字节跳动还是下场了。
7月13日，据财联社、36氪等多家媒体报道，字节跳动正在探索进入自动驾驶领域。这个项目目前由字节Seed旗下周畅的世界模型团队负责推进，业务方向上，字节有意布局无人物流场景，隶属于字节旗下火山引擎汽车行业线。
这个消息一出，很多人的第一反应是"怎么又是自动驾驶？这个赛道不是已经死了一批公司了吗？"确实，过去几年自动驾驶经历了一轮泡沫破裂——Starsky Robotics倒闭、Uber ATG出售、Waymo估值大砍、Tesla FSD迟迟不能真正无人驾驶。连创始人都公开说"自动驾驶被高估了"。
但字节选择在这个时间点入场，而且是用世界模型团队来做，说明它看到的不是"又一个自动驾驶故事"，而是一个完全不同的机会。
世界模型做自动驾驶：降维打击？
理解字节的自动驾驶布局，关键在于理解"世界模型团队"这六个字。
传统自动驾驶公司（Waymo、百度 Apollo、小鹏 XNGP）走的是"感知-预测-规划-控制"的模块化路线：摄像头和雷达感知周围环境→预测其他车辆行人的行为→规划行驶路径→控制车辆执行。这个路线安全但笨重，需要海量的路测数据、高精地图、手工设计的规则，而且corner case永远处理不完。
世界模型的思路完全不同。它不是"识别物体然后做决策"，而是直接学习物理世界的运行规律，像人一样理解和预测环境。就像人开车不需要识别每一个物体才能决策——人靠的是对"这个场景下通常会发生什么"的直觉理解。世界模型如果能学到这种直觉，就能用更少的数据、更低的算力、在更多的场景下实现自动驾驶。
字节Seed团队在世界模型上本来就有深厚积累——Seedance视频生成模型本质上就是一个世界模型：它理解物理规律、理解物体运动、理解场景演化。把视频生成的世界模型能力迁移到自动驾驶，是一个非常自然的延伸。这是典型的"我手里有锤子，看什么都是钉子"——但如果你的锤子真的足够好，钉子确实能被砸进去。
第三条路径：不是车企，不是Robotaxi，是场景服务商
更值得注意的是字节的业务定位：放在火山引擎汽车行业线下，方向是无人物流。这说明字节走的不是前两条路。
第一条路是车企路线：代表是特斯拉、小鹏、蔚来——自己造车，把自动驾驶作为车的卖点卖给消费者。字节不会造车，这是明确的。
第二条路是Robotaxi路线：代表是Waymo、百度萝卜快跑、小马智行——做无人出租车运营，直接服务C端出行。这条路烧钱太狠，商业化太慢，而且需要跟各地政府拿运营牌照，不符合字节的风格。
第三条路是场景服务商路线：这正是字节要走的路——不造车、不做运营，而是把自动驾驶能力打包成解决方案，卖给物流、园区、港口、矿山等B端场景客户。无人物流是一个完美的切入点：场景相对封闭（仓库到仓库、园区内配送）、路线相对固定、成本敏感度高、人力成本上涨快、对24小时运营需求强。
"自动驾驶的终局不是'让所有人不用开车'，而是'在特定场景下替代人类司机'。谁先在垂直场景赚到钱，谁就赢了。"—— 一位自动驾驶行业资深人士
放在火山引擎下面也很聪明——火山引擎本来就是做B端云服务和AI解决方案的，自动驾驶能力可以直接成为火山引擎汽车行业线的一个产品模块，卖给车企、物流企业、供应链公司。不需要重新建销售渠道，不需要重新做客户关系，直接用现有的B端体系就能落地。
当然，字节的自动驾驶还在非常早期的探索阶段，离真正的产品落地还有很长的距离。但互联网巨头用世界模型的思路入场，本身就是一个重要信号——自动驾驶的竞赛，可能会因为AI大模型的进步，迎来一个完全不同的玩法。
这个赛道沉寂了两年，现在又有新玩家带着新武器进场了。
明天见。
ByteDance finally got in.
On July 13, multiple outlets including Cailian Press and 36Kr reported that ByteDance is exploring entry into autonomous driving. The project is being led by Zhou Chang's world model team under ByteDance's Seed division, sitting within the VolcEngine automotive industry line, with unmanned logistics as the likely target scenario.
The immediate reaction for many: "Autonomous driving again? Hasn't that track already killed a batch of companies?" Fair point. The past few years saw an AV bubble burst — Starsky Robotics folded, Uber ATG sold, Waymo's valuation slashed, Tesla FSD still can't do true driverless. Even founders publicly said "autonomous driving was overhyped."
But ByteDance choosing to enter now — and doing it with a world model team — suggests it sees not "another AV story" but an entirely different opportunity.
World Models for Driving: A Dimensional Strike?
To understand ByteDance's AV play, the key phrase is "world model team."
Traditional AV companies (Waymo, Baidu Apollo, Xpeng XNGP) follow a modular "perception-prediction-planning-control" pipeline: cameras and radars perceive surroundings → predict behavior of other vehicles/pedestrians → plan a path → control the vehicle. This route is safe but cumbersome, requiring massive road test data, HD maps, hand-crafted rules, and never-ending corner cases.
World models take a completely different approach. Instead of "identify objects then decide," they directly learn how the physical world works, understanding and predicting environments the way humans do. Just as a human driver doesn't need to classify every object to make decisions — they rely on intuitive understanding of "what typically happens in this scenario." If a world model can learn that intuition, it can achieve autonomous driving with less data, less compute, across more scenarios.
ByteDance's Seed team already has deep world model expertise — the Seedance video generation model is essentially a world model: it understands physics, object motion, scene evolution. Transferring that world model capability from video generation to autonomous driving is a natural extension. Classic "when you have a hammer, everything looks like a nail" — but if your hammer is good enough, the nail actually goes in.
The Third Path: Not Carmaker, Not Robotaxi — Scenario Service Provider
More notable is the business positioning: under VolcEngine's automotive line, targeting unmanned logistics. That signals ByteDance isn't taking either of the first two paths.
Path one: the carmaker route — Tesla, Xpeng, NIO — building cars and selling autonomy as a feature to consumers. ByteDance won't build cars; that's clear.
Path two: the Robotaxi route — Waymo, Baidu Apollo Go, Pony.ai — operating driverless taxis directly serving consumer mobility. This path burns cash ferociously, commercializes slowly, and requires government operating licenses in every city. Not ByteDance's style.
Path three: the scenario service provider route — this is ByteDance's play. Don't build cars, don't operate fleets; package autonomous driving capability as a solution and sell it to B2B customers in logistics, industrial parks, ports, and mining. Unmanned logistics is the perfect entry point: relatively closed environments (warehouse-to-warehouse, campus delivery), relatively fixed routes, high cost sensitivity, rising labor costs, strong demand for 24/7 operation.
"The endgame for autonomous driving isn't 'nobody has to drive' — it's 'replacing human drivers in specific scenarios.' Whoever makes money in vertical scenarios first wins."— An autonomous driving industry veteran
Putting it under VolcEngine is also shrewd — VolcEngine already does B2B cloud and AI solutions, so autonomous driving capability becomes a product module in the VolcEngine automotive stack, sellable to carmakers, logistics firms, and supply chain companies. No need to rebuild sales channels or customer relationships from scratch; it plugs straight into the existing B2B apparatus.
Of course, ByteDance's AV effort is still in very early exploration, a long way from real product deployment. But an internet giant entering with a world model approach is itself an important signal — the AV race could be in for a whole new playbook thanks to LLM advances.
The track went quiet for two years. Now a new player with new weapons has entered.
See you tomorrow.
"自动驾驶的终局不是'让所有人不用开车'，而是'在特定场景下替代人类司机'。谁先在垂直场景赚到钱，谁就赢了。"
—— 一位自动驾驶行业资深人士
"The endgame for autonomous driving isn't 'nobody has to drive' — it's 'replacing human drivers in specific scenarios.' Whoever makes money in vertical scenarios first wins."
— An autonomous driving industry veteran
ByteDance · autonomous driving · Seed · world models · Zhou Chang · VolcEngine · unmanned logistics · third path · scenario service provider · video generation transfer
