Jeff Bezos Is Funding a Wild Hunt for the Brain’s ‘Core Algorithm’
Jeff Bezos Is Funding a Wild Hunt for the Brain’s ‘Core Algorithm’
杰夫·贝索斯正在资助一场寻找大脑“核心算法”的疯狂探索
Rob Williams knows how to pitch Jeff Bezos: You write a press release as if your product has already been built. Bezos reads it and gives a thumbs up or down. 罗伯·威廉姆斯（Rob Williams）深知如何向杰夫·贝索斯（Jeff Bezos）推销：你得写一份新闻稿，就好像你的产品已经研发出来了一样。贝索斯读完后，会给出赞成或反对的评价。
Williams went through this process a lot as an executive on Amazon’s “S-team,” in charge of software products such as Alexa, until his departure last fall. But the pitch he made a few weeks later—in December 2025—was different. Now he was collaborating with Thomas Reardon, a neuroscientist and repeat startup founder, and approaching Bezos as a funder, not a boss. 作为亚马逊“S-team”的高管，威廉姆斯在去年秋天离职前，曾多次经历这一流程，负责Alexa等软件产品。但他几周后（2025年12月）所做的推销却有所不同。这一次，他与神经科学家兼连续创业者托马斯·里尔登（Thomas Reardon）合作，是以融资者的身份，而非下属的身份接触贝索斯。
Here’s what Bezos, sitting on his yacht somewhere, read while Williams anxiously watched on Zoom: Flourish is a neuro AI company that is solving the two most difficult problems facing AI today: power efficiency and continuous learning. We are building Cortex AI, the first synthetic intelligence system designed to match the computational capacity, learning efficiency, and power budget of the human brain. 当时，贝索斯正坐在某处的游艇上，而威廉姆斯在Zoom上焦急地等待着。贝索斯读到了以下内容： Flourish是一家神经人工智能公司，致力于解决当今AI面临的两大最棘手问题：能效与持续学习。我们正在构建Cortex AI，这是首个旨在匹配人类大脑计算能力、学习效率和功耗预算的合成智能系统。
A month later, I’m lunching with Reardon and Williams in the Flatiron neighborhood in New York City. Reardon gets right to the point. AI has dug itself into a hole, he says. Though increasingly powerful, large language models are greedy consumers of computer power and data. 一个月后，我在纽约市熨斗区（Flatiron）与里尔登和威廉姆斯共进午餐。里尔登直奔主题。他说，人工智能已经陷入了困境。尽管大型语言模型（LLM）的功能日益强大，但它们是计算机算力和数据的贪婪消耗者。
Though the inspiration for LLMs was rooted in biology, current frontier models have little in common with the human brain. A person uses about 20 watts of energy to process information; a single chip in an AI training cluster uses more than 30 times that amount. The hyperscalers require thousands of chips and gigawatts of energy, enough to power small cities. And those models need to suck up virtually all of what humans have written. Each new model requires more, more, more. For all of that, the models don’t learn. Once you train them, they’re stuck. 虽然大语言模型的灵感源于生物学，但当前的前沿模型与人脑几乎没有共同之处。一个人处理信息大约只需要20瓦的能量；而AI训练集群中的单个芯片消耗的能量是这个数字的30多倍。超大规模计算中心需要数千个芯片和吉瓦级的能量，足以供应小型城市。而且，这些模型需要吸收人类写下的几乎所有内容。每一个新模型都需要更多、更多、更多。尽管如此，这些模型却无法学习。一旦训练完成，它们就停滞不前了。
The goal, Reardon tells me, is to build “a synthetic artificial intelligence brain that runs on 50 watts or less.” It should adapt to its conditions, be as nimble as a human mind, and burn a tiny fraction of an LLM’s compute power and energy. The proof of concept is thriving inside our skulls. “There’s something fundamentally wrong with saying, ‘I need to basically read every book ever written 20 times over in order to learn English,’” Reardon says. “A human baby does it with a couple hundred thousand utterances.” 里尔登告诉我，目标是构建“一个功耗在50瓦或更低的人工智能大脑”。它应该能够适应环境，像人类思维一样灵活，并且只消耗大语言模型算力和能量的一小部分。概念验证就在我们的颅骨内蓬勃发展。里尔登说：“如果说‘我需要把有史以来写过的每一本书读上20遍才能学会英语’，这在根本上是错误的。人类婴儿只需要几十万次的话语输入就能做到。”
Reardon and Williams haven’t figured out yet how to build systems that match the magic of a human brain. What they have is a belief that an expert, well-resourced team—of AI researchers and neuroscientists working essentially side by side—can find the answer. The neuroscientists will conduct original wet lab experiments with some of the most advanced lab equipment available, to hunt for usable intel on the brain’s architecture. They plan to release the models they’re currently developing as near-term products on the path to a full reinvention of AI. 里尔登和威廉姆斯尚未找到构建能够媲美人脑神奇之处的系统的方法。但他们坚信，一支由AI研究人员和神经科学家并肩工作的专家团队，且拥有充足的资源，一定能找到答案。神经科学家将利用最先进的实验室设备进行原创性的湿实验室实验，以寻找关于大脑架构的可用情报。他们计划将目前正在开发的模型作为短期产品发布，以此作为彻底重塑人工智能之路上的里程碑。
The fuzziness of the proposal didn’t bother Jeff Bezos. After reading Williams’ two-pager, he chipped in $50 million. Other funding came from Lux Capital, Google Ventures, and Catalio, among others. Bezos then almost doubled his initial stake and told Reardon he’d have given more if they’d asked. Now with a war chest of $500 million and a reported valuation of $2.5 billion, Flourish just needs to invent a new way to do AI. 提案的模糊性并没有困扰杰夫·贝索斯。在读完威廉姆斯的这份两页纸计划书后，他投入了5000万美元。其他资金来自Lux Capital、Google Ventures和Catalio等机构。随后，贝索斯几乎将他的初始投资翻了一番，并告诉里尔登，如果他们开口，他还会投更多。如今，Flourish拥有5亿美元的资金储备，据报道估值达25亿美元，他们现在只需要发明一种全新的AI实现方式。
Thomas Reardon IV doesn’t use his first name–too many Toms in the family tree. “My wife calls me Reardon, everyone calls me Reardon,” he says. He grew up one of 18 kids in a working-class family and dropped out of the University of New Hampshire at age 15. From there his résumé goes bonkers: He becomes a teenage programming wizard, gets hired to help build Microsoft’s first web browser, and starts and sells a wireless tech company. Next he goes to Columbia University for a degree in classics, gets into neuroscience and ultimately earns a doctorate in it (also from Columbia). He starts another company with some classmates, develops a mind-control wristband, gets acquired by Meta, and works there for six years. (The wristband comes with Meta’s latest smart glasses.) 托马斯·里尔登四世（Thomas Reardon IV）从不使用他的名字——因为家族树里叫“汤姆”的人太多了。“我妻子叫我里尔登，每个人都叫我里尔登，”他说。他在一个有18个孩子的工人阶级家庭长大，15岁时从新罕布什尔大学辍学。从那时起，他的履历变得不可思议：他成为了一名十几岁的编程天才，被聘请去帮助构建微软的第一款网络浏览器，并创办并出售了一家无线技术公司。接下来，他进入哥伦比亚大学攻读古典学学位，随后投身神经科学，并最终获得了该专业的博士学位（同样是在哥伦比亚大学）。他与同学创办了另一家公司，开发了一款意念控制腕带，被Meta收购，并在那里工作了六年。（这款腕带随Meta最新的智能眼镜一同推出。）
But Reardon was dissatisfied with how companies, including Meta, were building cutting-edge AI. Matching the brain’s ability to learn and energy parsimony isn’t a new idea. Both IBM and Intel have released neuromorphic chips inspired by the brain’s architecture. UC Berkeley computer scientist Ben Recht, who is a Flourish adviser, recalls that scientists decades ago were into neuromorphic approaches to software. Then LLMs took over. “They call those neural nets, but there’s nothing brain-like happening there,” Recht says. 但里尔登对包括Meta在内的公司构建尖端AI的方式感到不满。匹配大脑的学习能力和节能性并不是一个新想法。IBM和英特尔都曾发布过受大脑架构启发的神经形态芯片。加州大学伯克利分校的计算机科学家、Flourish顾问本·雷希特（Ben Recht）回忆说，几十年前科学家们就热衷于软件的神经形态方法。后来，大语言模型占据了主导地位。“他们称之为神经网络，但那里并没有发生任何类似大脑的事情，”雷希特说。
Reardon convinced Williams, the Amazon exec, whom he knew from their time at Microsoft, to join him. Another early recruit was Greg Wayne, a longtime researcher at DeepMind, who heads Project Astra, Google’s AI assistant initiative. “I didn’t know if they could achieve their goal, but I thought it would lead to interestingness, which probably will be useful,” Wayne says. DeepMind CEO Demis Hassabis fought to keep Wayne, and they forged an arrangement where Wayne kept his job but would spend 20 percent of his time at Flourish. 里尔登说服了他在微软时期就认识的亚马逊高管威廉姆斯加入。另一位早期招募的人才是DeepMind的资深研究员格雷格·韦恩（Greg Wayne），他负责谷歌的AI助手项目“Project Astra”。“我不知道他们是否能实现目标，但我认为这会带来有趣的结果，而且很可能会有用，”韦恩说。DeepMind首席执行官德米斯·哈萨比斯（Demis Hassabis）极力挽留韦恩，最终他们达成了一项协议：韦恩保留原职，但将20%的时间投入到Flourish。
By the end of March, Reardon had hired around two dozen top neuroscientists and AI researchers. I visited them the day the company moved into an office space in New York City’s West SoHo area, in a 10-story building with a built-in data center. People were setting up their computers; the lab equipment, like electron microscopes, were yet to arrive. 到3月底，里尔登已经聘请了大约两打顶尖的神经科学家和AI研究人员。在公司搬进纽约市西苏荷区（West SoHo）一栋带有内置数据中心的10层办公楼的那天，我拜访了他们。人们正在安装电脑；而电子显微镜等实验室设备尚未送达。
“The brain has a secret we haven’t found yet,” says Wayne. The team is focusing on structures called cortical columns, which one Flourish scientist calls “the canonical computational unit” of the brain. “大脑有一个我们尚未发现的秘密，”韦恩说。该团队正专注于一种被称为“皮层柱”（cortical columns）的结构，一位Flourish的科学家将其称为大脑的“规范计算单元”。