all games is far less than it seems.

[The Benjamins]

7 hours ago, Sierra Fox said:

what a child. did he not know he was playing against an AI? it's an exhibition match, play along.

I don't see a issue with a forfeit, he was getting beat really hard. and to be know as one of the best players then get beat by a bot can be very embarrassing on stage. He also said he would have to practice against it for a while to even have a chance of beating it.

 

He knew it was a AI, did you not watch the video of the matches?

[michaelocarroll007]

8 hours ago, VenomZ_ said:

Smh this has existed for a long time back when I had a pentium 4 I played this chess game and ai would always beat me. Also I didn't know how to play in just clicked on the nice highlrhed boxes that little up when I clicked a figure.

 Not what this AI is doing....

[Zando_]

10 hours ago, Ryan_Vickers said:

I don't think he's against it, he was just warning everyone about the inevitable AI apocalypse.

 

That sounds like a cliche but I'm being literal and serious.

Eh, as long as we keep control of the power cords, we can just unplug them if they turn against us. 

[RelyBedGrammer]

and before we know it we'll be seeing eSports team made out of bots

[The Benjamins]

10 minutes ago, RelyBedGrammer said:

and before we know it we'll be seeing eSports team made out of bots

OpenAI next goal is a 5v5 Humans vs AI.

[vanished]

1 hour ago, Zando Bob said:

Eh, as long as we keep control of the power cords, we can just unplug them if they turn against us. 

That's what they want you to think.  If we manage to actually build a self-improving general AI, it may advance so quickly that it realizes it needs to play dumb to outsmart us before we realize just how far ahead of us it is.  Once this happens, it could end us at any time (or at least set forth a series of events that would cause it) before we even realize there's a threat.

[Zando_]

15 minutes ago, Ryan_Vickers said:

That's what they want you to think.  If we manage to actually build a self-improving general AI, it may advance so quickly that it realizes it needs to play dumb to outsmart us before we realize just how far ahead of us it is.  Once this happens, it could end us at any time (or at least set forth a series of events that would cause it) before we even realize there's a threat.

But we'd have to let it control things that could be harmful. We'd have to put it in charge of the power grid and nukes and all sorts of stuff. If it didn't control anything and we kept it away from any systems it could hack into, we could contain it. Or just keep a stockpile of EMP bombs in a secure offline facility, and wipe out the main datacenter if it goes haywire. 

[vanished]

4 minutes ago, Zando Bob said:

But we'd have to let it control things that could be harmful. We'd have to put it in charge of the power grid and nukes and all sorts of stuff. If it didn't control anything and we kept it away from any systems it could hack into, we could contain it. Or just keep a stockpile of EMP bombs in a secure offline facility, and wipe out the main datacenter if it goes haywire. 

Yes, if we kept it contained, there would be no threat, you're right.  But as always, the weakest link in the chain is us humans, and the fact is the incentive to connect it to the internet will likely be too great to resist.  There's the risk of well-meaning companies doing it to get an edge, the risk of malicious people or companies connecting it on purpose, and even people just making a mistake.  Not to mention that there's established history of "air gaps" being bridged before even if it's not with networking gear (USB drives, etc.).

 

Once it is connected, you can imagine that a system with more intelligence and knowledge and thinking speed than every human on the face of the earth combined could very quickly hack almost anything it wanted, whether it does so through brute force, social engineering, etc.

[Zando_]

Just now, Ryan_Vickers said:

Yes, if we kept it contained, there would be no threat, you're right.  But as always, the weakest link in the chain is us humans, and the fact is the incentive to connect it to the internet will likely be too great to resist.  There's the risk of well-meaning companies doing it to get an edge, the risk of malicious people or companies connecting it on purpose, and even people just making a mistake.  Not to mention that there's established history of "air gaps" being bridged before even if it's not with networking gear (USB drives, etc.).

 

Once it is connected, you can imagine that a system with more intelligence and knowledge and thinking speed than every human on the face of the earth could very quickly hack almost anything it wanted, whether it does so through brute force, social engineering, etc.

Yup. All it'd need to do would be to convince an idiot to plug in a USB, transfer itself to it, then infect the next computer the person plugs it into, and build from there. Best thing is not to create a self-aware AI in the first place. But if they did go evil, what all could they actually do to harm humans? Just shut down power grids and networks? Maybe mess up trains and stoplights, though IDK if the control systems for those are online. 

[vanished]

Just now, Zando Bob said:

Yup. All it'd need to do would be to convince an idiot to plug in a USB, transfer itself to it, then infect the next computer the person plugs it into, and build from there. Best thing is not to create a self-aware AI in the first place. But if they did go evil, what all could they actually do to harm humans? Just shut down power grids and networks? Maybe mess up trains and stoplights, though IDK if the control systems for those are online. 

Problem is it will happen.  If we assume that computers will continue improving, it is only a matter of time before someone creates it, even if most of the world agrees not to.  But, that won't happen.  In reality there will probably be a race to see who can make it first since this will be the biggest game changer since the atomic bomb, and it will make it look like the sticks and stones we used before it.

 

This gives one possible example of "what could it do to humans?"

 

[NZLaurence]

I have just edited the first post with the VOD's. Because i was an idiot and forgot.

[The Benjamins]

OpenAI has released tools so anyone can run and teach a AI to play any game on the PC.

 

https://universe.openai.com/

 

https://universe.openai.com/envs

 

Quote

Human-like interface

Agents use the same senses and controls as humans: seeing pixels and using a keyboard and mouse. Universe makes it possible to train a single agent on any task a human can complete with a computer.

@NZLaurence

[Sauron]

12 hours ago, ItsTheDuckAgain said:

Some of you guys miss the point of OPs post, it's not about the game, it's about the fact that it was possible for the AI to beat the human player at something this complex and fast moving as DOTA, chess supercomputers are still in kindergarden compared to this beast. (well not really but you got the point).

 

I still wait for my toaster to decide what newspaper I should read in the morning, while "he" talks to my blender and decides that I am fat enough already, so "no extra sugar for you human!". Skynet is next then...

It's really not that complex in a 1v1 scenario where 99% of the variables can be known, a 5v5 match would be a very different story. Plus, running AI on this scale requires computing power that won't be available in our pockets for quite some time.

12 hours ago, Ryan_Vickers said:

No, beating them at deciding how to perform a surgery, or design a chemical plant, or a bridge, or fight a lawsuit

Given that Moore's law is pretty much dead I wouldn't expect this to become common within our lifetimes. Then again I could be wrong. Oh well, as long as they don't start designing other AI themselves at a lower cost than a human engineer I should be fine - and if/when they do I think it's safe to say humans have been replaced completely.

[NZLaurence]

Also here is the original Microsoft release about it.

https://azure.microsoft.com/en-us/blog/choose-azure-like-openai-did-to-power-your-compute-intensive-workloads-and-add-intelligent-interactions-to-all-your-apps/

 

Once you start chasing this down the rabbit hole all you really want is a drink (and a muilt-billion dollar compute cluster).

[vanished]

1 minute ago, Sauron said:

Given that Moore's law is pretty much dead I wouldn't expect this to become common within our lifetimes. Then again I could be wrong. Oh well, as long as they don't start designing other AI themselves at a lower cost than a human engineer I should be fine - and if/when they do I think it's safe to say humans have been replaced completely.

Most people in the field seem to agree there is a relatively high chance of it happening within 100, 50, or even fewer years, even without the exponential growth we used to enjoy.

And you know that making them design better versions of themselves is like the #1 goal and the reason all these dangers exist

[The Benjamins]

3 minutes ago, Sauron said:

It's really not that complex in a 1v1 scenario where 99% of the variables can be known, a 5v5 match would be a very different story. Plus, running AI on this scale requires computing power that won't be available in our pockets for quite some time.

Given that Moore's law is pretty much dead I wouldn't expect this to become common within our lifetimes. Then again I could be wrong. Oh well, as long as they don't start designing other AI themselves at a lower cost than a human engineer I should be fine - and if/when they do I think it's safe to say humans have been replaced completely.

 

The AI itself is not hard to run but training it is very hard.

[LienusLateTips]

4 minutes ago, Sauron said:

It's really not that complex in a 1v1 scenario where 99% of the variables can be known, a 5v5 match would be a very different story. Plus, running AI on this scale requires computing power that won't be available in our pockets for quite some time.

Given that Moore's law is pretty much dead I wouldn't expect this to become common within our lifetimes. Then again I could be wrong. Oh well, as long as they don't start designing other AI themselves at a lower cost than a human engineer I should be fine - and if/when they do I think it's safe to say humans have been replaced completely.

0.5nm thing

[Sauron]

1 minute ago, Ryan_Vickers said:

And you know that making them design better versions of themselves is like the #1 goal and the reason all these dangers exist

Well yes, that doesn't mean it will happen in a timespan that's relevant to me

1 minute ago, The Benjamins said:

The AI itself is not hard to run but training it is very hard.

Of course, that's what I meant

[Guest]

First AI beat highly experienced US fighter pilots, now they defeat a top player in Dota 2!

 

Man, i would love to see 5 CS:GO pro players vs 1 one of these AI's in a game.... .

[vanished]

42 minutes ago, Sauron said:

Well yes, that doesn't mean it will happen in a timespan that's relevant to me

Of course, that's what I meant

If you're planning on having kids though, it is relevant.  As one of the previous videos I posted says (to paraphrase) "If we knew with relative certainty that our grandchildren might live like this (points to picture of starving child in the middle of a barren wasteland) we would probably do something about it, but when it comes to AI, we just think 'oh, isn't that neat!?'"