Is Sophy 2.0 no longer competitive?

  • Thread starter rlx
  • 319 comments
  • 32,899 views
It's not a female name. Everyone seems to think it is, and it's often referred to as female, but it's not.

It's the Greek suffix meaning "knowledge". As in "philosophy" ("love of knowledge").
I'm going to message my kid now and let her know I want my grandson to be named Sophy.

I get what you're sayin'.
 
Folks, I know very well how it works.

What's funny is that this is pure marketing. Give it a female name, throw in some buzzwords like AI and Machine Learning (I wonder there's no Blockchain involved) and everybody is like "Oh, that's amazing".
Don't wanna spoil anyone's fun, if you like it that's great! I just think it's funny how things like this work those days.
You do realise machine learning and blockchain are completely unrelated technology terms, yeah?
 
I guess its simple, must try it when i get home, but the main goal of AI its to behave in a more natural way when racing the player, like in RL the middle field should be more chaotic and more mistakes from pilots while the cars in front should be a lot more competitive.

If people are not following the race rules, like same PP and same tyres compared to AI so the fault its on people not in AI implementations, thats why RL races have rules so everyone stick to them so cars can in a way be more balanced.

If people are racing AI with tunned cars, Racing tyres or cars that arent suit for those rules the blame its on people that dont understand simple rules...

If AI its trained to race the cars according those rules people must race using the same specs, if using the same specs compared to AI and the races are challeging and AI controled cars are more aware of ou presence and can defend/attack its a win.
 
Yes I do. But both are buzzwords used in marketing a lot lately.
OK, but what does that have to do with the price of tea in China?

Ultimately this is a machine-learning model as it has utilised many thousands of consoles attempting many thousands of laps so that the software can conduct trial and error and learn from itself. This is, by definition, machine learning.

If you spend a bit of time racing in close proximity to Sophy cars then you'll see what it means, as they behave very naturally in a lot of circumstances.

Utilisation of the term AI is a misnomer, as I've mentioned in other threads, as NPC racers in games have always been termed 'AI' despite not possessing any actual traits of being artificially intelligent. Whilst I agree they are not, this isn't used as a buzzword in this scenario, but as a colloquial terminology to express 'non playable competitive vehicles' in a term that everyone is familiar with.

So, to summarise:

  • not a girls name
  • not an attempt at buzzword bingo
  • not attempting to label it as AI specifically
  • not involved with the blockchain in any way

You just don't like it, and that's cool, but stop trying to crap on everyone's pancakes just because you're unimpressed by a technical solution to improve gameplay.
 
Last edited:
Basically, just trying to get rid of the choo-choo. If I was PD, it'd be a 'Senna Mode' next month.
 
Do people not understand how much work needs to be done .

Every individual car needs to be taught on that specific tire compound and a specific track and difficultly.

So for 1 car you need to train it three times per tire compound. ( Rain is included on rain tires)

THEN each car needs to be taught how to navigate on non rain tires tracks WITH rain to the pits . Easy solution is to default back to regular AI then switch to sophy AI once the car gets its rain tires on .

THEN for tuning to work PD needs to create tuning presets and train that car three times per tire compound on a track .

Now multiply this times the number of tracks and cars.

For simplicity Sophy can only use
Default tires and up too Racing hard tires + wet / intermediates

So comfort soft , Sport hard - soft , Racing hard - soft , wet and intermediates ( 9 tires)


That's 9 trained AI's + 3 difficulties .
27 variations per car per track .

Lets Say there will eventually be 400 Sophy cars

27 variations of a car X 400 cars 10,800 PER track .

Now multiply this by the amount of track layouts .

Now if you allow tuning and say add. Just 5 tune presets multiply that number by 5.

What PD should Do doing forward is use sophy for the carrier and only the Carrier , that makes their work so much easier.
 
I like to think of Sophy as a female. 😁
Ditto.

It's definitely a female's name, regardless of how it was derived - or truncated - in this case. If it was called Phil, we'd all think of it as a man's name. Regardless, I like thinking of the AI as female vs male so whatever... I'll leave you with this anyway. :D

feature-image-for-phils-philosophies-modern-family.jpg
 
Sophy has the potential for being something special in the future, but it’s not there yet and the reason is because once again PD nerfed it.

This is a comparison between my gr2 lexus and the AI on Suzuka. You can’t see it in the photo but the AI is in a slipstream and so am I. It’s the 6th position AI. The AI rpm is at maximum on 4th gear at 200km/h. I’m at 210km/h and I reached maximum rpm at 220km/h. We are both on RM tyres. The number 1 AI, also in a lexus revs at 205km/h on 4th gear but it had no slipstream.
 

Attachments

  • D16AFF81-8BFD-4B63-A03D-14F218786DEF.jpeg
    D16AFF81-8BFD-4B63-A03D-14F218786DEF.jpeg
    118 KB · Views: 9
  • 30C7328E-147E-4B88-90E6-9B6FAE91DF4F.jpeg
    30C7328E-147E-4B88-90E6-9B6FAE91DF4F.jpeg
    119.6 KB · Views: 16
Do people not understand how much work needs to be done .

Every individual car needs to be taught on that specific tire compound and a specific track and difficultly.

So for 1 car you need to train it three times per tire compound. ( Rain is included on rain tires)

THEN each car needs to be taught how to navigate on non rain tires tracks WITH rain to the pits . Easy solution is to default back to regular AI then switch to sophy AI once the car gets its rain tires on .

THEN for tuning to work PD needs to create tuning presets and train that car three times per tire compound on a track .

Now multiply this times the number of tracks and cars.

For simplicity Sophy can only use
Default tires and up too Racing hard tires + wet / intermediates

So comfort soft , Sport hard - soft , Racing hard - soft , wet and intermediates ( 9 tires)


That's 9 trained AI's + 3 difficulties .
27 variations per car per track .

Lets Say there will eventually be 400 Sophy cars

27 variations of a car X 400 cars 10,800 PER track .

Now multiply this by the amount of track layouts .

Now if you allow tuning and say add. Just 5 tune presets multiply that number by 5.

What PD should Do doing forward is use sophy for the carrier and only the Carrier , that makes their work so much easier.

It should be a lot less work than that.

The one thing I wouldn't expect to be possible (but I could be wrong) is to feed Sophy the complete geometry and tuning setup of the car as parameters and have it figure out the handling on its own. That would be a lot more car data than race data (location of surrounding opponents and track boundaries), so in a way, Sophy would have to learn more about car setups then about racing techniques.

But most of the rest should generalize quite well. Tyres can be one additional parameter with one of ~12 possible values. Lots of cars have similar performance characteristics (Gr.3, Gr.4, muscle cars, etc), as long as you keep them stock, so there's no need for retraining. And whatever Sophy learns about racing should generalize from one track to others. I can see why ovals might be a challenge, but in principle, once Sophy has seen all the tracks in the game, there is no reason why it shouldn't become capable of racing on any type of track layout or configuration.

Track surfaces (dirt, snow, rain) would be another parameter that Sophy needs to deal with, but while I can understand why that's not something Sony is focusing on in early iterations of the system, I don't see why it shouldn't be possible to add. As with most of the rest, this shouldn't require retraining from scratch.
 
Last edited:
You just don't like it, and that's cool, but stop trying to crap on everyone's pancakes just because you're unimpressed by a technical solution to improve gameplay.
Dude, I even wrote I don't wanna spoil anyone's fun. I just thought the way they are marketing this is funny and I guess I have the right to express my thoughts here. You have this right also ofc but I have the feeling that you misunderstood what I was trying to say.
Anyway, I will stop this discussion here and now.
Just came home from work, time to check out the update🙂
 
I think Sophie on Hard can be competitive only with a stock car using stock tires , without weight reductions or engine tuning.
Out of this , they move very smart , fast and fair , but impossible to match with a tuned car for a given PP level.
20 seconds in 3 laps against a 700PP car ..... no way .
They have to match the level or / and offer shuffle racing option again Sophie.
Other way , Sophie would be 100% useless.

F98cDfzWYAATfQo.jpeg
 
I think expecting Sophy to understand and apply tunings in order to be competitive with your car choice is a bit much. Heck there's 30% of the cars in the game it can't even drive yet.

Either use a racing class where the performance will be relatively equal or use a stock road car with Comfort tires. But you'll likely have the most competitive races with the Groups. It does feel like they dialed back Sophy's difficulty from the original preview period though.
 
Folks, I know very well how it works.

What's funny is that this is pure marketing. Give it a female name, throw in some buzzwords like AI and Machine Learning (I wonder there's no Blockchain involved) and everybody is like "Oh, that's amazing".
Don't wanna spoil anyone's fun, if you like it that's great! I just think it's funny how things like this work those days.
If you had played with the Feb preview build of Sophy, you’d have known that it was the real deal. Right now they’ve just scaled it down a bit too much.
 
Last edited:
No, and no that's not how it works.

The Sophy AI is trained on literally thousands of PS5s. We then get the "finished" versions of it (selected by Sony AI/PD for speed and racecraft) for the game.

Where? I've been covering it from the beginning and have no recollection of this claim being made.

It's always learning while it's being trained. The ones in the game are stable.
I don’t know where I saw or what was said. I just seem to remember something posted about it, that it was a self learning system. I don’t know how it works etc… so just assumed it was a system that worked off of players.

Not sure if this is what @Roggers was referring to but in the Feb 22 media event Kaz's presentation he talks about "how will GT Sophy appear in GT7" where he responds (via translator):



1. "GT Sophy will be there to coach the players of driving"
2. "the players can teach GT Sophy about sportsmanship and GT Sophy can appear as a friend in the races"

With a list of goals for GT Sophy on the slide:

  • can drive fast
  • appropriately handle various environments
  • appears to act naturally to people
  • behavior that exhibits meaning and intent
  • to be creative
  • to be ethical

So point 2 above implied GT Sophy could learn sportsmanship from players, but this didn't sound like a commitment just talking about what might be possible.

Could have been that, that I saw? I don’t really remember now as I’ve not really been following the developments much since that 25 year anniversary update when we had the option to race them.

Anyway hopefully Sophy improves over time.
Did a quick race at Spa today and I was able to win on professional in Gr3 cars and I’m a DR B/A driver. One thing I liked from the race was the competitiveness. The AI defending and having little looks up the inside was great. They can be a little dirty at times too I thought Sophy. As for race pace, they still have the issue of not been able to go flat through certain corners so have that shunt moment when your a lot faster
 
So I tried Spa on highest difficulty again, this time in a stock 911 (992) with power restricted to 70%, and still finished P2 without much of an effort.

Enjoyable race, but I feel a lot more challenged in a C lobby. If Sophy is an AI that beats world class drivers, then that's not the AI I'm racing against.
 
Last edited:
rlx
So I tried Spa on highest difficulty again, this time in a stock 911 (992) with power restricted to 70%, and still finished P2 without much of an effort.

Enjoyable race, but I feel a lot more challenged in a C lobby. If Sophy is an AI that beats world class drivers, then that's not the AI I'm racing against.
Are you still using soft tyres? You really need to use one of the cars the AI is using on the same tyres to get a proper comparison. Even down tuned the 911 (992) on soft tyres is going to be too quick.
 
Are you still using soft tyres? You really need to use one of the cars the AI is using on the same tyres to get a proper comparison. Even down tuned the 911 (992) on soft tyres is going to be too quick.
I'm on Sports Mediums, as implied by "stock 911". At 70% power restriction, my car should not be "too quick" -- definitely not in a race where each of my opponents is supposed to beat Kylian Drumont.
 
So we give you this amazing AI that has been trained via machine learning in tenths of thousands of hours!
It's amazing but it can't do this and it can't to that. It can't learn and it can't adapt to the players skills but hey it's really fast.....as long as you adapt to Sophy and use an appropriate car ofc.

Am I the only one who thinks this is really funny?
😄
Yeah the output of this collaboration 18 months into GT7 is a limited offering of an AI that a competent developer could've surpassed over 10 years ago. But with more emojis.
 
Last edited:
My 2c:

Is Sophy the answer to GT's absolutely ****** singleplayer AI, and is it fun to race against?

Yes.

Is Sophy implemented in a way right now that allows the user to scale difficulty or enjoy it to its full potential?

No.

A lot of naysayers on this initial limited roll-out. It's bizarre considering the majority of GT fans use this game as a car pokemon simulator, a singleplayer racing simulator, or a credit grinding simulator. You think there's be more enthusiasm for the forst non-braindead AI in 25 years.
 
rlx
definitely not in a race where each of my opponents is supposed to beat Kylian Drumont.
One of the points of the limited test was to establish a baseline for performance so they could pitch the difficulty correctly for the actual release. If the AI consistently runs faster than Julian Drumont then we'd all get our asses handed to us every time which makes for an equally frustrating game.

Not saying they got it right, and in an ideal world it would either scale to our pace after a couple of races or it would be selectable, but we are where we are.
 
Last edited:
One of the points of the limited test was to establish a baseline for performance so they could pitch the difficulty correctly for the actual release. If the AI consistently runs faster than Julian Drumont then we'd all get our asses handed to us every time which makes for an equally frustrating game.

Not saying they got it right, and in an ideal world it would either scale to our pace after a couple of races or it would be selectable, but we are where we are.

To get my ass handed to me by Sophy on highest difficulty is exactly what I would have wished for. As a rather average driver, I am rarely confronted with world class racing, so I'd have the opportunity to study replays and try to learn from Sophy. Or just admire what Sony/PD have achieved.

So while Sophy 2.0 is, obviously, a huge improvement over the vanilla AI, it still doesn't address one of the most glaring issues with singleplayer: the fact that the AI doesn't pose a challenge. And what is slightly frustrating about it is that Sony/PD made a conscious decision to keep it that way.
 
Back