Ex-Li Auto Exec Launches Kunlunxing: Embodied AI Prioritizes Understanding over Imitation
量子位 · wechat · 2026-08-15
QbitAI interviews Lang Xianpeng (former core head of Li Auto's intelligent driving), discussing his journey from Baidu and Li Auto to founding an embodied AI startup.
Key Insights and Experiences:
- Cognitive Iteration: Realized at Baidu that data is more critical than model structure; experienced the full paradigm shift from HD maps to map-less, then to end-to-end and VLA at Li Auto, achieving mass production for 1.5 million vehicles.
- Startup Opportunity: Views autonomous driving as a starting point, not the end goal, for embodied intelligence. The VLA framework is essentially designed for embodiment, which offers a higher ceiling. Partnered with Huawei's Ren Geng, valuing complementary skills and aligned cognition.
- Technical Route: Advocates "understanding over imitation," adopting a "data compilation" route to train physical causal world models rather than relying solely on imitation learning. Uses an Agent framework for engineering deployment instead of piecing together software and hardware.
- Industry Judgment: Believes embodied AI is currently similar to autonomous driving in 2016, with industry consolidation beginning after scale delivery. Must start with ToB before entering ToC households, emphasizing safety and mass production capabilities.
More from Companies & People
- Observation: Meta Shifts from Social Graph to Video-Ad-Attention Graph — johnmccrea · 2026-08-15
- Alexandr Wang: Data is the invisible ceiling for AI systems — r0ck3t23 · 2026-08-15
- Google's AI Exodus Signals a Shift in Capital and Compute Allocation — Exponential View (Azeem Azhar) · 2026-08-15
- Indie Developer's Overseas Company Turns 2: Tax-Free Last Year, Big Tax Bill This Year — yihui_indie · 2026-08-15
- Scholar Ends Sabbatical at Cornell, Thanks Hosts — windx0303 · 2026-08-15
- Dev builds AI app to help himself become a better friend — aaraalto · 2026-08-15