Gran Turismo Sophy: Sony AI x Polyphony Digital

[Lord Protector]

From my testing on Suzuka with Porsche Cayman GT4 Clubsport (the same car that was available in the initial Sophy Preview), Sophy 2.0 is much slower than the Preview version. I can’t find any video on YouTube that shows her true potential from the Preview version without her being blocked by someone (best I found was Super GT’s video where he’s leading the whole lap with Sophy being on his rear bumper 24/7, his lap time was 2:07. If he didn’t block her the whole lap she would have set a much faster lap, at least by another second.)

In my test (same cars, same tires, stock settings) Sophy 2.0 was able to do mere 2:11.265. And that’s without me blocking her.

And it’s not just slower, it also shares the same stupid behavior treats with the regular AI, i.e. early braking, letting the player pass her without a fight, giving up positions, unwilling to pass… also, when you set a one-make custom race, all the Sophies drive in one line, no fights, just like ducks in a row… extremely disappointing.

 

[Tired Tyres]

Try Gr3 but don't turn Bop on. The cars have different speeds and Sophy has no issue with passing.

 

[Lord Protector]

Try Gr3 but don't turn Bop on. The cars have different speeds and Sophy has no issue with passing.

Are you saying faster cars go faster and slower cars go slower?

 

[Tired Tyres]

Are you saying faster cars go faster and slower cars go slower?

Yes. There's a lot of difference when Bop is not on.

Anyway I'm doing a 9 hour race at Tsukuba in a Skyline R34. There's 5 and a half hours to go and I'm taking a break. I'm 3rd one lap down on the two leaders, first is a BMW M2 Competition and there's a Porsche 959 second. I've got 26 seconds lead over an NSX. I'm playing in VR. It's getting dark and headlights have come on.

 

[ShiftingGears]

From my testing on Suzuka with Porsche Cayman GT4 Clubsport (the same car that was available in the initial Sophy Preview), Sophy 2.0 is much slower than the Preview version. I can’t find any video on YouTube that shows her true potential from the Preview version without her being blocked by someone (best I found was Super GT’s video where he’s leading the whole lap with Sophy being on his rear bumper 24/7, his lap time was 2:07. If he didn’t block her the whole lap she would have set a much faster lap, at least by another second.)

In my test (same cars, same tires, stock settings) Sophy 2.0 was able to do mere 2:11.265. And that’s without me blocking her.

And it’s not just slower, it also shares the same stupid behavior treats with the regular AI, i.e. early braking, letting the player pass her without a fight, giving up positions, unwilling to pass… also, when you set a one-make custom race, all the Sophies drive in one line, no fights, just like ducks in a row… extremely disappointing.

Current sophy is just piggybacking off the old AI to dimension it for different cars and tracks. It isn't the raw, Sony AI Sophy.

Its failed to reach its potential because it relies on the same team that's failed to implement improvements to regular AI, to implement Sophy.

Yes. There's a lot of difference when Bop is not on.

I think you missed the point

 

[Galvston]

Some people are just throwing affirmations around without proof.

Please turn BOOST WEAK on. If not, it still has slowdown

 

[whiteusagi]

I don’t know how to feel about this. Did a couple 10 lap races around Suzuka with 7x tire, 10x fuel using the Super Formula cars in complete default setups (All cars fitted with RS tires). With boost on weak, Sophy, when not in the lead, easily hits 300km/h on the straights without using either the overtake function or slipstream, whereas I am only capable of 285km/h or so. When leading, Sophy also tops out around the same speed as I, showing off just how advantageous this “weak” boost is. Before the tires wear too much, Sophy is also capable of easily taking the fastest lap at high 1:33s or low 1:34s, all while actively avoiding using any of the curbs (Not using the full width of the track), demonstrating a high reliance on the significant speed boost to keep pace.

On the other hand, with boost set to off, Sophy is STILL capable of easily beating my fastest lap with a low 1:37 in race trim, but for whatever reason starts to slow down unnaturally as the race progresses. To be more specific, Sophy starts hitting a maximum top speed of under 260km/h on the straights all while fully on throttle, and cornering in sector 1 well under 200km/h even without significantly degraded tires.

Yes, boost set to “weak” allows Sophy to turn laps much faster than otherwise, and in this way is able to keep pace with (or beat) the human player. However, I personally believe there needs to be more nuance as to how Sophy achieves those lap times, and simply charging down the straights whenever it is convenient for it is not what I personally want from the AI drivers.

 

[Theufcveteran]

I think it's just a difference of what each player wants. I have no problem with the AI cheating with overly fast straight line speed or more grip (grip really should be kept to a minimum though) if it means their lap times are competitive, but I totally understand why others prefer a more natural approach with no artificial speed.

For me, I just want fast, close racing no matter what the AI does. I would like to be able to break away if I'm quick enough, but if it's not possible without the AI being overall too slow, I don't mind the pack always being close.

With that said it seems like you can get some separation with slipstream battles between small packs of cars, which can help. And I haven't raced enough yet to see if you could break away over a longer race distance.

Of course, the less rubberbanding/boost, the better, and Sophy overall doesn't seem (?) to use grip hacks from what I've heard and seen so far. Just sometimes has more speed on the straight.

Hopefully they can get Sophy to a place where she has the current speed without extra straight line speed, but if it has to be this way vs slower but more natural AI, I prefer it with the extra speed.

 

[Mathaios_C]

Honestly i am quite disappointed. Nothing to do with the first version. Not as fast, not as clean, not as aggressive, not as much awerness as before. Boohhh what they are even doing???

 

[Theufcveteran]

Honestly i am quite disappointed. Nothing to do with the first version. Not as fast, not as clean, not as aggressive, not as much awerness as before. Boohhh what they are even doing???

It is possible that the first versions were hyper tuned for a specific car, and to do that for the whole car roster would take an enormous amount of time. I think it's quite good at spotting gaps, avoiding contact. Not quite as good perhaps as the initial version but I'm quite happy with it, it's certainly a big upgrade over the regular AI in terms of pace... but you are losing some unpredictability in one-makes particularly.

 

[Galvston]

It is possible that the first versions were hyper tuned for a specific car, and to do that for the whole car roster would take an enormous amount of time. I think it's quite good at spotting gaps, avoiding contact. Not quite as good perhaps as the initial version but I'm quite happy with it, it's certainly a big upgrade over the regular AI in terms of pace... but you are losing some unpredictability in one-makes particularly.

Hey, Ive found it doesn't cheat on the straights if its around other cars, the only time it rubberbands is if it is very far from the action (behind). Weird how it seems to vary. But I had 4 tests and reached the same results

 

[Meythia]

It is possible that the first versions were hyper tuned for a specific car,

If it was, then still it is.
Machine learning models dont "forget" what they once were instructed to learn, unless what they learn recently is valued higher and in contradiction to what was previously good.

 

[05XR8]

At least the best thing is many players are pretty much having equal results. Before this update, results varied much between car set up, whether a player detuned their own car versus AI, race set up, preferred tyre choices for treaded tyre racers and slick tyre racers.

We can all set up cars the way we intended and will have SOPHY remain competitive. Regardless of any artificial(go figure ) programming of light rubberbanding.
There’s still a place for Reggie™️ as I have to use that until SOPHY learns certain cars.

 

[Bornhall]

Still amazed at the armchair experts on AI and machine learning in this thread.

 

[Theufcveteran]

If it was, then still it is.
Machine learning models dont "forget" what they once were instructed to learn, unless what they learn recently is valued higher and in contradiction to what was previously good.

Thing is, I think Sophy is tuned/learned per-car.

The general driving characteristics might be there but it still struggles with some cars now, like the Countach - to get that kind of super speed like Sophy Play Together probably takes a lot of work per car.

It can't drive the Vantage, or any of the recent update cars after the January update. I imagine it will probably take some time so it can effectively drive the cars.

That's not to say they couldn't further improve the general driving ability, so that it could both handle a wider range of cars, faster and with better racecraft, in which doing so would mean less time tuning per-car for Race Together kind of speed and racecraft.

Although Race Together speed would have to be a new difficulty option, or make that professional, and make the current Sophy the middle difficulty.

 

[sonnyboychris]

Honestly i am quite disappointed. Nothing to do with the first version. Not as fast, not as clean, not as aggressive, not as much awerness as before. Boohhh what they are even doing???

Understandable.

Maybe if you look at it from the perspective that we're all participating in a beta test for GT8, you'll feel a bit more optimistic. This has never been advertised as a finished product. They're iteratively making it better in some ways and in other ways, it's seemingly regressing.

As evident by the posts on this forum, there's a wide swath of skillset and desired playstyles. I wouldn't be surprised if they're trying to fine tune Sophy to accommodate as many players as possible in preparation for a full launch in GT8.

 

[KAlex122]

I too long for a better experience.

But: I expected from the beginning, that tuning it right will be harder than creating perfect Sophy in the first place.

 

[GTRedman]

It is possible that the first versions were hyper tuned for a specific car, and to do that for the whole car roster would take an enormous amount of time

In one of the media events a few years back they said the following about the training of the original Sophy 1.0

• It takes 1 day to train an AI agent from knowing nothing to driving around a specific track
• 2 days to get to the top 5% of drivers
• 10-12 days to reach the limit of the best drivers
• within that time Sophy drives around 300,000km

 

[Met Entso]

Do you guys think that Sophy / sim racing AI will experience a similar quick development as LLM AI is experiencing now? In this case in GT8 in 2027-2028 we could have some spectacular AI.

 

[sonnyboychris]

Do you guys think that Sophy / sim racing AI will experience a similar quick development as LLM AI is experiencing now? In this case in GT8 in 2027-2028 we could have some spectacular AI.

I dunno... I assume whatever developments occur will be limited to the hardware of the base PS5 for the foreseeable future. Unless PD offers a more advanced Sophy for PS6 users, which is a possibility given they've done it for the PS4/PS5 userbase now.

 

[KAlex122]

Do you guys think that Sophy / sim racing AI will experience a similar quick development as LLM AI is experiencing now? In this case in GT8 in 2027-2028 we could have some spectacular AI.

I think it's already super (1.0). We just don't like everything about the dialled down versions.

 

[Mathaios_C]

I dunno... I assume whatever developments occur will be limited to the hardware of the base PS5 for the foreseeable future. Unless PD offers a more advanced Sophy for PS6 users, which is a possibility given they've done it for the PS4/PS5 userbase now.

I think ps5 has enough power to do a huge step forward compared to gt7 which is held back from the ps4 hardware. GT7 is born already old like Gt6 was at his time. Same horrible mistake