人工智能的下一个前沿领域是人脑

如果不了解脑机接口，就无法了解人工智能的长期未来。

为什么会这样呢？因为脑机接口（BCI）将在定义人类智能和人工智能如何在拥有强大人工智能的世界中融合方面发挥核心作用。

对大多数人来说，脑机接口听起来像是科幻小说。但这项技术正在迅速成为现实。脑机接口在实际功能和应用方面正接近一个转折点。虽然听起来有些不可思议，但像心灵感应这样的能力很快就会成为现实。

脑机接口（BCI）领域可以分为两大类：侵入式方法和非侵入式方法。侵入式BCI方法需要手术，即将电子设备植入颅骨内，直接植入大脑内部或表面。而非侵入式方法则依赖于位于颅骨外部（例如耳机或帽子上）的传感器来解读和调节大脑活动。

在本系列文章的第一部分（10 月份发布）中，我们深入探讨了侵入式脑机接口技术和相关初创公司。在本文中，我们将重点关注非侵入式脑机接口。

脑机接口（BCI）和人工智能（AI）的结合，将在未来几年重塑人类和文明。现在正是我们认真关注这项技术的时候。

传感器的宝库

在深入了解当今非侵入式脑机接口 (BCI) 初创企业格局之前，让我们先花点时间探索一下使非侵入式脑机接口成为可能的核心技术。

无论你用大脑做什么——思考、阅读、说话、移动手臂——大脑内部都会发生一些可感知的物理事件，并遵循一定的模式。具体来说，信息通过微小的电脉冲在大脑神经元之间流动：这种基本的物理力也驱动着灯泡、厨房电器和iPhone等设备。这些微小的电信号还会触发大脑中的其他物理活动，包括磁场和血流的变化。

这些物理变化最终代表着信息。它们的模式编码着思想、概念、语言和行为。而编码的信息可以被解码。这正是脑机接口的目标。

为了以不同的方式解读（“读取”）和调节（“写入”）大脑的物理活动，人们开发了多种不同的非侵入式传感器。每种传感器都有其优势和劣势。为了理解非侵入式脑机接口（BCI）领域，必须了解这些不同类型的传感器（也称为“模态”）及其工作原理。

世界上最古老的脑电传感器是脑电图（EEG）。脑电图于1924年在德国发明，如今仍然是世界上应用最广泛的脑电传感器。

脑电图（EEG）通过放置在头皮上的小型电极直接测量大脑的电活动。（电极是一种可以检测电活动的简单装置。）脑电图在时间精度方面非常高：它可以以毫秒级的精度测量神经元活动。此外，它还具有价格低廉、便携、安全且易于使用等优点。

脑电图最大的缺陷在于其空间定位的不精确性。大脑的电信号在穿过颅骨和头皮到达脑电图电极的过程中会发生严重的失真，导致难以精确定位其在大脑中的起源位置。这是因为颅骨和大多数骨骼一样，导电性很差。

此外，脑电图测量的信噪比很低，因为大脑微弱的电脉冲很容易被附近许多其他电活动源所掩盖，例如：咬紧牙关、心跳，或者仅仅是环境电磁干扰。仅仅是眨眼就能产生比大脑电信号强10到100倍的电活动。

因此，从脑电图噪声数据中提取足够高保真度的信号，一直是脑电图应用于脑机接口技术的一大障碍。

另一种非侵入式脑机接口技术在这些方面远优于脑电图：脑磁图（MEG）。

你可能还记得高中物理课上讲过，电和磁是同一自然现象——电磁学——的两个统一表现形式。因此，当神经元放电并产生微弱的电信号时，它同时也会产生微弱的磁场。脑电图（EEG）测量的是电信号；脑磁图（MEG）测量的是与之相关的磁场。

与电场相比，磁场最显著的特点是它几乎可以完全不受干扰地穿过颅骨和头皮。因此，MEG的空间分辨率和定位精度远高于EEG。

有什么猫腻？

如今的MEG系统体积庞大，需要磁屏蔽腔和低温冷却，耗资数百万美元。这使得它们对于日常脑机接口应用而言根本不切实际。

但目前正在进行一些前景可观的研究，旨在使MEG系统更小更便宜。一种基于光泵磁力计（OPM-MEG）的新型MEG展现出巨大的潜力：它可在室温下工作，体积小巧，可以佩戴在头部，而且所需的屏蔽强度也较低。

OPM-MEG技术尚未成熟，但未来几年它有望成为一种重要的新型脑机接口技术，在避免侵入性手术的同时，提供比脑电图（EEG）更高保真度的脑部数据。

第三种值得一提的非侵入式脑机接口技术是功能性近红外光谱技术（fNIRS）。

与脑电图（EEG）测量电活动或脑磁图（MEG）测量磁活动不同，功能性近红外光谱（fNIRS）测量的是脑血流量。神经元放电时，血流量会增加，因为放电的神经元需要更多的营养物质。fNIRS传感器通过颅骨向大脑发射高波长光束，可以检测脑血流量的变化，并利用这些变化模式来解码脑活动。

近红外光谱成像（fNIRS）如今已成为全球第二大最常用的非侵入式脑机接口（BCI）传感器，仅次于脑电图（EEG）。这很大程度上要归功于布莱恩·约翰逊（Bryan Johnson）创立的初创公司Kernel在过去十年中所做的努力。Kernel的关键成就在于实现了fNIRS技术的微型化，首次将其转化为可穿戴设备，并实现了规模化商业化。与EEG一样，fNIRS安全、便携且价格相对低廉。fNIRS在定位方面比EEG更精确，但在时间精度方面则不如EEG；因此，这两种成像方式互为补充，并经常结合使用。

这就引出了目前最热门、最有前景的非侵入式脑机接口技术：聚焦超声。本文将更详细地探讨超声技术。请继续阅读！

要了解非侵入式脑机接口（BCI）领域的最新进展——哪些技术可行，哪些技术不可行，以及未来最大的机遇在哪里——最好的方法是探究当今领先的初创公司正在做的事情。让我们深入了解一下。

利用脑电图读取思想

一群低调的初创公司认为，不起眼的脑电图 (EEG) 有望从一种为人熟知但功能有限的传感器转变为脑机接口 (BCI) 的主流方法。

脑电图有很多优势。然而，几十年来，人们普遍认为脑电图的信号质量太差，无法支持先进的脑机接口功能。

那么，现代人工智能的一大优势就是它拥有超乎人类的能力，能够从嘈杂的数据中提取潜在信号，这真是太方便了。

如果你是一位铁杆深度学习信徒——一位“苦涩教训”的极端主义者——那么选择脑电图作为你的脑机接口（BCI）模式是有充分理由的。一言以蔽之：规模优势。

当前人工智能时代的特征在于规模化原则。OpenAI 在 2020 年普及了“规模定律”的概念：人工智能系统会随着训练数据、模型规模和计算资源的增加而稳步提升。此后五年间，人工智能的飞速发展主要归功于规模的全面扩展。大型语言模型之所以如此强大，是因为我们已经掌握了如何利用人类历史上几乎所有书面文本来训练它们的方法。

如果想把在生成式人工智能领域行之有效的策略应用到理解人脑，关键在于尽可能多地收集脑训练数据。而要收集尽可能多的脑训练数据，最佳传感器的选择显而易见：脑电图（EEG）。简而言之，脑电图比任何其他脑机接口（BCI）模式都更具可扩展性。

如今，全球脑电图（EEG）系统的数量比其他所有脑机接口（BCI）传感器加起来还要多几个数量级。世界上大多数医院都配备了脑电图设备；相比之下，全球范围内功能性近红外光谱（fNIRS）系统可能只有几千套，脑磁图（MEG）系统也只有几百套。基础型脑电图系统的价格不到1000美元。

Conduit是一家年轻的初创公司，它以人工智能为先导、规模化为先导，致力于开发非侵入式脑机接口（BCI）。这家公司由一位牛津大学的年轻研究员和一位剑桥大学的年轻研究员联合创立，旨在以最快的速度收集尽可能多的数据，以训练一个大型的大脑基础模型。该公司表示，到今年年底，他们将收集到来自数千名参与者的超过1万小时的脑电波记录。

虽然 Conduit 主要专注于收集脑电图数据，但它也通过其他非侵入式方式进行补充，因为该公司发现，如果使用来自每个用户的多种传感器方式进行训练，其人工智能的性能会显著提高，而不是仅仅使用一种传感器方式。

Conduit 设想其技术有哪些应用场景？

令人惊讶的是，该公司的目标是打造一款脑机接口产品，能够在用户将想法转化为语言之前就解码他们的想法。换句话说，他们正致力于开发意念转文本人工智能。

据该公司称，该系统已经开始运行。Conduit 目前的 AI 模型生成的文本输出与用户的想法在语义上匹配度约为 45%，而且无需事先针对任何特定个体进行微调即可实现。

举几个具体的例子能让这一点更具体一些。

例如，当一位参与者想到“房间似乎更冷了”这句话时，人工智能生成了“有微风，甚至一阵轻柔的风”。在另一个例子中，参与者想到“你有没有最喜欢的应用程序或网站”，人工智能生成了“你有没有最喜欢的机器人”。

这项技术尚未成熟，无法投入市场。45% 的准确率对于大众市场产品来说远远不够。而且，目前只有用户在头上佩戴一套笨重的传感器才能达到这样的准确率。但考虑到这项技术的目标是读取人心，这样的准确率仍然令人瞩目。而这家公司才刚刚起步。Conduit 公司几个月前才开始扩大数据收集规模；该公司计划未来将其训练数据集扩大几个数量级。

想象一下，如果仅仅通过思考就能将细微的想法传达给其他人或计算机，那将会有多么大的可能——社会将会发生怎样的变化。

Conduit联合创始人里奥·波普尔表示：“过去十年机器学习领域给我们带来的最大教训是规模和数据的重要性。与所有数据集中的个体都必须先接受脑部手术的情况相比，非侵入式方法使我们能够收集到更大、更多样化的数据集。”

她的联合创始人克莱姆·冯·施滕格尔补充道：“我们创立Conduit是因为我们意识到，如果我们都直接用想法而不是语言思考，人们就能更快地完成事情。而且我们也能更深入地了解彼此以及整个世界。”

另一家有趣的年轻创业公司 Alljoined 也在不断突破脑电图技术的应用极限。

与 Conduit 类似，Alljoined 也采用了以人工智能为先导的非侵入式脑机接口 (BCI) 技术，并押注脑电图 (EEG) 是合适的模态，因为它具有可扩展性和易用性。Conduit 的目标是将想法解码为语言，而 Alljoined 的初期重点是将想法解码为图像——也就是说，根据脑电图读数忠实地再现用户“脑海中”的图像，这项任务被称为图像重建。

Alljoined 的首席执行官兼联合创始人 Jonathan Xu 是开创性论文MindEye2 的合著者之一，该论文表明，基于生成式人工智能的方法仅需少量 fMRI 数据即可实现精确的图像重建。Alljoined 致力于将这项工作从 fMRI 扩展到 EEG 数据，并且已经取得了成功。

下图展示了Alljoined人工智能系统利用参与者脑电图数据重建的部分图像示例。正如您所见，重建结果并非完全精确，但这些结果代表了目前最先进的性能。而且——正如我们在人工智能的许多其他领域所观察到的那样——随着训练数据和计算规模的扩大，系统的性能必将持续提升。

上排代表人类参与者观看的图像，下排代表 Alljoined 的 AI 系统根据参与者的脑电图数据重建的图像。

来源：Alljoined

说到训练数据，Alljoined 去年开源了首个专门用于脑电图图像重建的数据集。该数据集包含 8 位参与者的脑电图数据，每位参与者观看 10,000 张图像。免费提供这些数据应该会极大地推动整个领域的发展。

Alljoined 最初专注于图像重建，但该公司也在探索其他应用领域。其中一个极具前景的领域是情感分析——即实时、精准地识别用户正在经历的情绪。直接从脑电数据中解码情感具有重要的商业价值，例如在市场营销和消费者行为研究领域，而且比目前让人们自我报告情绪的方式更加准确可靠。

最后值得一提的还有一家总部位于以色列的脑电图初创公司 Hemispheric。

Hemispheric公司由苹果Face ID技术的联合创始人之一创立，正全力探索脑电图（EEG）的可扩展性规律。该公司正在世界各地建立脑电图数据采集设施，并对这些设施的搭建方式进行系统化和模块化设计，以期尽快实现规模化。

这家公司计划在未来几个月内结束隐秘运营，多年来一直致力于开发一种新型模型架构，用于训练最先进的基础脑电图模型。该公司最近成功扩展并训练了其首个数十亿参数模型。

“一些公司专注于开发改进型非侵入式传感器，押注更先进的硬件将解锁高精度非侵入式脑机接口（BCI）产品，”Hemispheric 首席执行官兼联合创始人 Hagai Lalazar 表示。“我们则持相反观点：我们认为现有的非侵入式传感方式（脑电图、脑磁图、功能性近红外光谱）已经足够，突破将并非来自更先进的传感技术，而是来自对现有信号的更精准解码。人工智能是算法史上最伟大的革命，但迄今为止，还没有人能够大规模地收集脑活动数据并训练模型来解码神经数据。我们相信，在开发用于解码大脑电活动‘语言’的人工智能方面取得突破，是实现非侵入式脑机接口普及的关键所在。”

从更宏观的角度来看，值得注意的是，对于脑电图（EEG）与尖端人工智能相结合能否实现本文所述的宏伟愿景，仍然存在诸多不确定性和质疑。许多观察人士对能否从脑电图读数中提取足够高的信号数据以支持高级脑机接口（BCI）应用持怀疑态度，甚至完全否定这种观点。这种质疑主要来自那些专注于侵入式脑机接口方法的人、那些几十年来亲身经历并运用脑电图局限性的人，以及那些并非来自深度学习领域的人士。此外，一些近期研究也对利用脑电图进行语言解码的进展提出了质疑。

怀疑论者或许是对的。

然而，现实情况是，无论是怀疑论者、这些以人工智能为先导的脑电图初创公司，还是世界上任何一位脑机接口或人工智能专家，都无法确定答案。目前世界上还没有人大规模收集脑电图训练数据，并用这些数据训练大型神经网络，评估其性能。也没有人能够最终验证或证伪脑电图基础模型是否存在像大型语言模型那样的扩展规律这一假设。

2018 年OpenAI发布第一个 GPT 模型时，没有人能够想象，也没有人会相信，在接下来的几年里，仅仅通过规模化就能带来如此惊人的性能提升。

只有时间才能证明，在脑机接口（BCI）领域，规模化能否像在机器学习（LLM）领域那样卓有成效。如果确实如此，那就不要忽视脑电图（EEG）技术。

用于神经调控的消费级可穿戴设备

从 Fitbit（被 Google 以 21 亿美元收购）到 Ōura（最近估值 110 亿美元）再到 Apple Watch（年收入超过 100 亿美元），近年来许多消费可穿戴产品都取得了突破性的成功。

所有这些消费级可穿戴产品有什么共同点？它们都能测量你的个人健康指标，但无法改变这些指标。它们只能“读取”数据，而不能“写入”数据。（上文讨论的脑电图应用案例也同样只涉及读取，而不能写入。）

新一代消费级可穿戴设备公司正在打造以大脑为中心的产品，这些产品不仅能监测大脑状态，还能主动调节大脑活动。如果这些产品真能如预期般发挥作用，不难想象，其中一款产品可能会成为下一个Ōura。

一个有趣的例子是 Somnee Sleep，这是一家初创公司，它制造了一种头带，旨在改善用户的睡眠质量。

Somnee 由四位世界顶尖的睡眠科学家共同创立，其中包括加州大学伯克利分校教授马修·沃克博士，他是颇具影响力的著作《我们为什么要睡觉》的作者。

睡眠是人类最普遍、最重要的精神活动。一款能够显著改善用户睡眠的消费产品，将蕴藏着巨大的市场机遇：据统计，每年用于安眠药的支出高达800亿美元。

Somnee是如何运作的？

Somnee的头带利用脑电图（EEG）和其他传感器追踪睡眠期间的大脑活动，并通过人工智能学习您特定的睡眠模式和信号。然后，它会发出个性化的电脉冲，引导您的脑电波进入最佳节律，从而获得更深层、更高效的睡眠。这种神经调节技术被称为经颅电刺激（tES）。

研究表明，Somnee 的消费者头带在改善睡眠方面比褪黑素有效四倍，比安眠药（如安必恩）有效 1.5 倍。

来源：Somnee Sleep

它真的有效吗？

同行评审的研究表明确实如此。

最近一项临床研究表明，Somnee 的产品在提高睡眠效率方面比褪黑素有效四倍，比安眠药（如安必恩）有效 50%。

在该公司最近完成的另一项研究中，Somnee 的头带帮助用户入睡速度提高了一倍，睡眠时间延长了 30 多分钟，翻身次数减少了三分之一。

美国国家篮球协会（NBA）近日宣布与Somnee公司合作，将该公司的产品提供给NBA球员。Equinox健身中心和酒店也将很快提供Somnee的头带。

该领域另一家值得关注的初创公司是总部位于英国的Flow Neuroscience。与Somnee类似，Flow的产品也是一款可穿戴头带，它利用经颅电刺激技术产生轻柔的个性化电脉冲，从而调节用户的大脑活动。但Somnee专注于改善睡眠，而Flow的产品则旨在对抗抑郁症。

抑郁症会影响大脑中一个关键区域，即背外侧前额叶皮层。抑郁症患者的该区域脑细胞活动减弱。Flow 的头带可将精确校准的电刺激直接输送到背外侧前额叶皮层，从而刺激该区域并恢复健康的脑细胞活动模式。

Somnee 和 Flow 都依赖于经颅电刺激 (tES)。但 Somnee 使用的是经颅交流电刺激 (tACS)，而 Flow 使用的是经颅直流电刺激 (tDCS)。它们之间有什么区别呢？简而言之，像 Flow 这样的直流电产品会向大脑提供恒定电流，使神经元更容易放电；而像 Somnee 这样的交流电产品则会引入振荡脉冲，从而影响神经元放电的节律和频率。

与Somnee一样，Flow产品的疗效也已在同行评审的研究中得到验证。去年发表在《自然医学》杂志上的一项大型临床试验发现，Flow产品在治疗抑郁症方面的疗效是抗抑郁药物的两倍。该研究显示，57%使用Flow产品的临床抑郁症患者在10周后表示抑郁症状已消失。该公司报告称，在其数万名用户中，超过75%的用户在三周内就感受到了临床症状的改善。

Flow公司将其产品描述为“以电疗为药”，这个说法非常贴切。

Somnee 和 Flow 的头带均可在网上向公众购买。

最后值得一提的初创公司是Neurode。Neurode的头带利用电刺激来提高用户的专注力和注意力。该产品既适用于患有注意力缺陷多动障碍（ADHD）的人群，也适用于希望提升整体认知功能的普通人群。

Flow采用经颅直流电刺激（tDCS，一种恒流刺激），Somnee采用经颅交流电刺激（tACS，一种节律性振荡电流），而Neurode则采用经颅随机噪声刺激（tRNS），它提供的电流频率和振幅均随机波动。新兴研究表明，引入这种随机噪声可以增强神经回路中的信号检测能力，从而改善学习和注意力。

据该公司称，45% 的用户在使用该产品的第一周内就感受到注意力有所提高。

新兴的临床研究表明，像这些公司正在研究的那种使用消费级硬件对大脑进行电刺激，确实可以对大脑行为和个人体验产生显著影响，影响领域涵盖睡眠、抑郁和注意力等诸多方面。

“这些初创公司正值良机，”美国食品药品监督管理局（FDA）数字健康部门前驻场企业家安德里亚·科拉沃斯补充道，“监管体系尚未跟上步伐。FDA的首个人工智能/机器学习框架于2019年发布，此后已有近1000种人工智能设备获得批准。正是这一监管基础，使得企业能够更快地将研究成果应用于实际人体。”

但这些产品目前都尚未获得主流市场的认可。这些公司能否打造出足够令人愉悦的产品体验和足够有效的市场推广策略，从而将这些设备推向大众市场并获得成功，时间会给出答案。

聚焦超声：下一个伟大的脑机接口范式？

如果说有一种脑机接口技术最具发展潜力——一种能够超越现有解决方案（包括本文讨论的方案）并引领神经技术新范式的方案——那就是聚焦超声。在当今脑机接口领域，没有哪个方向比它更能引起人们的关注和兴奋。

聚焦超声究竟是什么？它为何如此具有发展前景？

从根本上讲，超声波只是声音的一个子类——也就是说，它是能在空气和其他物质的粒子中传播的波。人类可以听到特定频率范围内的声波。超声波就是频率高于人类耳朵可听范围（>20千赫兹）的声波，但其传播特性与可听声波类似。

超声波技术已用于医学成像超过 75 年（任何怀孕过或有亲人怀孕的人都会记得这一点）。

脑部聚焦超声是一项较新的创新技术——直到 2010 年代才开始逐渐成形。

聚焦超声的基本原理是精确地发射多束超声波，使它们汇聚于大脑中的某一特定点。所有超声波在该焦点处叠加，产生足够的能量密度和机械压力，从而以特定的方式调节该点的神经元活动，同时不影响超声波穿过的其他脑组织。（本页上的两个简单动画图很好地展示了这一现象，使其易于理解。）

聚焦超声作为一种脑机接口技术，具有几个独特且引人注目的优势。

首先，它的精度比任何其他非侵入式脑机接口技术都要高出几个数量级。脑电图（EEG）、功能性近红外光谱（fNIRS）和经颅电刺激（tES）的空间分辨率都只有几厘米。相比之下，聚焦超声可以以亚毫米级的精度靶向大脑中的特定区域。它可以被视为一束高精度光束，可以精确地瞄准大脑中想要定位的位置。

其次，聚焦超声比任何其他非侵入性技术都能更深入地进入大脑。

由于非侵入式传感器位于颅骨外，它们通常只能探测并与大脑最外层（即新皮层）进行交互。新皮层是高级认知和语言功能的中心，因此，能够探测到新皮层的传感器可以实现许多有用的应用。但是，许多重要的脑区和功能位于大脑更深层，因此脑电图（EEG）、功能性近红外光谱（fNIRS）、经颅电刺激（tES）和其他非侵入式传感器无法触及。

丘脑、下丘脑、海马体、基底神经节和杏仁核等深层脑结构调节着我们许多基本驱动力和功能：情绪、记忆、注意力、食欲、情绪、运动、动机和渴望。精准调控这些深层脑区的能力，有望为帕金森病、强迫症、抑郁症、阿尔茨海默病、癫痫、焦虑症、慢性疼痛和创伤后应激障碍等多种神经精神疾病带来强有力的新疗法——更不用说还能为普通人群带来认知增强。

此前，只有通过手术等侵入性方法，例如脑深部刺激（DBS），才能到达这些更深层的脑区。除超声波之外的所有非侵入性疗法——无论是电波、磁波、光波还是红外波——都会被人体组织衰减，这意味着它们只能传播有限的距离就会消散。相比之下，聚焦超声波是一种机械波，因此可以几乎不受组织衰减的影响地穿过人体组织。这使得它能够在保持高度聚焦的同时，深入大脑内部。

这些可能性并非仅仅停留在理论层面。近期研究表明，聚焦超声可以显著减轻患者的慢性疼痛；降低严重成瘾者的阿片类药物渴求；并最大限度地减少特发性震颤患者的震颤症状——所有这些都涉及对大脑深部结构的激活。

超声波的最后一个优势使其区别于其他所有非侵入性检查方式：它既能读取也能写入，而且都能以高分辨率完成这两项操作。没有任何其他单一的非侵入性检查方式能够同时实现这两项功能。脑电图（EEG）、功能性近红外光谱（fNIRS）和脑磁图（MEG）可以读取数据，但不能写入数据。经颅电刺激可以写入数据（尽管分辨率和深度均低于聚焦超声），但不能读取数据。

读写能力解锁了脑机接口的圣杯：闭环功能，即一个统一的系统可以读取和解码正在进行的神经活动，然后根据读取的内容以选择性和个性化的方式刺激大脑，然后观察大脑如何实时响应和适应，等等。

与使用一个设备进行传感，另一个设备进行调制相比，一个既能读取又能写入的设备可以实现传感和刺激之间的完美对齐、低延迟、简单的校准、更少的硬件复杂性、更高的空间效率、更低的成本，并最终实现更具可扩展性的产品。

超声波脑机接口领域的创业环境尚处于起步阶段，但发展速度惊人。

目前最受瞩目的聚焦超声初创公司是 Nudge，该公司最近宣布完成由 Thrive 和 Greenoaks 领投的 1 亿美元融资。

Nudge 的首席执行官兼联合创始人 Fred Ehrsam 此前曾联合创立 Coinbase 和 Paradigm，这两家公司都是加密货币领域最成功的企业之一。Nudge 由此延续了亿万富翁创办脑机接口 (BCI) 初创公司的传统，此前 Elon Musk 创立了 Neuralink，Bryan Johnson 创立了 Kernel，Sam Altman 创立了 Merge Labs（下文将详细介绍 Merge）。Nudge 的另一位联合创始人 Jeremy Barenholtz 此前曾领导 Neuralink 的产品和技术工作。

Nudge 的使命是推进聚焦超声技术在硬件、人工智能和神经科学领域的全面发展，从而实现精准、强大的非侵入式神经调控。公司初期专注于治疗成瘾、慢性疼痛和焦虑等神经精神疾病，但其最终目标是让大众都能增强认知能力。Nudge 致力于让每个人都能精准、便捷地调节自身在学习、记忆和睡眠等领域的心理行为。

Nudge 的初始形态是一个嵌入核磁共振成像仪中的超声头盔。（核磁共振成像仪用于“读取”图像。虽然超声本身也可以用于高分辨率读取，但 Nudge 最初的核心重点是推进聚焦超声“写入”技术的最新发展。）

该公司的产品功能齐全，几乎每天都被用于人体研究。这款初始产品不便携带，不适合消费者使用，但Nudge公司已经在研发一款更小巧的架构，旨在方便在家中和日常生活中使用。

Nudge 的首款聚焦超声设备是 Nudge Zero。

来源：Nudge

正如该公司所说：“想象一下，未来无需阿片类药物即可缓解慢性疼痛，创伤后应激障碍患者可以实时调节创伤记忆，临床医生可以像检查患者心率一样轻松地对大脑回路进行成像和调节。想象一下，未来无需咖啡因即可提高注意力，学习一门新语言或一项新技能只需几天或几周，而不是几个月或几年。这并非科幻小说，而是一份工程路线图。而我们正在着手实现它。”

该领域另一家颇具潜力的初创公司是Sanmai，由亚利桑那大学教授、聚焦超声技术早期先驱Jay Sanguinetti领导。Sanmai的主要投资人是Reid Hoffman，他主导了该公司近期1200万美元的融资。

与Nudge类似，Sanmai也专注于超声波的神经调控能力（即其“写入”大脑的能力），而非其传感能力（即其“读取”大脑的能力）。与Nudge相比，Sanmai更注重严谨的临床应用，较少面向消费者。

三麦的经颅聚焦超声设备目前正在进行临床研究，有望成为世界上首个获得FDA批准的经颅聚焦超声设备。

Sanmai 的聚焦超声设备是可穿戴的，最初专注于治疗帕金森病。

来源：三枚

三麦制药的首要治疗目标是帕金森病。全球约有1000万人患有帕金森病，仅美国每年就新增9万例，这使其成为一个重要的市场机遇。三麦制药的联合创始人之一泰勒·库恩（Taylor Kuhn）发表了最早一批研究成果，证明了聚焦超声在治疗帕金森病方面的疗效。

我们尚未探讨的一个问题是，聚焦超声究竟是如何治疗帕金森病等脑部疾病的——也就是说，这项技术的作用机制是什么。简而言之，就像大多数与大脑相关的问题一样，我们尚未完全了解其中的细节。但帕金森病的案例引人入胜，值得我们深入研究。

帕金森病的主要诱因被认为是大脑深部多个区域神经元内α-突触核蛋白的错误折叠蛋白的积累。研究表明，利用聚焦超声的集中机械能靶向这些深部脑区，可以减少α-突触核蛋白的毒性积累，从而可能有助于缓解帕金森病的症状。

Sanmai计划在近期内利用聚焦超声治疗的其他疾病包括临床焦虑症，这将涉及针对患者的杏仁核进行治疗。

“我大约在15年前就开始研究超声神经调控，”Sanmai首席执行官兼联合创始人Jay Sanguinetti说道。“当时，大多数人都怀疑低强度超声的微弱机械能是否真的能影响大脑活动。作为一名研究生，我阅读了一些早期的论文——其中一些甚至有近百年的历史——并感觉其中蕴含着一些真实的东西。早期，我不得不努力争取让人们关注这些数据。如今，这个领域已经取得了巨大的进步。我们创立Sanmai的初衷是打造首款专为临床应用而设计的超声神经调控设备，它将严格的安全标准、人工智能辅助的个性化靶向定位以及实际的临床应用相结合，旨在让临床医生在诊疗过程中充满信心。”

该领域另一家前沿创业公司是 Forest Neurotech。

Forest Neurotech是一家非营利机构——更确切地说，它是一种新型的非营利创业公司，称为聚焦研究组织（FRO）。FRO是一种创新的新型融资结构，旨在支持那些规模庞大或成本高昂，对传统学术实验室而言过于庞大或昂贵，但商业化程度又不足以进入产业界的特定且雄心勃勃的科学里程碑的实现。FRO通常拥有类似创业公司的团队和文化，但其资金来源是慈善捐赠，而非风险投资。因此，Forest不追求商业化，而是专注于推进基础超声技术的尖端发展。

具体来说，福里斯特专注于将超声波硬件小型化，这是使这项技术得到广泛应用的关键一步。

而且，该公司在这方面取得了令人瞩目的成功。Forest公司最近发布了其首款设备——Forest 1脑机接口，该设备比传统的超声波扫描仪小1000倍，比标准钥匙扣还要小。

Forest Neurotech 公司的 Forest 1 设备比标准钥匙扣还要小，可以使用超声波进行读写，设计用于植入患者的颅骨内。

来源：Forest Neurotech

值得注意的是，Forest 1 设备能够利用超声波进行读写操作，这使其区别于 Nudge 和 Sanmai 设备。它能够基于血流动力学生成整个大脑（深度达 20 厘米）的高分辨率三维图像，并且还可以进行精确的神经调控。

Forest 1 设备凸显了超声波技术的一个重要特性。此前，我们一直将超声波视为一种无需手术的非侵入式脑机接口 (BCI) 技术。事实上，超声波可以而且经常以非侵入式的方式应用：Nudge 和 Sanmai 都采用了非侵入式的超声波技术。

但福雷斯特的装置属于侵入性操作：需要通过手术切开患者的头骨，并将装置植入其中。

这是为什么呢？

颅骨对于超声波来说是一个很大的挑战，因此将超声波设备放入颅骨内有很大的优势。

超声波穿过大脑等软组织时衰减很小，但颅骨则不然。颅骨由骨骼构成，对超声波的传播效果很差。颅骨会反射部分超声波，吸收部分超声波，还会散射和扭曲剩余的超声波。

如何解释超声波与颅骨相互作用以及受颅骨影响的不可预测性，是聚焦超声领域面临的最大未解工程难题之一。像Nudge和Sanmai这样的初创公司正在投入大量资源来解决这一难题。

Forest公司针对这个问题提出的解决方案是，直接将设备植入用户的颅骨内。这种方法的优点在于完全避免了超声波穿过颅骨这一棘手问题。缺点是，任何想要使用Forest设备的患者都必须先接受脑部手术。天下没有免费的午餐。

Forest公司称其植入手术为“微创手术”，因为虽然植入设备需要打开患者的颅骨，但该设备不会穿透患者的脑组织；相反，它位于大脑的保护性硬脑膜层之上。这使其与Neuralink和犹他阵列等完全侵入式脑机接口技术截然不同，后者会穿透大脑组织。

FRO（前沿研究组织）通常设定了时间限制，其理念是，如果团队实现了特定的科研目标，就可以孵化出一家传统的营利性初创公司，实现商业化。因此，不久之后，如果看到一家或多家营利性初创公司从Forest Neurotech组织中涌现出来，也不要感到惊讶。

Forest联合创始人威尔·比德曼表示：“多年来，设备小型化和计算能力的提升为我们在医疗保健领域带来了更强大的技术。现在，借助超声波技术，我们拥有了实现无创脑机接口梦想所需的保真度、精确度和理解力。”

我们将要讨论的最后一家超声波脑机接口初创公司是所有初创公司中最具雄心和前沿性的：Sam Altman 的 Merge Labs。

Merge公司尚未正式上线，因此目前公开的信息很少。（不过，未来几天内情况可能会有所改变，请不要感到惊讶！）

据报道，Sam Altman 将担任该公司联合创始人之一，OpenAI 已向该公司投入巨资，估值达 8.5 亿美元。

Merge公司将以近期超声波技术的突破为基础，对人脑进行读写操作。但它的目标是进一步拓展这项技术的边界：公司的愿景是将聚焦超声波与基因编辑相结合，从而实现更强大的脑机接口（BCI）功能。没错，你没看错：超声波加基因编辑！

这是怎么回事？

简而言之，基因编辑可以使大脑中特定的神经元群以特定方式对聚焦超声波产生反应。这一新兴科学领域被称为声遗传学。

首先，可以通过基因工程将一个特殊基因插入大脑中特定神经元亚群的DNA中。该特殊基因可以编码一种对机械力敏感的特定蛋白质。由于聚焦超声会产生微小的机械扰动，因此这些特定神经元中的这种特定蛋白质会对聚焦超声的作用产生反应。具体来说，这种蛋白质通常是一种离子通道，当受到聚焦超声作用时，它会按需打开或关闭。

与不涉及基因编辑的聚焦超声相比，声致遗传学方法能够对大脑活动进行更加精准和个性化的控制。它能够靶向大脑中的特定神经元和神经元类型，同时不影响其他神经元：例如，仅影响兴奋性神经元而不影响抑制性神经元，或者仅影响表达特定受体的神经元，或者仅影响特定的脑回路（例如，与某些成瘾行为相关的特定投射通路）。

声波成像方法还可以更直接地定义和控制聚焦超声作用于大脑神经元的机制，从而确定其作用效果，并据此将新的基因和蛋白质引入神经元。

加州理工学院著名教授米哈伊尔·夏皮罗是这一新兴研究领域的先驱之一。据报道，夏皮罗已加入 Merge Labs，这对该公司来说无疑是一项重大胜利。

即使在脑机接口这一前沿领域，Merge Labs 正在探索的方法也堪称最具前沿性和“科幻色彩”。一些基本的科学问题仍有待解决。即便最终能够成功，这一愿景的实现也至少需要十年或更长时间。

无声的语言

最后值得讨论的一种非侵入式创业类别是无声演讲。

无声语言技术能够感知并解码某人试图说或想象要说的话，即使他们没有大声说出这些话。（因此，它也被称为默读。）

无声语言技术与本文讨论的其他技术和初创公司有一个关键区别：它不涉及直接解码大脑信号。相反，它关注的是大脑下游的物理信号——特别是与说话意图相关的面部和嘴部信号。

无声语言是如何运作的？其基本原理是，当一个人试图说话时，即使没有发出任何声音，其言语系统中的各种电生理和肌肉机制（例如舌头、嘴唇、下颌）也会启动。这些生理机制是可以被检测和解码的。

目前对于实现无声说话的最佳技术方案尚未达成共识。不同的公司正在探索不同的方法，而且一般来说，无声说话初创公司对其技术细节高度保密。我们可以确定的是：从人脸解码试图说话或想象说话的物理特征的可行方法包括基于生物磁、光学和射频数据的技术。

在思考无声言语时，设想一系列可能性会很有帮助：从（1）完全正常的言语，到（2）耳语但仍然可以听到的言语，到（3）听不到但嘴巴完全张开的言语，到（4）部分张开的言语（例如，人的嘴巴保持闭合，但舌头在嘴里移动），到（5）几乎不涉及任何身体动作的“言语”，只是在脑海中构思和发出词语。

所有无声语音识别公司都在致力于开发能够解码低声语音的技术，即上文第 (3) 类和第 (4) 类语音。无声语音识别技术能否可靠地破解第 (5) 类语音——通常被称为“想象语音”——还有待观察。

近年来，语音作为一种高效、便捷且直观的交互方式，在人工智能时代迅速普及。无声语音的优势在于，它能让人们以语音作为交互界面——与他人交流、搜索互联网、记笔记、回复电子邮件等等——而且无论身处何地，无论是在办公室、拥挤的咖啡馆、地铁还是街头，都能私密且隐蔽地进行这些操作。

大多数致力于研发无声语音技术的公司都设想将这项技术嵌入到耳机或蓝牙耳机等消费产品中。在产品外形中加入某种耳塞至关重要，因为它能实现私密的低声输入与私密音频输出的闭环连接——例如，用户不仅可以隐蔽地查询人工智能模型，还可以隐蔽地接收回复。

虽然目前有不少前景看好的初创公司正在研发无声语音技术，但迄今为止只有一家公司公开亮相：那就是麻省理工学院的衍生公司 AlterEgo。AlterEgo 两个月前发布了一段 3 分钟的宣传视频，值得一看，可以帮助你更直观地了解无声语音的概念。

AlterEgo 首席执行官兼联合创始人 Arnav Kapur 表示：“目前与计算机和人工智能交互的方式受限于你在屏幕和键盘上点击和打字的速度。在智能时代，我们需要一个从零开始构建的全新界面——一个感觉像是人类思维自然延伸的界面。为了实现这一点，我们必须发明一些全新的东西。”

预计到 2026 年，会有更多资金雄厚、实力雄厚的无声演讲竞争者从幕后走向台前。

坊间盛传，苹果和谷歌等科技巨头正在认真探索将无声语音功能作为未来消费硬件产品的核心技术。同样，也有传言称，由苹果前传奇设计师乔纳森·艾维领衔设计的OpenAI即将推出的原生AI消费设备也将具备无声语音功能。

因此，我们预计在中短期内，该初创企业领域将会出现一些备受瞩目的并购交易。

但无声语言要成为人机交互领域的一个重要新范式，还需要克服一些障碍。

首先，无声语音产品的广泛应用将需要消费者行为和社会规范发生重大改变。如今，有多少人会乐意使用一种需要在办公室或咖啡馆里默默地用嘴型说话的产品呢？

无声言语面临的更根本风险在于，其他能更直接地与大脑交互的脑机接口技术可能会超越它，并使其功能黯然失色。如果能够直接从大脑提取高保真度的语言信号——例如，如果人工智能驱动的脑电图或下一代超声成像技术能够充分发挥其潜力（如上所述）——那么我们为什么还要费心研究默念呢？无声言语或许比可听见的言语更私密，延迟也比打字更低，但思想的私密性和延迟都远胜于上述所有方式。

事实上，这些技术都尚未成熟，无法真正投入市场。它们各自发展迅速，潜力巨大，但都可能面临性能瓶颈或难以实现产品化。这些技术究竟能以多快的速度发展，最终融入人们使用和喜爱的产品中，时间会给出答案。

结论

纵观人类文明，技术进步的一个显著特征就是通信和信息传输的速度、带宽和准确性的提升。文字的发明、古腾堡的印刷机、电报、无线电、电话、互联网——所有这些技术飞跃的本质都是为了增强人类共享信息的能力。

总的来说，当更多的人能够更有效地相互交流更多信息时，就会带来各种各样事先无法预测的积极影响：科学进步、健康进步、生产力提高、教育进步、以及我们对彼此和宇宙的理解加深。

脑机接口代表了数千年来技术进步的必然下一步。

信息在人与机器之间直接往返于大脑是最有效的传递方式。它消除了对有损信息的中间环节的需求，包括语言本身。毕竟，语言本身就是一种高度有损的压缩：想想看，你内心深处的心理体验，其所有细节，与你能用语言表达的程度之间，存在着多么巨大的差异。

高性能脑机接口（BCI）将开启各种奇妙而宝贵的可能性。近期影响将体现在医疗领域，这将为全球数百万患有各种神经精神疾病或心理健康问题的患者带来深远的益处。但这仅仅是个开始。试想一下，只需将新技能——比如空手道、潜水或高尔夫——“上传”到大脑，直接强化相应的神经通路，就能瞬间掌握这些技能。试想一下，能够以完美的“感官保真度”回忆和重温任何记忆。试想一下，能够重新编程大脑，使其看到或感受到如今人类大脑无法直接感知的事物：Wi-Fi信号、无线电波，甚至是“正北”方向。

更重要的是，我们甚至还无法想象脑机接口将带来的最深刻的变革和机遇——就像十四世纪的人们无法想象印刷书籍将以各种方式改变社会（民主、科学方法、启蒙运动）；或者 20 世纪 80 年代的人们无法想象互联网将以各种方式改变社会（比特币、云计算、优步）一样。

从长远来看，脑机接口技术在社会上的普及是不可避免的。然而，目前远未确定的是，脑机接口技术的主流方法究竟是无创的、有创的，还是两者兼而有之。

如今，鲜有其他技术领域像脑机接口（BCI）这样，让众多见多识广的观察者对该领域的未来发展方向持有如此截然相反的观点。一些专家基于简单的物理定律，提出了令人信服的论证，认为最先进的脑机接口技术始终需要与大脑进行直接的物理连接，因此必然需要手术。另一些专家则同样令人信服地指出，鉴于非侵入式技术在可扩展性、安全性和易用性方面的巨大优势，它们才是该领域发展的必然趋势；而且，传感、解码和调制技术的进步只是时间问题，最终即使是最先进的脑机接口应用也能以非侵入式的方式实现。还有一些专家则认为，直接作用于大脑本身并非必要，像无声语音这样的具有划时代价值的产品将基于大脑下游的信号构建。

未来几年，这些技术将从实验室走向我们生活的方方面面。做好准备吧。

The Next Frontier For AI Is The Human Brain

ByRob Toews,Contributor. I write about the big picture of artificial intelligence.

Follow Author

Dec 07, 2025, 05:45pm ESTDec 07, 2025, 11:45pm EST

Sam Altman and Elon Musk's rivalry has grown beyond AI to brain-computer interfaces.

SOURCE: GETTY

It is not possible to understand the long-term future of artificial intelligence without understanding brain-computer interfaces.

Why is that? Because brain-computer interfaces (BCI) will play a central role in defining how human intelligence and artificial intelligence fit together in a world with powerful AI.

To most people, brain-computer interfaces sounds like science fiction. But this technology is getting real, quickly. BCI is nearing an inflection point in terms of real-world functionality and adoption. Far-fetched though it may sound, capabilities like telepathy will soon be possible.

The world of BCI can be divided into two main categories: invasive approaches and non-invasive approaches. Invasive approaches to BCI require surgery. They entail putting electronics inside the skull, directly in or on the brain. Non-invasive approaches, on the other hand, rely on sensors that sit outside the skull (say, on headphones or a hat) to interpret and modulate brain activity.

In the first part of this article series, published in October, we dove deep into invasive BCI technologies and startups. In this article, we turn our attention to non-invasive BCI.

Together, BCI and AI will reshape humanity and civilization in the years ahead. Now is the time to start paying serious attention to this technology.

MORE FOR YOU

Disappointing News About The ‘Pluribus’ Season 2 Release Date On Apple TV

Bitcoin Braced For $30 Trillion Fed Bombshell After Trump Confirms ‘Immediate’ Price Game-Changer

Google Confirms Android Attacks—No Fix For Most Samsung Users

A Cornucopia of Sensors

Before we walk through today’s non-invasive BCI startup landscape, let’s spend a moment exploring the core technologies that make non-invasive BCI possible.

Whenever you use your brain to do anything—think a thought, read a book, speak a sentence, move your arm—detectable physical events take place inside your brain in certain patterns. Specifically, information flows through your brain’s neurons via tiny pulses of electricity: the same basic physical force that powers lightbulbs and kitchen appliances and iPhones. These tiny electrical signals trigger other physical activities in your brain as well, including changes in magnetic fields and blood flow.

These physical changes ultimately represent information. Their patterns encode thoughts, concepts, words, actions. And information that is encoded can be decoded. That is what brain-computer interfaces seek to do.

A number of different non-invasive sensors have been developed in order to both interpret (“read”) and modulate (“write”) the brain’s physical activities in different ways. Each has strengths and weaknesses. In order to understand the field of non-invasive BCI, it is essential to understand these different sensor types (also referred to as “modalities”) and the mechanisms by which they operate.

Read More

The world’s oldest brain sensor is the electroencephalogram, or EEG. Invented in 1924 in Germany, EEG today remains the most widely used brain sensor in the world.

EEG directly measures electrical activity from the brain using small electrodes placed on the scalp. (Electrodes are simple devices that can detect electrical activity.) EEG is highly precise from a timing perspective: it can measure neuronal activity with millisecond-level accuracy. It is also inexpensive, portable, safe and easy to use.

EEG’s great weakness is how imprecise it is from a spatial perspective. The brain’s electrical signals get heavily distorted as they pass through the skull and scalp on the way to the EEG’s electrodes, making it difficult to pinpoint exactly where in the brain they originated. This is because the skull, like most bone, is a terrible conductor of electricity.

Relatedly, EEG measurements have poor signal-to-noise ratio because the brain’s tiny electrical pulses can easily be drowned out by many other nearby sources of electrical activity: a jaw clenching, a heart beating, or just ambient electromagnetic interference. Simply blinking your eyes can generate electrical activity that is 10 to 100 times stronger than the electrical signals from your brain.

Extracting sufficiently high-fidelity signal from EEG’s noisy data thus represents a long-standing obstacle to using EEG for BCI technology.

Another non-invasive BCI modality is vastly superior to EEG on these dimensions: magnetoencephalography (MEG).

As you may remember from high school physics, electricity and magnetism are two unified aspects of the same underlying natural phenomenon: electromagnetism. So when a neuron fires and generates a tiny electrical signal, it generates a tiny magnetic field at the same time. EEG measures the electrical signal; MEG measures the associated magnetic field.

Compared to electrical fields, the remarkable thing about magnetic fields is that they pass through the skull and scalp almost completely undistorted. As a result, MEG has far greater spatial resolution and localization accuracy than EEG.

What’s the catch?

Today’s MEG systems are room-sized, requiring a magnetically shielded chamber and cryogenic cooling. They cost millions of dollars. This makes them hopelessly impractical for everyday BCI applications.

But promising research is underway to make MEG systems smaller and cheaper. A newer type of MEG based on optically pumped magnetometers (OPM-MEG) shows great promise: it works at room temperature, is small enough to wear on the head and requires less intensive shielding.

OPM-MEG technology is not yet ready for primetime. But it could become an important new BCI modality in the years ahead, offering higher-fidelity brain data than EEG while still avoiding invasive surgery.

A third non-invasive BCI modality worth mentioning is functional near-infrared spectroscopy, or fNIRS.

Instead of measuring electrical activity like EEG does, or magnetic activity like MEG does, fNIRS measures blood flow. Blood flow increases to neurons when they fire because neurons that are firing require more nutrients. By beaming high-wavelength light through the skull and into the brain, fNIRS sensors can detect changes in blood flow and use those patterns to decode brain activity.

fNIRS is today the second most common non-invasive BCI sensor in the world, behind only EEG. This is thanks in large part to the efforts of Bryan Johnson’s startup Kernel over the past decade. Kernel’s key achievement was to miniaturize fNIRS technology, turning it for the first time into a wearable device that could be commercialized at scale. Like EEG, fNIRS is safe, portable and comparatively cheap. fNIRS is more accurate than EEG in terms of location but less accurate than EEG in terms of timing; the two modalities are thus complementary and often used in tandem.

This brings us to today’s buzziest and most promising non-invasive BCI modality of all: focused ultrasound. We will have much more to say about ultrasound in this article. Read on!

The best way to understand the state of the art in non-invasive BCI—what is possible, what is not possible, where the biggest future opportunities lie—is to explore what today’s leading startups are doing. Let’s dive in.

Reading Minds with EEG

A cohort of stealthy startups believes that humble EEG is poised to transform from a familiar but limited sensor into the dominant approach to BCI.

EEG has many advantages. For decades, though, conventional wisdom has held that EEG’s signal quality is simply too poor to support advanced BCI capabilities.

How convenient, then, that one of modern AI’s great strengths is its superhuman ability to extract latent signal from noisy data.

If you are a hardcore deep learning disciple—a “Bitter Lesson” maximalist—there are good reasons for EEG to be your BCI modality of choice. In one word: scale.

The current era of AI has been defined by the principle of scaling. OpenAI popularized the concept of “scaling laws” in 2020: the idea that AI systems predictably improve as training data, model size and compute resources increase. AI’s dramatic advances in the half-decade since then have resulted, more than anything else, from scaling everything up. The reason that large language models are so astonishingly capable is that we figured out how to train them on more or less all the written text that humanity has ever produced.

If one wanted to take the playbook that has worked so well for generative AI and apply it to understanding the human brain, the key would be to collect as much brain training data as possible. And if one wanted to collect as much brain training data as possible, the best sensor to choose would be obvious: EEG. EEG is, put simply, far more scalable than any other BCI modality.

There are several orders of magnitude more EEG systems in the world today than every other kind of BCI sensor combined. EEG devices can be found in most hospitals in the world; by contrast, there are perhaps a few thousand fNIRS systems and a few hundred MEG systems globally. Basic EEG systems are available for under $1,000.

One young startup that exemplifies this AI-first, scaling-first approach to non-invasive BCI is Conduit. Cofounded by one young Oxford researcher and one young Cambridge researcher, Conduit is collecting as much data as possible as quickly as possible in order to train a large foundation model for the brain. The company says it will have collected over 10,000 total hours of brain recordings from several thousand participants by the end of the year.

While Conduit is focused primarily on collecting EEG data, it supplements this with other non-invasive modalities because the company has found that its AI’s performance improves dramatically when trained on multiple sensor modalities from each user rather than just one.

What use case is Conduit envisioning for its technology?

The company’s goal is—astonishingly—to build a BCI product that can decode users’ thoughts before they have even formulated those thoughts into words. In other words, they are seeking to build thought-to-text AI.

And according to the company, the system is already beginning to work. Conduit’s current AI model produces text outputs that achieve ~45% semantic matches with users’ thoughts, and can do so zero-shot (meaning that the AI system is not fine-tuned on any particular individual ahead of time).

A few specific examples will help make this more concrete.

In one example, when a human participant thought the phrase “the room seemed colder,” the AI generated the phrase “there was a breeze even a gentle gust.” In another example, the participant thought “do you have a favorite app or website” and the AI generated “do you have any favorite robot.”

This technology is not yet ready for primetime. 45% accuracy is not good enough for a mass-market product. And, for now, these results are only possible when users put an unwieldy suite of sensors on their heads. But this level of accuracy is nonetheless remarkable when considering that the task at hand is reading people’s minds. And the company is just getting started. Conduit only began scaling its data collection efforts a few months ago; the company plans to increase its training data corpus by several orders of magnitude moving forward.

Imagine what might become possible—imagine how society might change—if it were possible to communicate nuanced ideas to other people and to computers merely by thinking them.

"The biggest lesson from ML in the last decade has been the importance of scale and data,” said Conduit cofounder Rio Popper. “Noninvasive approaches let us collect a much larger and more diverse dataset than we’d be able to if everyone in our dataset had to get brain surgery first.”

Added her cofounder Clem von Stengel: “We founded Conduit because we realized that people could get things done so much faster if we all thought directly in ideas rather than in words. And we could have a much richer understanding of each other and of the world in general.”

Another interesting young startup pushing the limits of what is possible with EEG is Alljoined.

Alljoined, like Conduit, is taking an AI-first approach to non-invasive BCI and is betting on EEG as the right modality given its scalability and accessibility. While Conduit’s goal is to decode thoughts into language, Alljoined’s initial focus is to decode thoughts into images—that is, to faithfully reproduce an image that a user has in his or her “mind’s eye” based on EEG readings, a task known as image reconstruction.

Alljoined’s CEO/cofounder Jonathan Xu co-authored the seminal MindEye2 paper, which showed that generative AI-based methods could achieve accurate image reconstruction based on only modest amounts of fMRI data. Alljoined set out to extend that work from fMRI to EEG data—and has already had success doing so.

The graphics below show some examples of images that Alljoined’s AI system reconstructed from participants’ EEG data. As you can see, the reconstructed outputs are not fully accurate, but these results represent state-of-the-art performance today. And—as we have observed in so many other fields in AI—it is a safe bet that the system’s performance will continue to improve as training data and compute scale.

The top row represents the image that a human participant looked at, and the bottom row represents the image that Alljoined's AI system reconstructed based on the participant's EEG data.

SOURCE: ALLJOINED

Speaking of training data, last year Alljoined open-sourced the first-ever dataset specifically built for image reconstruction from EEG. The dataset contains EEG data from 8 different participants looking at 10,000 images each. Making this data freely available should serve as a helpful catalyst for the entire field.

While Alljoined’s initial focus has been on image reconstruction, the company is also exploring other application areas. One promising area is sentiment analysis—the ability to accurately and granularly identify the emotion that a user is experiencing in real-time. Decoding sentiments directly from brain data could have significant commercial relevance, for instance in marketing and consumer behavior research, and would be far more high-fidelity than the current status quo of asking individuals to self-report their emotions.

One final EEG startup worth mentioning is Israel-based Hemispheric.

Founded by one of the co-creators of Apple’s FaceID technology, Hemispheric is going all in on the pursuit of scaling laws for EEG. The company is establishing EEG data collection facilities around the world, systematizing and modularizing how these facilities are set up in order to scale as quickly as possible.

The company, which plans to come out of stealth mode in the coming months, has spent years developing a novel model architecture to train a state-of-the-art foundation EEG model. The company recently successfully scaled and trained its first multi-billion-parameter model.

“Some companies are focused on developing improved non-invasive sensors, betting that better hardware will unlock high-precision non-invasive BCI products,” said Hemispheric CEO/cofounder Hagai Lalazar. “We are making the opposite bet: that current non-invasive sensing modalities (EEG, MEG, fNIRS) suffice, and that the breakthrough will come not from better sensing but from better decoding of existing signals. AI is the biggest revolution in the history of algorithms, but so far no one has scaled brain activity data collection and model training for decoding neural data. We believe that a breakthrough in developing AI for decoding the ‘language’ of the brain’s electrical activity is the missing link to making non-invasive BCIs pervasive.”

Zooming out, it is important to note that plenty of uncertainty and skepticism still exist as to whether EEG paired with cutting-edge AI will be able to deliver on the lofty visions outlined here. Many observers are doubtful or downright dismissive of the idea that sufficiently high-signal data can ever be extracted from EEG readings to enable advanced BCI use cases. Much skepticism comes in particular from those who focus on invasive approaches to BCI, those who have witnessed and worked with EEG’s limitations first-hand for decades, and/or those who do not come from the world of deep learning. And some recent research has cast doubt on progress in language decoding from EEG.

The skeptics may prove to be right.

The reality, though, is that no one—not the skeptics, not these AI-first EEG startups, not any BCI or AI expert in the world—knows for sure. No one in the world has yet collected EEG training data at massive scale and trained a large neural network on it and assessed its performance. No one has yet definitively validated or falsified the hypothesis that scaling laws exist for EEG foundation models like they do for large language models.

When OpenAI published the first GPT model in 2018, no one could have conceived of, and no one would have believed, the breathtaking performance gains that would result over the next few years from sheer scaling.

Only time will tell whether scaling will prove anywhere near as productive in the world of BCI as it has for LLMs. If it does, don’t sleep on EEG.

Consumer Wearables for Neuromodulation

From FitBit (acquired by Google for $2.1 billion) to Ōura (recently valued at $11 billion) to Apple Watch (generating well over $10 billion in annual revenue), a number of consumer wearable products have achieved breakout success in recent years.

What do all these consumer wearable products have in common? They measure your personal health metrics, but they cannot change them. They can only “read”; they cannot “write”. (The EEG use cases discussed above all likewise involve only reading, not writing.)

A new generation of consumer wearable companies is building brain-focused products that don’t just monitor your brain state but actively modulate it. If these products work as expected, it’s not hard to imagine that one of them could become the next Ōura.

One intriguing example is Somnee Sleep, a startup that has built a headband to improve the quality of its users’ sleep.

Somnee was co-founded by four of the world’s leading sleep scientists, including UC Berkeley professor Dr. Matthew Walker, author of the influential book Why We Sleep.

No mental activity is more universal or more important than sleep. A consumer product that could significantly improve users’ sleep could unlock a massive market opportunity: as a point of reference, $80 billion is spent on sleeping pills annually.

How does Somnee work?

Somnee’s headband uses EEG and other sensors to track your brain’s activity during sleep, learning its particular sleep patterns and signals using AI. It then sends out personalized electrical pulses that nudge your brainwaves into their optimal rhythms for deeper, more efficient sleep. This neuromodulation technology is known as transcranial electrical stimulation, or tES.

Research shows that Somnee's consumer headband is four times more effective than melatonin and 1.5 times more effective than sleeping pills like Ambien at improving sleep.

SOURCE: SOMNEE SLEEP

Does it actually work?

Peer-reviewed research suggests that it does.

One recent clinical study showed that Somnee’s product is four times more effective than melatonin and 50% more effective than sleeping pills like Ambien at improving sleep efficiency.

In another study that the company recently completed, Somnee’s headband helped users fall asleep twice as fast, stay asleep more than 30 minutes longer and reduce tossing and turning by one-third.

The National Basketball Association recently announced that it is partnering with Somnee to make the company’s product available to NBA players. Equinox will also soon make Somnee’s headbands available in its gyms and hotels.

Another noteworthy startup in this category is UK-based Flow Neuroscience. Similar to Somnee, Flow’s product is a wearable headband that uses transcranial electrical stimulation to generate gentle personalized electrical pulses that modulate its user’s brain activity. But while Somnee focuses on improving sleep, Flow’s product is designed to combat depression.

Depression affects a key region of the brain called the dorsolateral prefrontal cortex. In depressed individuals, the brain cells in this region become less active. Flow’s headband delivers precisely calibrated electrical stimulation directly to the dorsolateral prefrontal cortex in order to stimulate this region and restore healthy brain cell activity patterns.

Both Somnee and Flow rely on transcranial electrical stimulation (tES). But while Somnee uses transcranial alternating current stimulation (tACS), Flow makes use of transcranial direct current stimulation (tDCS). What’s the difference? In short, direct current products like Flow provide a constant current to the brain that make neurons generally more likely to fire, while alternating current products like Somnee introduce an oscillating pulse that influences the rhythms and frequencies at which neurons fire.

Like Somnee, the efficacy of Flow’s product has been validated in peer-reviewed studies. A large clinical trial published last year in Nature Medicine found that the Flow product is twice as effective at addressing depression as antidepressant drugs. According to the study, 57% of clinically depressed patients who used the Flow product reported that they no longer had depression after 10 weeks. The company reports that, of its total user base of tens of thousands of customers, over 75% see some clinical improvement within three weeks.

Flow describes its product as “electricity as medicine.” It is an apt phrase.

Both Somnee and Flow’s headbands are available online to the general public.

One final startup worth mentioning is Neurode. Neurode’s headband uses electrical stimulation to improve its users’ focus and attention. The product is intended both for individuals with ADHD and for members of the broader population looking to boost their overall cognitive functioning.

While Flow uses tDCS (a constant current) and Somnee uses tACS (a rhythmically oscillating current), Neurode uses transcranial random noise stimulation, or tRNS, which delivers current that fluctuates randomly in both its frequency and amplitude. Emerging research suggests that introducing this random noise can boost signal detection in neural circuits, thus improving learning and focus.

According to the company, 45% of its users experience an increase in focus within the first week of using the product.

An emerging body of clinical research indicates that electrical stimulation of the brain with consumer-grade hardware, like these companies are pursuing, can indeed meaningfully influence brain behavior and individual experience in areas as diverse as sleep, depression and focus.

“These startups are building at the right moment,” added Andrea Coravos, a former Entrepreneur in Residence in the FDA’s Digital Health Unit. “The regulatory infrastructure isn’t playing catch-up. The FDA’s first AI/ML framework came out in 2019, and nearly 1,000 AI-enabled devices have been authorized since. That regulatory foundation is what lets companies move from research to real humans, faster.”

But none of these products have yet won mainstream adoption. Time will tell whether these companies are able to craft product experiences that are delightful enough and go-to-market strategies that are effective enough to turn these devices into mass-market successes.

Focused Ultrasound: The Next Great BCI Paradigm?

If there is one BCI technology that offers the greatest upside potential—one approach that could transcend the existing landscape of solutions (including those discussed in this article) and usher in a new paradigm for neurotechnology—it is focused ultrasound. No area within the world of brain-computer interfaces is generating more buzz and excitement today.

What exactly is focused ultrasound, and why is it so promising?

At a basic level, ultrasound is just a subcategory of sound—that is, waves that travel through particles in air and other materials. Humans can hear sound waves that fall within a certain range of frequencies. Ultrasound waves are simply sound waves with a higher frequency than humans can detect with their ears (>20 kilohertz), but that otherwise behave similarly to audible sound waves.

Ultrasound technology has been used for medical imaging for over 75 years (as anyone who has ever been pregnant or had a pregnant loved one will recall).

Focused ultrasound for the brain is a much newer innovation—one that began to take shape only in the 2010s.

The basic concept of focused ultrasound is to aim and launch many ultrasound waves in a precise sequence such that they all converge at one particular point in the brain. All the individual waves add together at that one focal point, creating enough energy density and mechanical pressure to modulate the neurons in particular ways in that one spot while leaving unaffected the rest of the brain tissue that the waves travel through. (The two simple animated graphics on this page do an excellent job of visualizing this phenomenon, making it intuitive to grasp.)

Focused ultrasound offers several unique and compelling advantages as a BCI modality.

The first is that it is orders of magnitude more precise than any other non-invasive BCI modality. EEG, fNIRS and tES all offer spatial resolution of a few centimeters. Focused ultrasound, by contrast, can target a particular spot in the brain with sub-millimeter precision. It can be thought of as a high precision beam that can be aimed at the exact location in the brain that one wants to target.

Second, focused ultrasound can reach deeper into the brain than any other non-invasive technology.

Because non-invasive sensors sit outside the skull, they generally are only able to access and interact with the outermost layer of the brain, known as the neocortex. The neocortex is the seat of advanced cognition and language, so plenty of useful applications are achievable with sensors that can only reach the neocortex. But many important regions and functions sit deeper inside the brain and are therefore out of reach for EEG, fNIRS, tES and other non-invasive sensors.

Deep brain structures like the thalamus, hypothalamus, hippocampus, basal ganglia and amygdala regulate many of our fundamental drives and functions: emotions, memory, attention, appetite, mood, movement, motivation, cravings. The ability to precisely modulate these deep brain regions could enable powerful new treatments for neuropsychiatric disorders as diverse as Parkinson’s, OCD, depression, Alzheimer’s, epilepsy, anxiety, chronic pain and PTSD—not to mention unlocking cognitive augmentation for the general population.

Up until now, access to these deeper regions could only be achieved via invasive methods that require surgery, like deep brain stimulation (DBS). All non-invasive modalities other than ultrasound—whether electrical, magnetic, optical or infrared—are attenuated by human tissue, which means they can make it only a limited distance before they dissipate. Focused ultrasound, by contrast, is a mechanical wave and as a result can pass through human tissue with very little attenuation. This enables it to travel deep into the brain while maintaining its concentrated focus.

And these possibilities are not just theoretical. Recent research has shown that focused ultrasound can, for instance, meaningfully reduce chronic pain in patients; decrease opioid cravings in participants with serious addictions; and minimize tremors for those who suffer from essential tremor—all of which involve accessing deep brain structures.

One final advantage of ultrasound that sets it apart from every other non-invasive modality: it can both read and write, and it can do both with high resolution. No other individual non-invasive modality can carry out both of these functions. EEG, fNIRS and MEG can read, but they cannot write. Transcranial electrical stimulation can write (though at lower resolution and shallower depth than focused ultrasound), but it cannot read.

The ability to both read and write unlocks the holy grail for brain-computer interfaces: closed loop functionality, whereby one unified system can read and decode ongoing neural activity, then stimulate the brain in selective and personalized ways based on what it reads, then see how the brain responds and adapt in realtime, and so on.

Compared to using one device to sense and a different one to modulate, a single device that can both read and write enables perfect alignment between sensing and stimulation, low latency, straightforward calibration, less hardware complexity, greater space efficiency, lower cost, and ultimately more scalable products.

The startup landscape for ultrasound BCI is nascent but developing at breakneck speed.

Today’s most high-profile focused ultrasound startup is Nudge, which recently announced a $100 million fundraise led by Thrive and Greenoaks.

Nudge CEO/cofounder Fred Ehrsam previously cofounded both Coinbase and Paradigm, two of the most successful organizations in the world of crypto. Nudge thus continues the lineage of billionaires launching BCI startups, following Elon Musk with Neuralink, Bryan Johnson with Kernel, and Sam Altman with Merge Labs (more on Merge below). Nudge’s other cofounder Jeremy Barenholtz previously led product and technology at Neuralink.

Nudge’s mission is to advance the state of the art in focused ultrasound across the full stack of hardware, AI and neuroscience in order to enable precise and powerful non-invasive neuromodulation. The company’s initia...