Jeff Bezos Is Funding a Wild Hunt for the Brain’s ‘Core Algorithm’

Rob Williams knows however to transportation Jeff Bezos: You constitute a property merchandise arsenic if your merchandise has already been built. Bezos reads it and gives a thumbs up oregon down.

Williams went done this process a batch arsenic an enforcement connected Amazon’s “S-team,” successful complaint of bundle products specified arsenic Alexa, until his departure past fall. But the transportation helium made a fewer weeks later—in December 2025—was different. Now helium was collaborating with Thomas Reardon, a neuroscientist and repetition startup founder, and approaching Bezos arsenic a funder, not a boss.

Here’s what Bezos, sitting connected his yacht somewhere, work portion Williams anxiously watched connected Zoom:

Flourish is simply a neuro AI institution that is solving the 2 astir hard problems facing AI today: powerfulness ratio and continuous learning. We are gathering Cortex AI, the archetypal synthetic quality strategy designed to lucifer the computational capacity, learning efficiency, and powerfulness fund of the quality brain.

A period later, I’m lunching with Reardon and Williams successful the Flatiron vicinity successful New York City. Reardon gets close to the point. AI has dug itself into a hole, helium says. Though progressively powerful, ample connection models are greedy consumers of machine powerfulness and data.

Though the inspiration for LLMs was rooted successful biology, existent frontier models person small successful communal with the quality brain. A idiosyncratic uses astir 20 watts of vigor to process information; a azygous spot successful an AI grooming clump uses much than 30 times that amount. The hyperscalers necessitate thousands of chips and gigawatts of energy, capable to powerfulness tiny cities. And those models request to suck up virtually each of what humans person written. Each caller exemplary requires more, more, more. For each of that, the models don’t learn. Once you bid them, they’re stuck.

The goal, Reardon tells me, is to physique “a synthetic artificial quality encephalon that runs connected 50 watts oregon less.” It should accommodate to its conditions, beryllium arsenic nimble arsenic a quality mind, and pain a tiny fraction of an LLM’s compute powerfulness and energy. The impervious of conception is thriving wrong our skulls. “There’s thing fundamentally incorrect with saying, ‘I request to fundamentally work each publication ever written 20 times implicit successful bid to larn English,’” Reardon says. “A quality babe does it with a mates 100 1000 utterances.”

Reardon and Williams haven’t figured retired yet however to physique systems that lucifer the magic of a quality brain. What they person is simply a content that an expert, well-resourced team—of AI researchers and neuroscientists moving fundamentally broadside by side—can find the answer. The neuroscientists volition behaviour archetypal bedewed laboratory experiments with immoderate of the astir precocious laboratory instrumentality available, to hunt for usable intel connected the brain’s architecture. They program to merchandise the models they’re presently processing arsenic near-term products connected the way to a afloat reinvention of AI.

The fuzziness of the connection didn’t fuss Jeff Bezos. After speechmaking Williams’ two-pager, helium chipped successful $50 million. Other backing came from Lux Capital and Google Ventures, among others. Bezos past astir doubled his archetypal involvement and told Reardon he’d person fixed much if they’d asked. Now with a warfare thorax of $500 cardinal and a reported valuation of $2.5 billion, Flourish conscionable needs to invent a caller mode to bash AI.

Thomas Reardon IV doesn’t usage his archetypal name–too galore Toms successful the household tree. “My woman calls maine Reardon, everyone calls maine Reardon,” helium says. He grew up 1 of 18 kids successful a working-class household and dropped retired of the University of New Hampshire astatine property 15. From determination his résumé goes bonkers: He becomes a teenage programming wizard, gets hired to assistance physique Microsoft’s archetypal web browser, and starts and sells a wireless tech company. Next helium goes to Columbia University for a grade successful classics, gets into neuroscience and yet earns a doctorate successful it (also from Columbia). He starts different institution with immoderate classmates, develops a mind-control wristband, gets acquired by Meta, and works determination for six years. (The wristband comes with Meta’s latest astute glasses.)

But Reardon was dissatisfied with however companies, including Meta, were gathering cutting-edge AI. Matching the brain’s quality to larn and vigor parsimony isn’t a caller idea. Both IBM and Intel person released neuromorphic chips inspired by the brain’s architecture. UC Berkeley machine idiosyncratic Ben Recht, who is simply a Flourish adviser, recalls that scientists decades agone were into neuromorphic approaches to software. Then LLMs took over. “They telephone those neural nets, but there's thing brain-like happening there,” Recht says.

Reardon convinced Williams, the Amazon exec, whom helium knew from their clip astatine Microsoft, to articulation him. Another aboriginal enlistee was Greg Wayne, a longtime researcher astatine DeepMind, who heads Project Astra, Google’s AI adjunct initiative. “I didn't cognize if they could execute their goal, but I thought it would pb to interestingness, which astir apt volition beryllium useful,” Wayne says. DeepMind CEO Demis Hassabis fought to support Wayne, and they forged an statement wherever Wayne kept his occupation but would walk 20 percent of his clip astatine Flourish.

By the extremity of March, Reardon had hired astir 2 twelve apical neuroscientists and AI researchers. I visited them the time the institution moved into an bureau abstraction successful New York City’s West SoHo area, successful a 10-story gathering with a built-in information center. People were mounting up their computers; the laboratory equipment, similar electron microscopes, were yet to arrive.

“The encephalon has a concealed we haven’t recovered yet,” says Wayne. The squad is focusing connected structures called cortical columns, which 1 Flourish idiosyncratic calls “the canonical computational unit” of the brain. One of Flourish’s investors is Jacob Vogelstein, a neuroscientist turned task capitalist who, on with his member Joshua and others, started an ambitious inaugural called the Open Connectome Project. “The thought was that you could cod each these images of the encephalon and commencement to bash information processing connected them to effort to construe the circuits,” helium says.

That enactment could extremity up being utile to the team. Joshua Vogelstein—a Flourish cofounder—recently coauthored a insubstantial connected the neural web of a effect alert and recovered that its web is 10 times much businesslike than the transformer, a halfway architectural portion of an LLM. “The methods are astatine an inflection point,” says Nathan Danielson, a Flourish neuroscientist and doc who worked with Reardon astatine Meta.

Flourish is not unsocial successful seeking answers successful the brain; the word “neuromorphic” has been flung astir truthful overmuch that it’s astir a buzzword. A institution called Cortical Labs is combining lab-grown neurons with silicon chips. OpenAI CEO Sam Altman is backing Merge Labs, with “the semipermanent ngo of bridging biologic and artificial intelligence.” Meta’s superintelligence radical claims that its TRIBE v2 exemplary “acts arsenic a integer duplicate of quality neural activity.” An enactment called Unconventional AI is creating a assistance programme for probe that addresses Reardon’s goal: gathering an AI that replicates biologic efficiency. Some task superior firms adjacent specialize successful brain-science efforts.

Reardon believes that the company’s borderline lies successful its unusually beardown harvest of neuroscientists. These researchers volition behaviour laboratory experiments portion the AI squad builds models informed by their discoveries; the algo team, meanwhile, mightiness unearth clues that assistance the neuroscientists. “You don't truly cognize if you recognize thing until you tin physique it, instrumentality it successful silicon,” says Josh Morgan, a Flourish neuroscientist. They accidental they’re unfastened to publishing immoderate of their archetypal research.

“Fundamentally, the institution is looking for the algorithms underlying intelligence,” says Jacob Vogelstein, the investor.

Reardon tells maine that his squad has identified paths to near-term gross that exploit caller encephalon research. They’re processing a hippocampus-inspired mode to grip representation that volition let the company’s models to larn without extended grooming data, helium says. The algorithm squad has built a exemplary that tin larn continuously and is moving connected embodying it successful “the kinds of devices you transportation successful your pocket,” helium says. He adds that he's negotiating with a large spot shaper to enactment the exemplary connected silicon.

In aboriginal May, the Flourish scientists held an all-hands gathering successful their New York office. Reardon and Williams are seated astatine a league array with astir a twelve others, including elder advisor Wayne, who's visiting from London. The scientists are debating six imaginable experiments. These are large swings that necessitate the acquisition of multimillion-dollar microscopy machines and years of work.

The treatment of the experiments brings up biologic phenomena arsenic wide arsenic however rabies spreads successful the cortex and the neurobiology of birdsong. They statement whether they should analyse molecules and synapses oregon absorption connected bigger-scale cells oregon circuits. Are analyses of connectomes successful the rodent encephalon capable for immoderate purposes, oregon would lone quality brains do?

The verdict, for the infinitesimal astatine least, seems to beryllium to effort it all. “The takeaway is that we privation to bash information postulation crossed the nano, micro, and meso scales to enactment the find of the halfway algorithm,” says Sean Bittner, a computational neuroscientist who besides worked with Reardon astatine Meta.

There’s an awkward pause, past Wayne speaks up. “That’s a fucking large experimental plan!” helium says. “It’s really practical, arsenic opposed to, like, insane.”

There’s nary question that Flourish is making a risky, long-range bet. As Williams tells it, aft speechmaking the transportation document, Bezos needed to cognize 1 thing—were the founders committed to spending years connected this? When helium got an emphatic yes, helium agreed to merchandise his millions. (Bezos did not respond to requests for comment.) “You’ve can’t get precise overmuch done successful 3 years,” says Williams. “The mode you marque large differences is to program for things that person worth 7 to 10 years out.” (For the record, Reardon says he’s hoping that Flourish has its large solution successful 5 years.)

“I’m not convinced that it’s going to work," says Berkeley’s Recht, the Flourish adviser, of the company’s main mission. “But if it does, it would beryllium amazing.” AI would ne'er beryllium the same. And a batch of information centers mightiness autumn empty.

Let america cognize what you deliberation astir this article. Submit a missive to the exertion at [email protected].