It has been a while since we’ve heard anything about Gran Turismo Sophy, the all-new artificial intelligence system developed by Sony AI in collaboration with Polyphony Digital. After making its debut as part of a limited-time “Race Together” special event in Gran Turismo 7 update 1.29 back in March, the Sony AI team has been quietly evaluating all of the data they collected and figuring out how to make the new agent better.
GTPlanet had a chance to speak exclusively with Dr. Pete Wurman, Director of Sony AI America and GT Sophy project lead, to find out more about what his team learned, what they thought about all the feedback shared by Gran Turismo 7 players, and how they are aiming to improve it in the future.
Read the transcript below, or watch the video above for GTPlanet’s full analysis and discussion of all of Dr. Wurman’s responses.
![](https://www.gtplanet.net/wp-content/uploads/2023/02/Peter-Wurman-Sony-AI-800x450.jpg)
GTP: What were some of the more interesting data points that you collected with Sophy’s debut in Gran Turismo 7?
Dr. Wurman: Releasing Gran Turismo Sophy in GT7 allowed us, for the first time, to evaluate how well the neural network performed when embedded in the game. We learned quite a bit about the specific engineering tasks involved in making the agent fully integrated, as well as the game design itself. GT Sophy is such a good driver that it was challenging to find combinations of cars that made the races fun for a variety of skill levels.
GTP: What are your thoughts on the various reactions of the public to Sophy? What do you think the general consensus was from players? Did any comments or opinions surprise you?
Dr. Wurman: We saw a lot of very positive comments by players about GT Sophy. They voiced strong support for rolling out the technology throughout the campaign mode. Many said that it provided a game-changing racing experience and that racing GT Sophy resembled the best experience they had racing other humans in the online mode, but without the hassles. Many also requested the ability to turn down or modulate the skill so that they could find a level that matched their playing skill.
![](https://www.gtplanet.net/wp-content/uploads/2023/02/RaceTogether_in-race-1-800x450.jpg)
GTP: What do you believe are the biggest misconceptions that users have about Sophy or how it works?
Dr. Wurman: A lot of players understood the technology quite well, but there were also quite a few misconceptions. For one, theories were shared about how easy it should be to make a slower version of GT Sophy. We’ve been working on that, and it turns out to be a lot harder than it may initially seem to make an agent that is slower in a real human way.
It was also interesting how much the players anthropomorphized the agent, attributing intentions or desires to the agent’s behavior when it was really just reacting to the situation at hand. Of course, our goal is to create an experience that mimics racing against good, clean human drivers, so, in a sense, having players describe the agent in human terms was a testament to how well it worked.
GTP: Have you made any changes to the direction of Sophy’s research since this first public test as a result of the public’s reactions or experiences?
Dr. Wurman: We’ve been looking at some of the anomalous behaviors that were reported to see if we can improve the training. We’ve also been working with PDI to make a version of the agent that can drive many more cars and be used in more flexible gameplay scenarios to give players more control over the races.
![](https://www.gtplanet.net/wp-content/uploads/2023/02/RaceTogether_replay-1-800x450.jpg)
GTP: The progress of AI research in general has rapidly accelerated over the past year. How have these changes impacted Sophy’s development? Are you using any new techniques or technologies to further refine it?
Dr. Wurman: The Gran Turismo Sophy ‘Race Together’ mode was one of the first times a system trained through reinforcement learning has been delivered into a production environment. This demonstrates that this technology can provide robust solutions to complex domains.
The rapid advances in generative AI are certainly something we watch closely, but we haven’t changed our approach. After all, you can ask ChatGPT how to drive a racecar fast, and it will give you reasonable advice like “hit the apex of the corners.” But ChatGPT can’t actually control the car, give you specific instructions for specific racing scenarios, or tell you whether a racing incident deserves a penalty.
GTP: The first public test for Sophy was only available on PlayStation 5 consoles; can PlayStation 4 users have any expectations to see Sophy in the future?
Dr. Wurman: We do not have anything to share at this time.
See more articles on Gran Turismo Sophy and Sony AI.