加州新创 Sabi 正在开发一款脑机介面穿戴设备，形状像保暖帽，目标是将人的内心语音解码为萤幕上的文字。其 CEO Rahul Chhabra 表示，首款脑读取 beanie 将在今年年底上市，并且也在设计棒球帽版本。与 Neuralink 等著重植入式脑晶片、供重度运动障碍患者使用的方案不同，Sabi 的路线是非侵入式可穿戴，Vinod Khosla 认为若要让数十亿人每日接近广泛使用，技术必须是非侵入式、可大众化的。

Sabi 的感测器采用头皮表面的 EEG（脑电图）金属圆盘，这让讯号必须穿过头皮与骨骼后再解码，较植入式设备的神经讯号品质低。为提升准确率，Sabi 以高密度感测设计弥补这一缺点：一般 EEG 仅有十几到数百个感测器，但其帽体计划部署 70,000 到 100,000 个微型感测器。公司初步目标是每分钟约 30 字（words）输入，虽低于一般打字速度，但 Chhabra 预期随著使用熟悉度会提升。另一步大挑战是同一句想法在不同人脑中的发放模式有差异，故 Sabi 正在建构「脑基础模型」以跨使用者泛化。

Sabi 已搜集 100 位受试者共 100,000 小时的脑讯号资料，作为训练这类跨人模型的基础。JoJo Platt 指出，消费级 BCI 在实际采用上必须做到即插即用，不能每次都重新校准，因为疲劳、专注度会让讯号日变。为了日常穿戴，也必须兼顾美观与舒适，以免像医疗设备般突兀。资料安全方面，Chhabra 表示装置外传至云端时会端到端加密，AI 可在加密资料上训练并避免直接读取原始神经资料；公司也正与史丹福等神经安全专家共同审核其技术堆叠，以强调神经资料的高度隐私性。

California startup Sabi is developing a wearable brain-computer interface in the form of a beanie to decode internal speech into on-screen text. CEO Rahul Chhabra says the first brain-reading beanie will launch by the end of the year, with a baseball-cap version also planned. Unlike implant-focused firms such as Neuralink, which target severe motor disabilities, Sabi pursues a noninvasive consumer path, and investor Vinod Khosla argues that only a noninvasive approach can scale to daily use by billions of people.

The system uses scalp EEG, so brain signals must pass through skin and bone before being decoded, which weakens them compared with implanted devices. Sabi’s response is extreme sensor scaling: most EEG systems use dozens to a few hundred sensors, while the cap is planned to include about 70,000 to 100,000 miniature sensors to improve spatial precision and decoding reliability. The company targets about 30 words per minute initially—slower than average typing—but expects speed gains with user adaptation. Because imagined-speech patterns vary across people, Sabi is building a “brain foundation model” trained on many users rather than person-specific models.

Sabi reports 100,000 hours of brain data collected from 100 volunteers. Independent consultant JoJo Platt says consumer BCIs must be ready out of the box, since fatigue and focus shifts change neural signals and would make per-session calibration impractical. Comfort and low visibility are also important, because wearables must be practical for daily life. On privacy, Chhabra says neural data is uploaded only in end-to-end encrypted form, and models are trained on encrypted data rather than raw neural traces. The company is also working with neurosecurity experts, including groups at Stanford, to audit security across its stack.