I watched it racing against top drivers, and I'm not sure people are going to love it. There's a fundamental problem with it not understanding the rules of racing that doesn't exist in most other types of game. For example, a lot of work has been put into DOTA2 AI, and I'm not aware there are any "sportsmanship" issues in that game, i.e. I think probably you can do anything the game lets you do. Driving games may well be unique in having rules that you are supposed to follow that the game doesn't enforce, which makes it tough to apply machine learning. Also, some of Sophy's techniques such as deliberately sliding the car on corner entry from a wide entry point with reduced grip will be a bit frustrating to race against, as it may be impossible for humans to replicate, given the latency in their internal feedback loop for applying corrections.
I'm not saying it can't be done, but IMO what I saw in those races wouldn't be 100% appropriate for a full roll-out into the game. There is a machine learning AI for Rocket League now, called Necto, that is pretty amazing. Its standard of play can't challenge top players yet, it's around top 10% standard, but the style of its play is fantastic for players around that standard to play with and against. It's quite mild-mannered, and doesn't use aggressive techniques that the game permits, such as demoing players, I'd guess that might have been achieved as a deliberate thing through the choice of metrics for reinforcement. I'd say Sophy needs a similar sort of attention to the style of how it drives and competes for places, rather than purely focusing on outright pace.
