GT7 B-Spec Thread

[jmopborba]

So does this mean it is still coming to GT7? Looks likely.

 

[Bornhall]

So does this mean it is still coming to GT7? Looks likely.

Most likely it would mean “coming to GT7 on PS5 only”, just to be clear.

 

[Bojador]

But we have seen it working. I don't remember much about b-spec in older games, but in the GTWS they were able to tell Sophy when to push so this means they have it working to some extent and that would probably be good enough.

 

[kjb]

TBH if I could just have Sophy drive me and do it's own thing I would be cool with that. It would interesting to me if you just tell it how long you want it to drive then it goes does that. I know most people like to be able to give instructions but I would just like to see how it behaved on its own while racing for me. Would it try to keep the same pace as me if it is an endurance? I guess it would depend on how much time is left in the race. What if when the switch occurs and a lap down would Sophy try to gain that lap back? Just some things to ponder.

 

[Voodoovaj]

“I can’t say that it’s impossible, but there are a few hurdles involved in that,” Yamauchi explained. “In terms of controllability, the Sophy AI is a lot more difficult to work with. If you do B-Spec, obviously you want to direct the driver. That’s why it’s a difficult time to do [it] with the neural network.”

Oh for the love of all that is holy, WHAT ARE THEY TRYING TO DO?????? This is the worst possible direction for B Spec. The GT5 B-Spec was heading in the right direction.

B-Spec should lean into Sophy's learning. Give the player to tools/gameplay to train up Sophy drivers and create a team of Sophy's. Train up a team, like a bunch of Tamagotchi, and then let them loose with some vague orders. B-Spec should be more passive than active. Metal Gear 5 is a great example. You assemble an army, and manage the staff, but the army fights without you being present. You just send them into battle. It's a great meta game.

B-Spec should be 100% emergent gameplay, not directed game play. I don't want to be staying awake 14 hours into a 24 hour race to tell my Sophy driver to speed up or slow down!! I want them to know what to do! I want to build talent. I want to manage a team. Who wants to micromanage drivers? Then you'd just drive for yourself.

I'd give my left 🥜 to be given the opportunity to go direct this for them.

 

[tonidavid5]

B-Spec should be 100% emergent gameplay, not directed game play. I don't want to be staying awake 14 hours into a 24 hour race to tell my Sophy driver to speed up or slow down!! I want them to know what to do! I want to build talent. I want to manage a team. Who wants to micromanage drivers? Then you'd just drive for yourself.

You're spot on!!

C'mon PD!

You're giving us endurance races,... but don't include (yet) a proper in-race 'save' feature ... and are planning/working towards a B-Spec where we have to still manage the A.I. driver instead of being resting (virtually, there are other things like eating, working, even interacting with other humans not called Sophy 😅).

This thing needs some further thinking. Not only is not more 'realistic' racing (unless you are a 1-man team, owner, director, race engineer and driver 1, 2 & 3 feeding from Red-Bulls aplenty) but you are also telling us it's more complicated Sophy-wise.

 

[Bornhall]

Thing is, that is not how Sophy works, and naively thinking that we could train the model on-console within any reasonable time is folly. Sony AI (more than PD) knows what they’re doing, and no, you can’t just tell the model to “push harder” or “take it slow, save the tyres” without actually training the model - exactly as Kaz mentions.

That said, given extra parameters, you could analyse how much you as a driver pushes the limits (kerbs, oversteer and so on) and apply it to a trained model - but for each parameter the training will take longer and be more prone to errors.

I trust that we will get B-spec whenever they feel it is ready, regardless of how long it takes. Anything else would basically be shooting yourself in the foot in the public opinion.

 

[Voodoovaj]

Thing is, that is not how Sophy works, and naively thinking that we could train the model on-console within any reasonable time is folly. Sony AI (more than PD) knows what they’re doing, and no, you can’t just tell the model to “push harder” or “take it slow, save the tyres” without actually training the model - exactly as Kaz mentions.

Actually ya it is. You're thinking of training as going from 0-100. In reality, Sophy is already 100. B-spec should give you a "0" version, like what we see in the power pack. Sophy that is intentionally slowed. Then, the "training" that happens on console is to restore the drivers to their full selves.

You can break it down a myriad of ways by impeding Sophy on the straight, the corners, etc. You can introduce mistakes, collisions, etc. You build incompetence on top of Sophy, and then slowly remove it.

The WORST past of B-Spec was telling the driver to push harder, or slow down, the driver should KNOW to do this itself based on personality that you have "trained" into it. The game should be in the growth of the driver, not in the direction during a race. NO ONE likes micromanagement.

I want a field of Kimi Raikkonens

 

[tonidavid5]

@Bornhall

But that's exactly why I don't understand Kaz's comments. We don't need to be able to direct Sophy, we just want her to take over her duties when requested to drive the car on its stint so we can rest for a while.

Let her do her thing strategy wise (aka Sophy 3.0 theoretically already does this).

I know some people would always complain that the strategy followed by Sophy was not the one they were calling for, but, hey, it would be an awesome first big step to enjoy long endurance races. In the meantime, PD / Sony can work on the wrinkles to enable Sophy to be given some sort of direction in the future.

 

[Bornhall]

@tonidavid5 That is at least doable, but for most other ideas here, it’s simply various kinds of “nerfing” or “buffing” the car that the model is driving, as per what @Voodoovaj mentions above. Yes, you can improve models in various ways, reinforcing the behaviour of the model, but to introduce the things that people (seemingly) ask for, you would indeed have to introduce new parameters into the model, essentially at the very least a “safety” parameter (less or more risk taking rewards the model differently, even if it sometimes crashes but lap times are overall lower). And for that to happen, the model would need a lot of training, even as an “addition” to the model.

It’s definitely not impossible, but highly improbable. From my perspective, Kaz does know what he is talking about.

For anyone curious about how Sophy works (many seem to draw parallels to LLMs), I posed a number of relevant questions to ChatGpt, with regard to what is possible or not, and what can and can not be done with the regression learning Sophy is based on. You should be able to see the conversation here: https://chatgpt.com/share/69481f94-d8fc-8010-adaa-9103d8610184

 

[Voodoovaj]

@tonidavid5 That is at least doable, but for most other ideas here, it’s simply various kinds of “nerfing” or “buffing” the car that the model is driving, as per what @Voodoovaj mentions above. Yes, you can improve models in various ways, reinforcing the behaviour of the model, but to introduce the things that people (seemingly) ask for, you would indeed have to introduce new parameters into the model, essentially at the very least a “safety” parameter (less or more risk taking rewards the model differently, even if it sometimes crashes but lap times are overall lower). And for that to happen, the model would need a lot of training, even as an “addition” to the model.

It’s definitely not impossible, but highly improbable. From my perspective, Kaz does know what he is talking about.

People generally have a basic lack of understanding regarding what Sophy already does because few in the community spend enough time with the Sophy custom races. Sophy is already capable of doing everything that the community is asking for.

When you say "not impossible, but highly improbable", you should know that it's already in game. All B-spec needs is three things:

1 - Races that can be completed without the player participating. This is trivial since it's the basic of B-spec.
2 - "Inexperienced" B-spec drivers need to make more mistakes. 99% of this can be achieved by introducing variation in the braking zone. Sophy already varies in the braking zone, "inexperienced" B-spec drivers need a larger window of error. I would add that inexperienced drivers remove performance points from a car (behind the scenes), making them slower. 10-15PP should do the trick.
3 - "Experienced" B-spec drivers have smaller windows of error and they add to the PP of a car (again, behind the scenes). Again 10-15pp should do it, since this is what I use to make it happen now.

This is the state of the game right now. This is 15 minutes of me at Dragon's Tail. I spend the second half of the race fighting with "George Russell" for 4th, and he beats me. Watch as Sophy and I trade position. Sophy already does this throughout the field if you set it up with enough car variation to allow for it.



The only variation in this field is that I remove 1 point from either the ECU or power limiter for each finishing position in the 2025 F1 world championship. So, the "Mercedes" is 5 PP more than my "Ferrari"

 

[Bornhall]

@Voodoovaj that is possible, but as I mentioned, it is a virtual nerfing that I’m not sure will work in all situations. Same goes for tweaking for example the inputs to the model, say introduce a delay to what the model “sees”, reaction time and so on, but you run the risk of sending a trained model off the rails, yielding very unexpected results. Which is why I think they want to do it the right way rather than some temporary kludge that they may have to go back on at a later time.

 

[Voodoovaj]

@Voodoovaj that is possible, but as I mentioned, it is a virtual nerfing that I’m not sure will work in all situations.

Again, it's exactly what happens now. EXACTLY. If you set up a race, the drivers in the first 3 positions are "better" than the last 3 are "worse". These evolve during the race as positions change.

Same goes for tweaking for example the inputs to the model, say introduce a delay to what the model “sees”, reaction time and so on, but you run the risk of sending a trained model off the rails, yielding very unexpected results.

Again, this is already in game. Before they moved from 1.65 to 1.66, Sophy made these errors. It already dive bombs you if the situation calls for you. Variation is already there. Adding a little more is no big deal.

Which is why I think they want to do it the right way rather than some temporary kludge that they may have to go back on at a later time.

This is the right way. I would bet my life on it.

 

[Ryan81]

Really not bothered about the absence of B-Spec. Practically never used it.