Support NeoGAF

[Tripolygon]

What I got from this is that we need a live commentator for all online games

Yes, especially FromSoft Games.

Also

Spoiler: Yup

Proving again that you don't need specific hardware accelerators or INT8 and lower precision do inference.

 

[Tomeru]

While it is cool af, and this is one of the reason I like Sony so much (in terms of tech and games) - is this or some parts of this be in gt7?

Edit:
Answered already I see.

Imagine having this in an oine b-spec race. How weird will that be...

 

[Elog]

hm, I'm not convinced this is "far more complex" than starcraft or dota AI

Yeah, that is why the behaviour of this AI is on the front cover of Nature - one of the highest impact scientific journals globally...

 

[EverydayBeast]

Microsoft didn't think Gran Turismo would still be hanging around after they put distance between them with the recent Forza entries.

 

[PharaoTutAnchAmun]

I’m still waiting for the PS3’s folding at home to cure cancer.

Thats was not the task of the PS3, the PS3 was used as a tool to help supercomputers folding, the idea was how more computers in the world where helping this enourmus task the more data they would get it would help to understand how cancer works, its nothing to make jokes about. When i had the PS3, i was always participating with this program..cancer is a bitch, and how sooner whe find a cure, how better.

 

[Loxus]

GT Sport

Yes, especially FromSoft Games.

Also

Well, seeing how Sony going this hard with AI & ML. It's hard to believe the PS5 doesn't support it.

 

[Anos Voldigoad]

The main takeaway from this is that SIE plans to put Ai in more games in the future. I'm excited to see the possibilities

 

[SLB1904]

sounds like it will be multiplat tech

 

[REDRZA MWS]

Thats was not the task of the PS3, the PS3 was used as a tool to help supercomputers folding, the idea was how more computers in the world where helping this enourmus task the more data they would get it would help to understand how cancer works, its nothing to make jokes about. When i had the PS3, i was always participating with this program..cancer is a bitch, and how sooner whe find a cure, how better.

I agree cancer sucks, but the Sony and all it’s bullshit tech marketing nonsense isn’t curing cancer. Started with “emotion engine”, then “reality synthesizer” and onto now Sophy…… it’s all bullshit. It’s a video game.

 

[Florin4k4]

Sportspersonship?! Really?
Even had a hard time typing it as my autocorrect had no idea what I was trying to say

 

[Aenima]

@17:50 is the first race.

This was really entertaining. Loved the 1st race, seeing the AI use a more risky driving line, that human players dont used, just to make the corner faster.

Curious to see how Kaz is going to implement it in GT7

 

[Tripolygon]

I agree cancer sucks, but the Sony and all it’s bullshit tech marketing nonsense isn’t curing cancer. Started with “emotion engine”, then “reality synthesizer” and onto now Sophy…… it’s all bullshit. It’s a video game.

Folding at home is a real thing. You can donate your PC time to the project. PS3 Cell processor was really good at doing those sorts of calculations. It was a project in collaboration with Stanford University and 15 million PS3 users donated 100 million computing hours to the project.

 

[PharaoTutAnchAmun]

@17:50 is the first race.

So when you drive behind Sophie, you are learning to take more advantage of the whole circuit, taking corners better, better braking points, where you dont have to break and so on.....

 

[ethomaz]

ionInteresting stuff.

Would be interesting to note if this is due to some "cracks" in the "simulation" or if these sorts of techniques would actually work on a real track with real cars.

Typically the "lines" that drivers use in these games are meant to mimic those used in real life.

But honestly.. from a machine learning standpoint... not sure how interesting this is lol Folks working on car AI aren't trying to make cars win races on controlled track conditions.

Imagine you are a professional real life racer.
You train every day the same curve to near perfection... but even so your can do it perfectly just 20% of times, of course some curves are easier than others so some you will have a less than 5% change to make it perfect while other ever over 20% change... now imagine a full lap has 20 curves... so you have 20 points per lap to miss the perfect.
At the end you train everyday to be the most close possible to the perfection in each lap... so that way you do your best time.
So in a race with the pressure and physical tired you will in a race with 60 laps reach near your best time maybe one or two times... all the others times will be a bit or a lot lower depending of the condition.
A perfect lap time? Maybe 1 in 1 million... just a super luck lap.

That is a normal Champion Racer.

That AI runs without mistakes making perfect lap times after perfect lap times without get tired... so with 60 laps it will be few seconds always ahead you to the point that at the end it will have a lap over the best Champion Racer.

I won't call it "cracks"... it is just it can do the best time the track allow without mistakes.
It still uses all the physical rules and race regulations... it just it is so good and makes no mistake in all the curves and laps that at the end it can really be called "super-human".

 

[SLB1904]

the ai got involved in an investigation lol

 

[IntentionalPun]

Imagine you are a professional real life racer.
You train every day the same curve to near perfection... but even so your can do it perfectly just 20% of times, of course some curves are easier than others so some you will have a less than 5% change to make it perfect while other ever over 20% change... now imagine a full lap has 20 curves... so you have 20 points per lap to miss the perfect.
At the end you train everyday to be the most close possible to the perfection in each lap... so that way you do your best time.
So in a race with the pressure and physical tired you will in a race with 60 laps reach near your best time maybe one or two times... all the others times will be a bit or a lot lower depending of the condition.
A perfect lap time? Maybe 1 in 1 million... just a super luck lap.

That is a normal Champion Racer.

That AI runs without mistakes making perfect lap times after perfect lap times without get tired... so with 60 laps it will be few seconds always ahead you to the point that at the end it will have a lap over the best Champion Racer.

I won't call it "cracks"... it is just it can do the best time the track allow without mistakes.

I'm not sure why you think I needed this explained to me... don't think you understand really what I'm saying.

This AI has learned the fastest way to race tracks in Gran Turismo 7, doing things no actual car racer would ever do.. my point was that is probably revealing where the "simulation" is flawed, as I doubt these things being done in the real world would work. So what has this AI/ML project actually really done?

It honestly sounds like pretty basic ML stuff... "fun" but why is this groundbreaking?

 

[ethomaz]

I'm not sure why you think I needed this explained to me... don't think you understand really what I'm saying.

This AI has learned the fastest way to race tracks in Gran Turismo 7, doing things no actual car racer would ever do.. my point was that is probably revealing where the "simulation" is flawed, as I doubt these things being done in the real world would work.

I think it can be done in real world at the same conditions... same tracks, same cars, weather conditions, etc.
GT7 tracks are very accurate to real life ones.

Outside that... I don't know because the AI while being perfect needs to deal with more obstacles and even different speeds.

 

[PharaoTutAnchAmun]

I agree cancer sucks, but the Sony and all it’s bullshit tech marketing nonsense isn’t curing cancer. Started with “emotion engine”, then “reality synthesizer” and onto now Sophy…… it’s all bullshit. It’s a video game.

Its not bullshit tech marketing, it was serious.and emotion engine and RS, was just name giving of tech every company does do that. Even MS with theyr Velocity Architechture and Intelligent Delivery.
But Sophy is not bullshit, its a real thing. And Sophy will be used in videogames, why are you so triggert?

 

[IntentionalPun]

I think it can be done in real world at the same conditions... same tracks, same cars, weather conditions, etc.
GT7 tracks are very accurate to real life ones.

And I'm highly doubting that.

 

[SLB1904]

And I'm highly doubting that.

dave perel profissional gt3 driver said the same thing. whats your doubts based on?

 

[IntentionalPun]

dave perel profissional gt3 driver said the same thing. whats your doubts based on?

I'm not saying it isn't a really great simulation.

I'm saying that considering this AI re-invented the "lines" drivers would normally take, that probably represents where the simulation is a bit off, not where every driver of these tracks ever has miscalculated what the best "lines" would be.

It's absolutely not a perfect simulation and an AI would absolutely be able to find where that allows for driving that normally wouldn't work on a real track to have superior results, is all I'm saying.

 

[nemiroff]

I don't have time to wade through all the PR bullshit, but I am curious, so for those who did; What's unique about this compared to the numerous other ML/AI projects, training and papers we've had in the last decade?

 

[SLB1904]

I'm not saying it isn't a really great simulation.

I'm saying that considering this AI re-invented the "lines" drivers would normally take, that probably represents where the simulation is a bit off, not where every driver of these tracks ever has miscalculated what the best "lines" would be.

It's absolutely not a perfect simulation and an AI would absolutely be able to find where that allows for driving that normally wouldn't work on a real track to have superior results, is all I'm saying.

yeah, my comment was about how accurate the gt tracks are to their real world counterpart. of course the simulation isnt 1 to 1.

 

[REDRZA MWS]

Its not bullshit tech marketing, it was serious.and emotion engine and RS, was just name giving of tech every company does do that.

Sony and big tech also have a bridge to sell you. It’s a video game, going to be a damn good one and will sell gangbusters. But Sophy isn’t Terminator AI from the future.

 

[SLB1904]

I don't have time to wade through all the PR bullshit, but I am curious, so for those who did; What's unique about this compared to the numerous other ML/AI projects, training and papers we've had in the last decade?

i guess you have to read this to find out



if i had to guess, they just use all the studies through the decades to create sophy. the impressive part is using in videogames.
imagine if they use this ai in other action games. like 3rd person shooters and stuff.

 

[IntentionalPun]

yeah, my comment was about how accurate the gt tracks are to their real world counterpart. of course the simulation isnt 1 to 1.

Oh gotcha; I was saying I highly doubt these tactics invented by the AI would work in the real world.

It'd be interesting to test some of them; but it's not a 1:1 simulation, and these drivers have already tried adjusting every line imaginable, likely have race teams employing all kinds of tech to calculate those lines/etc.

 

[Faithless83]

Isn't it ironic that they used a female name for a driver AI?

As someone who's studying AI as a side project, this is extremely impressive.

To the "DOTA/RTS/Chess is harder", not it isn't.
Even though there are a lot of variables on these, there is a lot more adjustment and fine tuning on the fly in racing sims.
Weather, tires, cars, course lines and other drivers are way more complex than fixed outcomes in chess/RTS.

 

[IntentionalPun]

i guess you have to read this to find out



if i had to guess, they just use all the studies through the decades to create sophy. the impressive part is using in videogames.
imagine if they use this ai in other action games. like 3rd person shooters and stuff.

I still don't really get it.

Why does Sony think creating an AI that can beat anyone is some useful innovation?

Why do they even really think this is something others haven't done?

 

[SLB1904]

I still don't really get it.

Why does Sony think creating an AI that can beat anyone is some useful innovation?

Why do they even really think this is something others haven't done?

thats the thing, this ai is learning how to play gran Turismo, isnt just a gran turismo ai. basically can learn any game. the breakthrough isn't supposed to be mechanic stuff. if you watch the race, the behavior was almost like a human, the other racing games ai is very clinical and precise.

 

[Faithless83]

I still don't really get it.

Why does Sony think creating an AI that can beat anyone is some useful innovation?

Why do they even really think this is something others haven't done?

It's not about beating the top players, it's about simulating a real person playing.
Choosing racing sims as a start point helps a lot since there are so many variables in it, much more than other genres.

 

[systemr123]

I still don't really get it.

Why does Sony think creating an AI that can beat anyone is some useful innovation?

Why do they even really think this is something others haven't done?

It is very useful for creating self driving cars, which might be perfect timing since Sony is bringing its own EV cars to the market soon. I own a self driving Tesla and while it’s absolutely not perfect yet, it’d be hard to go back to having to steer and accelerate/brake by yourself

 

[IntentionalPun]

thats the thing, this ai is learning how to play gran Turismo, isnt just a gran turismo ai. basically can learn any game. the breakthrough isn't supposed to be mechanic stuff. if you watch the race, the behavior was almost like a human, the other racing games ai is very clinical and precise.

Well that's what neural networks do... Sony didn't invent the concept. I understand how actual machine learning based AIs work.

 

[IntentionalPun]

It is very useful for creating self driving cars, which might be perfect timing since Sony is bringing its own EV cars to the market soon. I own a self driving Tesla and while it’s absolutely not perfect yet, it’d be hard to go back to having to steer and accelerate/brake by yourself

Those AIs are also first built by using machine learning against simulations, that include what you'd see in real world driving scenarios like pedestrains and stoplights and all of that jazz.

And they are made more complex by also having to have AI that attempts to recognize what is actually in front/beside/behind them using cameras/sensors, vs. a video game or simulation where all of that data is perfectly accurate.

I'm just not seeing any breakthroughs here.

 

[SLB1904]

Well that's what neural networks do... Sony didn't invent the concept. I understand how actual machine learning based AIs work.

I didn't say that lol
What I'm saying it's fantastic the the games are able to take advantage of this stuff.
For example for people that plays cod, bots are shit. If in the future bots behave like human it will be fun. It could be for anything really

 

[systemr123]

Those AIs are also first built by using machine learning against simulations, that include what you'd see in real world driving scenarios like pedestrains and stoplights and all of that jazz.

And they are made more complex by also having to have AI that attempts to recognize what is actually in front/beside/behind them using cameras/sensors, vs. a video game or simulation where all of that data is perfectly accurate.

I'm just not seeing any breakthroughs here.

I’m quite sure this AI knows much better how to overtake and avoid collisions than the current software in actual cars. Revolutionary? Maybe no. But every step is a step into self driving bliss

 

[Thirty7ven]

I still don't really get it.

Why does Sony think creating an AI that can beat anyone is some useful innovation?

Why do they even really think this is something others haven't done?

I know you're busy being the smartest guy in the room, and the folks at Nature, A.I programmers, and University doctorates don't have a hope to get where you are right now, but try to look at in the same way as the A.I that succeeded at GO.

Except here its not just about winning, but an A.I that needs to react very quickly to what's happening around it while trying not just to be the fastest, but abiding by sportsmanship principles.

And instead of asking "Why do they even really think this is new?", maybe you should look inwards and ask yourself... "What do I actually know?"

Otherwise, show the examples you have. And please don't just show an example of machine learning.... how do you think this works? You tell a machine you want an A.I and it does the rest of the work?

 

[Loxus]

Those AIs are also first built by using machine learning against simulations, that include what you'd see in real world driving scenarios like pedestrains and stoplights and all of that jazz.

And they are made more complex by also having to have AI that attempts to recognize what is actually in front/beside/behind them using cameras/sensors, vs. a video game or simulation where all of that data is perfectly accurate.

I'm just not seeing any breakthroughs here.

If Sophy was perfectly accurate.
Don't you think all 4 Sophy would of placed first in all 3 tracks?

Sophy was born using Machine Learning on a Super Computer to act like a human. Ain't nothing basic about that.

 

[Inviusx]

Pretty much what I expected although I didn't think this would roll into GT7. I guess we will wait and see.

 

[SLB1904]

 

[IntentionalPun]

I know you're busy being the smartest guy in the room, and the folks at Nature, A.I programmers, and University doctorates don't have a hope to get where you are right now, but try to look at in the same way as the A.I that succeeded at GO.

Except here its not just about winning, but an A.I that needs to react very quickly to what's happening around it while trying not just to be the fastest, but abiding by sportsmanship principles.

And instead of asking "Why do they even really think this is new?", maybe you should look inwards and ask yourself... "What do I actually know?"

Otherwise, show the examples you have. And please don't just show an example of machine learning.... how do you think this works? You tell a machine you want an A.I and it does the rest of the work?

I'm asking questions because I'd like to know what the actual breakthrough is.

I also have no knowledge of whether "Nature Portfolio" are some AI experts either..

To your point, maybe I'm coming across more dismissive than I need to be, but I do understand ML.. I understand the concepts being talked about in the abstract of the paper they published as well.

I'm simply asking, WHAT is the breakthrough? What was accomplished beyond what others have done before? What will the application be?

I don't see how that is some unfair question...

 

[SLB1904]

I'm asking questions because I'd like to know what the actual breakthrough is.

I also have no knowledge of whether "Nature Portfolio" are some AI experts either..

To your point, maybe I'm coming across more dismissive than I need to be, but I do understand ML.. I understand the concepts being talked about in the abstract of the paper they published as well.

I'm simply asking, WHAT is the breakthrough? What was accomplished beyond what others have done before? What will the application be?

I don't see how that is some unfair question...

the only way to know is to read the magazine i guess

 

[SLB1904]

second race. huge improvement on ai behavior

 

[Anos Voldigoad]

I wouldn't mind smashing Sophy

 

[Three]

Interesting stuff.

Would be interesting to note if this is due to some "cracks" in the "simulation" or if these sorts of techniques would actually work on a real track with real cars.

Typically the "lines" that drivers use in these games are meant to mimic those used in real life.

But honestly.. from a machine learning standpoint... not sure how interesting this is lol Folks working on car AI aren't trying to make cars win races on controlled track conditions.

I don't think they will work on a real track. There are just too many variables in the real world. A small pebble can change the input required but not in a game. Your input is always the same.

 

[DJ12]

Will come in an update.

So basically I need to beat the game before they implement this or I'm never winning a race

 

[IntentionalPun]

I don't think they will work on a real track. There are just too many variables in the real world. A small pebble can change the input required but not in a game. Your input is always the same.

I'm not really talking about having the AI control cars on a track; I mean the strategies the AI came up with, would those influence real world drivers?

They claim the AI came up with new "lines", meaning the path the car travels, when to accelerate, when to break, not typically used. My understanding is the recommended "lines" in these games are based on what real world drivers do on these very tracks.

Would be cool to see a real world driver try these same "lines" and see if they can improve track time.

Or if it doesn't improve track times; Sony could try to figure out why their simulation doesn't match reality, and adjust the simulation accordingly.

 

[nemiroff]

I'm just not seeing any breakthroughs here.

That's because there is most likely none.. But we'll see.

Would be cool to see a real world driver try these same "lines" and see if they can improve track time.

Or it could simply be more akin to an aerodynamically unstable fly-by-wire aircraft which is impossible to fly manually by a human because it relies on computers to do hundreds of corrections per second thus drawing advantages out of being inherently unstable. Maybe a human doesn't have the processing power to make the same lines as "sophy", IDK.. How much human restrictions did they put into this. "Ideal" racing lines are also dynamic according to a whole range of factors, like car, setup, track conditions, tires, fuel amount, and much more. So I'd also like to know how much of this data they've taken into consideration when doing the ML training.

 

[Raven117]

Its got nothing to do with reaction time, its about making extremely complex intelligent decisions, in Dota/LoL that could be running away to buy something, engaging, supporting someone else who just turned up, chasing, staying put. In GT it has to avoid the other cars, its not following a live changing economy as the game/race goes on.

There's only so much that could happen in a race since its a linear journey round a set track where the speed of opponents will be within a predictable range, I'm not saying there aren't other factors to consider in GT outside of the immediate logistics of the cars around you, but I think its fair to say its nowhere near as complex as a free moving open map where any player can be anywhere and decide to retreat or advance for many different reasons.

I'd say the decisions that could be made by an AI in a high-level game of DotA/LoL would change at such a pace that the AI could often never make a decision that wasn't too out of date to be useful, just because it can't reasonably predict what the opponents will be doing that far in the future. Compared to a racing game where you can predict what the opponent will do further into the future because there are so many less avenues of decision the opponent can go down.

Lol! Effing wow. You know absolutely nothing about racing.

 

[Anos Voldigoad]

We gonna learn these tracks and we gonna take that to be better than a Machine, exciting times we are in here. Can you imagine the fun times we can have with intelligent ai in co op games either as your companion or nemesis. Games stopped being fun with dumb ai, this will revolutionise the gameplay experience

 

[SLB1904]

Nerf sophy