Gran Turismo 7: Latest news and discussion thread

[RandomCarGuy17]

As for why not swap it for the 2018 BMW Motorsport M4 GT4 when that was launched in 2018, I can only guess increased license cost and time resource, simply to introduce a second Gr.4 M4 for BMW. Cost to benefit ratio?

I would like to believe the 2nd answer of cost to benefit ratio, but then I'm reminded of the company I'm talking about. That answer just sounds way too rational for PD.

 

[MatskiMonk]

I would like to believe the 2nd answer of cost to benefit ratio, but then I'm reminded of the company I'm talking about. That answer just sounds way too rational for PD.

Can't add 1929 Mercedes S. Barker Tourer if you've spent all your money on a second third fourth F82 M4.

 

[RandomCarGuy17]

Can't add 1929 Mercedes S. Barker Tourer if you've spent all your money on a second third fourth F82 M4.

I hear your point, but still going to say it. Too rational for Polyphony Digital, I feel like there's a more ridiculous reason than that.

 

[ddm]

Beacause it's an extremely new technology and it needs time to be devoleped, changhing the whole game AI would be too risky, it's also a self-learning machine and it probably needs to learn many things before working properly, probably pd will use everything they learned in GT7 to fully implenent Sophy in GT8.

Not at all the case, note that it's usually wise to not comment about how deep learning AI works if you don't know yourself...

Sophy as we know it right now is 5-60 year old AI tech (depending on how you're counting),
and it is honestly totally unreasonable for it to ever be a complete replacement for all AI.

And it has already learnt everything that it will ever learn, until the scoring criteria / model structure is changed and it is retaught.
It never "learns more" after at most a week of training: instead it has to learn from scratch after every tweak.

I don't see Sophy ever being implemented as a complete replacement.
Given this has been feasible to achieve on modern hardware for the past 10 years, why do you think no game developer has ever used a deep-learning AI for it's primary AI opponents?
Quick hint: the answer is not "because PD are so forward-thinking that they're gonna be first to do it!".

 

[Saargasm]

Not at all the case, note that it's usually wise to not comment about how deep learning AI works if you don't know yourself...

Sophy as we know it right now is 5-60 year old AI tech (depending on how you're counting),
and it is honestly totally unreasonable for it to ever be a complete replacement for all AI.

And it has already learnt everything that it will ever learn, until the scoring criteria / model structure is changed and it is retaught.
It never "learns more" after at most a week of training: instead it has to learn from scratch after every tweak.

I don't see Sophy ever being implemented as a complete replacement.
Given this has been feasible to achieve on modern hardware for the past 10 years, why do you think no game developer has ever used a deep-learning AI for it's primary AI opponents?
Quick hint: the answer is not "because PD are so forward-thinking that they're gonna be first to do it!".

I don't see Sophy ever being implemented as a complete replacement.

Not sure if you meant for your reply to come off a little snarky, but it did. You can however, decide to maybe explain why you feel this way rather than being vague and slightly condescending that we don't know.

 

[Grimm6Jack]

With all the talk of Tomica, Decided to look how much a Tomica diecast that matched my car were going for so I can maybe buy one... oof. View attachment 1178178

I wanted to make a collection of my favourite cars and race cars (1/18 or 1/24 models), I didn't look at Tomica models, mostly maisto and autoart... And when I saw something like an LFA costing 700€ on ebay... Nope. And since I like these models perfectly accurate/detailed and not made of crap plastic, I had to leave that dream... Just too expensive. Certainly these aren't the prices when you buy these models new right?

I don't remember the AI being so mistake prone in the old games like they are in GT7. I suppose it gives them a bit of a human element to an extent.

That's actually to have AI behave like humans. There are some absolutely ridiculous situations though, like the way they slow down to almost an alt when there's a yellow flag, and some other bizarre moments I've seen in a ton of youtube clips.

Compared to past GTs, GT7 has definitely improved the AI, but there's still a loooot to improve. The way they just stick to the racing line is not natural... It pains me to see something like a LaFerrari or Veyron having difficulties overtaking much slow cars on a straight like the SLR McLaren or something because they are stuck literally behind them and don't know how to move to one side and overtake...

Yeah... it's why I keep coming back to this game despite the lack of SP content. The graphics and the driving simply sell.

PD has the core of the game settled. The driving, taking into account that it has to appease to a large audience and not strict sim gamers, is exceptionally good (and it keeps on improving) and the graphics, especially when it comes to car detail is unmatched by any other racing/driving game.

They fix the plethora of issues they have, which quite frankly, most are so simple to the point it infuriates me why they haven't fixed them already (almost all of them are design decisions that can easily be changed, added or removed), and just keep on adding tracks and cars over time, and you have the perfect racing game.

 

[ddm]

Not sure if you meant for your reply to come off a little snarky, but it did. You can however, decide to maybe explain why you feel this way rather than being vague and slightly condescending that we don't know.

A PS4 doesn't have anywhere near enough processing power or memory to run 15 deep learning agents.
A PS5 most likely also can't achieve that.
Note that during all prior demonstrations, Sophy has been run with one agent per PS5.
This is also not a genericised model, but a single car, single track, single environment model: it only handles a few trillion times less environments than a genericised model would need.

And Sophy (unless Sony AI make an actual new major breakthrough in AI tech) will be as bad if not worse at adapting to racing against players slower than the absolute best, as the current AI is at adapting to racing against, well, anyone.

The snark is because I'm sick and tired of people thinking Sophy is anything special, which it isn't.
Obviously that problem occurred because Polyphony Digital & Kaz have thought that Sophy is something that it is not, because they don't fundamentally understand how it works.
The entire set of slides with how Sophy can be "a friend" and all the other nonsense says nothing other than "whoever wrote these slides doesn't understand deep learning".

It has been known for 10+ years that deep learning AI cannot be taught to be a natural AI opponent for anything other than a top skill player, because you can't quantify that requirement. There is no logical statistic for tracking "fun".
Hence the only logical statistic Sophy has been trained with so far is "time" (well, that and collision penalties).

 

[Lux Klonoa]

They fix the plethora of issues they have, which quite frankly, most are so simple to the point it infuriates me why they haven't fixed them already (almost all of them are design decisions that can easily be changed, added or removed), and just keep on adding tracks and cars over time, and you have the perfect racing game.

Eh, my only issue with the game is, aside from the lack of events and event variety, the design decision of merging the career mode into the World Circuits menu. It's a pain in the butt trying to find the group of events that you're trying to replay when you have to go to the circuits themselves to find them. Why aren't they grouped together on a single list (a.k.a the GT League screen) for quick access? It's mind-boggling. =/

 

[05XR8]

Eh, my only issue with the game is, aside from the lack of events and event variety, the design decision of merging the career mode into the World Circuits menu. It's a pain in the butt trying to find the group of events that you're trying to replay when you have to go to the circuits themselves to find them. Why aren't they grouped together on a single list (a.k.a the GT League screen) for quick access? It's mind-boggling. =/

I can only think it’s about globetrotting. However, even with that, they still stuff it up by not allowing us to rotate the globe. We have to jump out and click on a menu instead of spinning the planet.

 

[Saargasm]

A PS4 doesn't have anywhere near enough processing power or memory to run 15 deep learning agents.
A PS5 most likely also can't achieve that.
Note that during all prior demonstrations, Sophy has been run with one agent per PS5.
This is also not a genericised model, but a single car, single track, single environment model: it only handles a few trillion times less environments than a genericised model would need.

And Sophy (unless Sony AI make an actual new major breakthrough in AI tech) will be as bad if not worse at adapting to racing against players slower than the absolute best, as the current AI is at adapting to racing against, well, anyone.

The snark is because I'm sick and tired of people thinking Sophy is anything special, which it isn't.
Obviously that problem occurred because Polyphony Digital & Kaz have thought that Sophy is something that it is not, because they don't fundamentally understand how it works.
The entire set of slides with how Sophy can be "a friend" and all the other nonsense says nothing other than "whoever wrote these slides doesn't understand deep learning".

It has been known for 10+ years that deep learning AI cannot be taught to be a natural AI opponent for anything other than a top skill player, because you can't quantify that requirement. There is no logical statistic for tracking "fun".
Hence the only logical statistic Sophy has been trained with so far is "time" (well, that and collision penalties).

I don't mean for this to come off as attacking you, ddm. You clearly are very intelligent, based off your previous posts - they bring much needed insight that almost all wouldn't know otherwise. At the same time, your tone can be very elitest/snarky/condescending when it doesn't need to be. You would do a better job making your point without it.

Are you really, "sick and tired" of people restating what was presented to them from Sony - because ddm says said entity doesn't understand the AI model they themselves created? Ok, maybe, but am I supposed to blindly trust your opinion and discard the peer-reviewed articles, the YT presentations (full-length one), the sizable investment and spurious marketing of Sophy?

 

[ddm]

discard the peer-reviewed articles

The peer-reviewed articles contain nothing relevant to how Sophy could be implemented in an enjoyable way in the game.
They describe a system that makes an incredibly efficient AI agent for GT7: which is nothing particularly new, considering we've had incredibly efficient deep learning AI agents for specific video games for 5 years.

I'm not telling you to discard them: rather I'd prefer for people to actually read them, as well as other articles on and around how reinforcement learning works.
Comparing the articles to PD's presentations on Sophy shows there is very little relevance: the presentations are clearly full of claims that the article does not claim, and the presentations should not be treated as peer-reviewed content themselves.

Absolutely you shouldn't blindly trust me: but you shouldn't blindly trust PD either on this subject.
Deep learning AI is a complicated subject and it is basically a completely different concept to standard game AI: hence why I believe PD could easily have been originally "oversold" on the idea by Sony AI.
And as far as I'm aware, Sony AI themselves haven't claimed anything extraordinary from Sophy: only PD has.

 

[snc]

The peer-reviewed articles contain nothing relevant to how Sophy could be implemented in an enjoyable way in the game.
They describe a system that makes an incredibly efficient AI agent for GT7: which is nothing particularly new, considering we've had incredibly efficient deep learning AI agents for specific video games for 5 years.

I'm not telling you to discard them: rather I'd prefer for people to actually read them, as well as other articles on and around how reinforcement learning works.
Comparing the articles to PD's presentations on Sophy shows there is very little relevance: the presentations are clearly full of claims that the article does not claim, and the presentations should not be treated as peer-reviewed content themselves.

Absolutely you shouldn't blindly trust me: but you shouldn't blindly trust PD either on this subject.
Deep learning AI is a complicated subject and it is basically a completely different concept to standard game AI: hence why I believe PD could easily have been originally "oversold" on the idea by Sony AI.
And as far as I'm aware, Sony AI themselves haven't claimed anything extraordinary from Sophy: only PD has.

Im not sure is nothing special, remember google made quite a big halo that their ai Alpha Zero beat not ai cheess software and imo racing game is more complex than chess game at least in context of ai.

 

[Samus]

GT7 seems to have received its first publisher discount in the UK. PS5 version now running at around £45, PS4 around £35 at most shops. Down from £65 and £55 respectively.

https://www.currys.co.uk/products/playstation-gran-turismo-7-ps5-10234767.html

https://www.currys.co.uk/products/playstation-gran-turismo-7-10234771.html

 

[ddm]

Im not sure is nothing special, remember google make quite a halo that they ai Alpha Zero beat not ai cheess software and imo racing game is more complex than chess.

Have you heard of OpenAI? Prior to GPT, DALL-E and GitHub Copilot, OpenAI made a few iterations upon a Dota 2 bot that learns in a roughly similar way to how Sophy does for GT7.
In a reduced feature set game of Dota 2, this bot could defeat professional players.
This bot was first demonstrated live in 2017.
(They later greatly expanded the feature set of the games, which reduced it's efficiency but it still held a 99.4% winrate over 42,729 games.)

You're probably right about a racing game being more complex than chess.
But ... Sophy really doesn't "play a racing game".
It has significantly reduced data passed to it, some of which could be considered cheating if a player had access to them.
Sophy also only plays "60 turn games". It only finds the best solution for the next 6 seconds at 10Hz.
It does not see the track, but instead how much track is to the left & right, lap distance & incline.
It does not see the movements of a car, but instead tyre load, tyre slip angle, angular velocities, car velocity & acceleration.
This comes to around 100 data inputs, and there are only 3 data outputs. (Throttle, brake, steer. It uses AT gears.)

A game like chess has ~400 different states every 2 moves.
AlphaZero must calculate every move until the end of the game, because a move in chess can invisibly cause disaster more than 100 turns later. There's a whole lot more possible moves too, which means a whole lot more data outputs.

Obviously, this doesn't mean making Sophy is "easy".
And you'll still need to be a research company to train it because you still need a few spare hundred-thousand $ supercomputers laying around.
But nothing Sophy does couldn't have been done 5+ years ago.
(And it's not going to be your friend, and likely won't be a particularly enjoyable experience if it was your B-spec driver either.)

 

[Panoz]

Sometimes I find some comedic gold after watching a custom race replay for photos.

 

[eran0004]

Not at all the case, note that it's usually wise to not comment about how deep learning AI works if you don't know yourself...

Sophy as we know it right now is 5-60 year old AI tech (depending on how you're counting),
and it is honestly totally unreasonable for it to ever be a complete replacement for all AI.

And it has already learnt everything that it will ever learn, until the scoring criteria / model structure is changed and it is retaught.
It never "learns more" after at most a week of training: instead it has to learn from scratch after every tweak.

I don't see Sophy ever being implemented as a complete replacement.
Given this has been feasible to achieve on modern hardware for the past 10 years, why do you think no game developer has ever used a deep-learning AI for it's primary AI opponents?
Quick hint: the answer is not "because PD are so forward-thinking that they're gonna be first to do it!".

The concept of machine learning may be old, but the actual technology isn’t because there hasn’t been enough processing power and server capacity before to do what we’re able to do today. Claiming that Sophy is 60 year old tech is kind of like claiming that the moon landings were old tech because scientists and authors had been dreaming about it since the 1800’s.

Sure, the HARDWARE of ten years ago would have been able to perform the processes required to do machine learning, but the infrastructure required to make it happen did not exist, which is why no game developer has attempted a project like this before.

 

[Artremis]

GT7 seems to have received its first publisher discount in the UK. PS5 version now running at around £45, PS4 around £35 at most shops. Down from £65 and £55 respectively.

https://www.currys.co.uk/products/playstation-gran-turismo-7-ps5-10234767.html

https://www.currys.co.uk/products/playstation-gran-turismo-7-10234771.html

On sale on the PS Store starting tomorrow too. 30% off

 

[HarVee]

I expect to see some increase in players during this sale time, though I expect a bigger influx during the holiday season.

 

[DefNotJake]

These are the type of fun, light-hearted races I missed from the PS3 online era. Look Tat that grid diversity!

 

[GBalao888]

Nobody got the credits reward from predicting the Nations and Manufacturers' Cup winners yet?

I only guessed Subaru winning the Manufacturers' Cup correctly.

 

[linksracer]

Nobody got the credits reward from predicting the Nations and Manufacturers' Cup winners yet?

I only guessed Subaru winning the Manufacturers' Cup correctly.

I received 500,000cr for voting Miyazono in the Nations Cup, went into the rewards area of the garage to be redeemed. Haven't received anything for the Manufacturer's cup although I can't remember who I voted for and if they did well or not.

 

[Dibs]

For those who don't own the game yet, it's -25 % on PS Store

 

[A_Higher_Place]

Anyone having issues with liveries loading? Mine won't load in the garage

 

[Nuschel01]

The concept of machine learning may be old, but the actual technology isn’t because there hasn’t been enough processing power and server capacity before to do what we’re able to do today. Claiming that Sophy is 60 year old tech is kind of like claiming that the moon landings were old tech because scientists and authors had been dreaming about it since the 1800’s.

I think he meant 5-6 years, not 5-60

 

[snc]

Have you heard of OpenAI? Prior to GPT, DALL-E and GitHub Copilot, OpenAI made a few iterations upon a Dota 2 bot that learns in a roughly similar way to how Sophy does for GT7.
In a reduced feature set game of Dota 2, this bot could defeat professional players.
This bot was first demonstrated live in 2017.
(They later greatly expanded the feature set of the games, which reduced it's efficiency but it still held a 99.4% winrate over 42,729 games.)

You're probably right about a racing game being more complex than chess.
But ... Sophy really doesn't "play a racing game".
It has significantly reduced data passed to it, some of which could be considered cheating if a player had access to them.
Sophy also only plays "60 turn games". It only finds the best solution for the next 6 seconds at 10Hz.
It does not see the track, but instead how much track is to the left & right, lap distance & incline.
It does not see the movements of a car, but instead tyre load, tyre slip angle, angular velocities, car velocity & acceleration.
This comes to around 100 data inputs, and there are only 3 data outputs. (Throttle, brake, steer. It uses AT gears.)

A game like chess has ~400 different states every 2 moves.
AlphaZero must calculate every move until the end of the game, because a move in chess can invisibly cause disaster more than 100 turns later. There's a whole lot more possible moves too, which means a whole lot more data outputs.

Obviously, this doesn't mean making Sophy is "easy".
And you'll still need to be a research company to train it because you still need a few spare hundred-thousand $ supercomputers laying around.
But nothing Sophy does couldn't have been done 5+ years ago.
(And it's not going to be your friend, and likely won't be a particularly enjoyable experience if it was your B-spec driver either.)

Its obvious they have to simplify complex problem to have solution in realtime. And even if Sophy could be done 5 years ago (and in terms of software algorythm is rather not that long ago) it wasn't so yeah, still kudos for Sony and PD to make it happen (tough we still wait for gt7 implementation for gamers).

 

[DS13_supergt]

Anyone having issues with liveries loading? Mine won't load in the garage

Can you please explain what you mean for loading (some decals don't show up, gt auto doesn't load, etc..)

 

[A_Higher_Place]

Can you please explain what you mean for loading (some decals don't show up, gt auto doesn't load, etc..)

The game doesn't load any car with a custom livery in the garage. The home screen in the garage doesn't load if a custom liveried vehicle is chosen. Additionally, race C specifically doesn't load mine or other players' liveries correctly. My Gr3 jag ends up with a half purple half green livery instead of the custom GT2 one I have chosen.

 

[Greatbee]

The game doesn't load any car with a custom livery in the garage. The home screen in the garage doesn't load if a custom liveried vehicle is chosen. Additionally, race C specifically doesn't load mine or other players' liveries correctly. My Gr3 jag ends up with a half purple half green livery instead of the custom GT2 one I have chosen.

You are on ps4 or ps5?

 

[A_Higher_Place]

You are on ps4 or ps5?

PS5

 

[DS13_supergt]

The game doesn't load any car with a custom livery in the garage. The home screen in the garage doesn't load if a custom liveried vehicle is chosen. Additionally, race C specifically doesn't load mine or other players' liveries correctly. My Gr3 jag ends up with a half purple half green livery instead of the custom GT2 one I have chosen.

This os extremely strange, i had some minor problems, but on ps4 and i wouldn't know hkw to help.