HyprLabs在旧金山用两辆被改装的Tesla Model 3进行了约一年半的路测，每辆车加装了5个额外摄影机，并搭载一台手掌大小的超级电脑。它想回答一个核心问题：在AI加速的当下，一家公司究竟能多快打造可用的自动驾驶软体，以及能否把「开得不错」推进到「比人类更安全」。

这家公司首次公开行动：团队共17人，其中只有8人是全职，分布在巴黎与旧金山；由Zoox共同创办人Tim Kentley-Klay领导，他在2018年离开该公司。Hypr自2022年以来仅募得550万美元，却宣称其Hyprdrive训练方式能降低成本与人力投入，并试图在产业回温（robotaxi扩城、车厂再度许诺量产自驾）时取得后发优势。

Hypr主打「run-time learning」：以transformer模型为起点，在人类监督下边开边学，只把新奇资料回传「mothership」微调，再把变更的部分（deltas）下发回车端。两辆车合计只搜集4,000小时、约65,000英里的数据，其中约1,600小时（约40%）用于实际训练；相较之下，Waymo在十多年中累积了1亿英里的全自动驾驶里程。Hypr承认尚未达到上路营运的安全与量产水准，但声称以极少的计算工作展现了令人印象深刻的驾驶能力，并计划在明年推出一款非传统机器人。

HyprLabs has spent about the last year and a half testing two modified Tesla Model 3 sedans in San Francisco, each fitted with five extra cameras and a palm-sized supercomputer. Its core question is speed: how fast a company can build autonomous-vehicle software today, and whether AI-driven methods can move from “driving pretty well” to “driving more safely than a human.”

The company is going public with its work for the first time, led by Zoox cofounder Tim Kentley-Klay, who left Zoox in 2018. Hypr has 17 people total, only 8 full-time, split between Paris and San Francisco, and has raised just $5.5 million since 2022. It is pitching Hyprdrive as a cheaper, lower-labor training leap timed to an industry rebound as robotaxis expand and automakers renew self-driving promises.

Hypr’s approach, “run-time learning,” starts with a transformer model that learns while driving under human supervision, sending only novel data back to a “mothership” for fine-tuning and pushing only the changed pieces (deltas) back to the cars. Across two vehicles it has collected 4,000 driving hours (about 65,000 miles) but used roughly 1,600 hours—about 40%—for training. That is tiny next to Waymo’s 100 million fully autonomous miles over a decade-plus, and Hypr says it is not yet production- or safety-ready; it argues the point is an unusually strong signal achieved with very little computational work, ahead of a planned new robot next year.