Zhao Yili

Zhao Yili inquired about the main factors driving the faster expansion of Pony AI's operational areas beyond technology, and whether large language models (LLMs) are being used to advance L4 autonomy.

Answer

Tiancheng Lou, CTO, explained that Pony AI's L4 native tech architecture is inherently built for generalization, allowing rapid expansion into new areas like Shanghai Pudong and Shenzhen Nanshan within weeks without additional model training, as corner cases are consistent across regions. He noted that the speed of expansion is primarily limited by the number of robotaxi vehicles to maintain density. Regarding LLMs, Lou stated they are not suitable for L4 onboard driving models due to non-negotiable requirements for uncompromising safety (LLMs have 'model health' issues) and low latency. He added that LLMs rely on human data, which can introduce errors. However, Pony AI uses LLMs extensively in R&D for AI-enhanced human-machine interaction, engineering productivity tools, and rider feedback analysis.

Zhao Yili asked about the main factors behind the faster expansion of Pony AI's operational areas beyond technology, and whether large language models (LLMs) are being used and how they contribute to advancing L4 autonomy.

Answer

CTO Tiancheng Lou explained that Pony AI's L4 native tech stack is inherently built for generalization, allowing rapid expansion into new areas (e.g., Shanghai Pudong, Shenzhen Nanshan District in weeks) without additional model training, as corner cases are consistent across regions. He noted that the key to expansion is the number of robotaxi vehicles, as expanding too many areas without sufficient cars dilutes density. Regarding LLMs, he stated they are not suitable for onboard driving models due to critical requirements for uncompromising safety (LLMs have 'model health' issues) and low latency (LLMs are optimized for throughput, not low latency on low-power chips). LLMs also rely on human data, which limits them to existing human knowledge and can pick up human errors. Pony AI uses LLMs for R&D efforts like AI-enhanced human-machine interaction, engineering productivity tools, and rider feedback analysis, but not for the core driving model.