Using a chatbot AI to more efficiently teach another AI... it's a bit odd. I watched the series about the creation of Sophy a while ago, and the very clever programmers involved seemed quite proud of their work in guiding this racing AI to be so fast and adaptive. In this case: making hundreds, and eventually a limitless amount of customised driving styles, I concede that it can't be done by people. Or... well, people aren't really doing it anyway, deep reinforcement learning is strange like that.
Paraphrasing a very long video I made, Sophy was taught to drive in a way equivalent to how a dog learns an agility course: a clear goal and a handful of treats. There are differences in that dogs can't read their split times nor fully understand clean overtaking, but that's the gist of it.
There've already been quite a few types of Sophy: the superhuman Sophy that beat GTWS drivers at Le Mans (by taking extremely aggressive, almost illegal and very unusual racing lines with nearly the accuracy of a tool-assisted speedrun), the four colours of the Race Together event Sophy (Violette was a particularly sharkish driver), Sophy "devious grin emoji" 2.0, and more relaxed, but still very good at overtaking Power Pack Sophy. Exactly how much human work was going into making Sophy variants with different paces, I'm not sure. The way it was presented made it seem that once it had become the fastest bot in the world, it was (not so simply) a matter of telling it to aim for more reasonable lap times on other difficulty settings (which still requires training time). I've witnessed Power Pack Sophy having a sugar rush and putting in a lap 2 seconds faster than its own pole time, then doing the next lap 6.5 seconds slower once it realised it was pulling too far away, and I'm not sure if that was "Boost" or sandbagging.
If AI can turn a text prompt for how to drive into new instructions, then into new weights for the model, great! I don't know how well a prompt can generate a driving personality, though. Each agent would be different, but exactly how different would be open to interpretation and a lot of confirmation bias. "I told Sophy to drive recklessly, like Burt Reynolds, and I can see the resemblance!"
Something like "Drive like Sophy v4, but take the risky overtake option 30% more often" could make a virtual personality for a future Gran Turismo AI opponent, I suppose, but sadly not "Drive like (famous racing driver)".
How would an LLM even describe a driver to Sophy? If it only gleans the web for written descriptions of driving styles, it seems pointless. Would Sophy Michael Schumacher crash into Sophy Damon Hill to maintain championship standings? Not from a prompt. Anything less than turning real, recorded laps into in-game telemetry to study (somehow) wouldn't really capture a person's driving style. Still, hopefully this leads to something fruitful, and
not Sony AI's humans being made redundant.
Sophy is already machine learning, i wouldn't be against using AI tech in back end stuff like this, it seems to be a "harmless" use of the tech.
I would be out in a heartbeat if PD started using Gen AI in their car models tho.
Don't even wanna think about it
I think PD already use Houdini and/or other procedural generation tools like it for some environment details (like the road markings on Grand Valley and Trial Mountain, background buildings in Tokyo). Unless some Disneyesque executive greed intervenes, I'd say the world's best car modelling team has secure employment and is going to keep doing things their way.
