DoorDash 联合创始人讨论 AI 驱动的自然语言订购如何重塑食品和杂货配送，以及机器人和数据在智能体商务愿景中的作用。
DoorDash co-founders discuss how AI-driven natural language ordering is reshaping food and grocery delivery, and the role of robotics and data in their agentic commerce vision.
要点 · TL;DR
DoorDash 自 2018 年起从客户用例出发推进自动驾驶，打造了 DOT，并正迈向多模式配送车队。 DoorDash has pursued autonomy from a customer-use-case perspective since 2018, building DOT and heading toward a multimodal delivery fleet.
自然语言下单已在推动业务增长：半数餐厅搜索会尝试新商家，杂货购物篮规模扩大 40%。 Natural-language ordering already moves the business: half of restaurant searches try new merchants, and grocery baskets are 40 percent larger.
DoorDash 预计十年后人类 Dasher 数量会更多而非更少，因为需求增长需要所有配送方式共同参与。 DoorDash expects more human Dashers in ten years, not fewer, because demand growth requires every delivery modality.
核心观点 · Key points
DoorDash 自 2018 年起以实验和合作方式投入自主技术，之后从客户用例出发自研了 DOT 机器人。 DoorDash has invested in autonomy since 2018 starting with experiments and partnerships, then built DOT from a customer-use-case perspective.
自然语言下单已带来显著变化：餐厅搜索中一半是尝试新商家，杂货订单金额平均高出约 40%。 Natural-language ordering already moves the business: half of restaurant searches try new merchants, grocery baskets are 40 percent larger.
实体商业世界杂乱且微妙，DoorDash 数十亿次配送与交付点数据构成难以复制的优势。 Physical-world commerce is messy and nuanced, so DoorDash's billions of deliveries and drop-off data provide an unmatched edge.
未来是 Dasher、DOT、无人机和人行道机器人组成的多模式配送网络，统一由 DoorDash 本地商业平台协调。 The future is a multimodal fleet—Dashers, DOT, drones, and sidewalk robots—all coordinated through DoorDash's local-commerce platform.
自主技术跑通后，最难的规模化问题转向运营、硬件、制造以及衡量 AI 投入产出比。 Once autonomy works, the hardest scaling problems shift to operations, hardware, manufacturing, and measuring AI return on investment.
反共识 · Contrarian takes
DoorDash 原本押注语音是自然交互界面，结果发现对话式文本才能真正吸引用户。 DoorDash originally bet on voice as the natural interface, but found conversational text is what actually landed with users.
真正合适的配送机器人既不是人行道机器人，也不是 RoboTaxi，而是面向 3~5 英里配送的摩托车式自动驾驶车辆。 The right delivery robot is neither a sidewalk bot nor a robo-taxi, but an autonomous motorcycle-like vehicle for three-to-five-mile runs.
十年后 DoorDash 的人类 Dasher 很可能比今天更多，因为需求增长需要动用每一种配送方式。 Ten years from now DoorDash will likely have more human Dashers than today, because demand growth requires every delivery modality.
在物理世界 AI 中，单靠通用模型远远不够；先造技术再找问题，这条路是反的。 In physical-world AI, a general model alone is not enough; starting from technology and then finding a problem is backwards.
真实企业数据往往让在清洗后基准上表现出色的前沿模型失灵，暴露被干净测试集掩盖的数据分布差距。 Real enterprise data often defeats frontier models that ace scrubbed benchmarks, exposing distribution gaps hidden by cleaned test sets.
本期章节 · Chapters（共 6）
现实世界AI招聘Recruiting for real-world AI
自主配送规模化Scaling autonomous delivery
与Also的自主与车辆规模化Autonomy and vehicle scaling with Also