Jeff Bezos is funding the Wild Hunt of the Core Algorithm of the Brain

Rob Williams knows how to sell Jeff Bezos: You write a press release like your product has been built. After Bessos has read it, he raises his thumb or his thumb.

Williams asAmazonS-team's executives went through a lot of this process, taking responsibility for software products such as Alexa until they left last fall. But a few weeks later (December 2025) his propaganda was different. He now works with neuroscientists and re-entrepreneur Thomas Reardon to contact Bessos as a financier rather than a boss.

Here's what Bezos read on his yacht, while Williams watched anxiously on Zoom:

Flourish is a neuro-intellectual company dedicated to two of the most difficult issues facing artificial intelligence today: efficacy and continuous learning. We're building Cortex AI, the first artificial intelligence system to match human brain computing, learning efficiency and power budget.

A month later, I had lunch with Reardon and Williams in the ironing area of New York City. Reardon opened the door. He said artificial intelligence had dug a hole in itself. Despite the growing power of large language models, it is a greedy consumer of computer capabilities and data.

♪ Though ♪LL.M.The inspiration comes from biology, but current frontier models have little in common with the human brain. (b) A single chip in AI training cluster uses more than 30 times that amount. Mega-scale enterprises require thousands of chips and gigawatts of energy to supply electricity to small cities. These models need to incorporate virtually all human-generated content. Each new model needs more, more and more. Nevertheless, the model does not learn. Once you train them, they'll be in trouble.

Reardon told me that our goal is to build "a synthetic artificial brain with a power of less than 50 watts." It should adapt to the environment, be as flexible as human thinking, and only consume. LLM A small fraction of computing capacity and energy. Conceptual validation is flourishing in our minds. ""In order to learn English, I basically need to read all the books I've written 20 times." Reardon said. “A human baby can do this in hundreds of thousands of words.”

Reardon and Williams have yet to figure out how to build systems that match the magic of the human brain. They believe that a well-resourced team of experts, consisting of artificial intelligence researchers and neuroscientists, works largely side by side to find answers. Neurologists will use some of the most advanced laboratory equipment for original wet laboratory experiments to find available information on brain structure. They plan to publish the model that is currently being developed as a recent product for a comprehensive re-engineering of artificial intelligence.

The vagueness of the proposal does not bother Jeff Bessos. After reading Williams' two pages, he donated $50 million. Other funds are from Lux Capital and Google Ventures etc. Subsequently, Bessos almost doubled his initial shares and told Reardon that he would give more if they so requested. Now, with $500 million in money and $2.5 billion in valuation, we just need to invent a new artificial intelligence method.

Thomas Reardon IV did not use his name — too many Tom's names in the family history. "My wife calls me Reardon, everyone calls me Reardon," he says. He was one of 18 children in a working-class family who dropped out of New Hampshire at the age of 15. Since then, his résumé has become crazy: he's become a teenage programming genius, hired to help build it. Microsoft The first web browser and a wireless technology company were established and sold. He then went to Columbia University to pursue his degree in classical literature, enters the field of neuroscience and eventually obtains a doctorate (also from Columbia University). He started another company with some classmates, developed a mind-control wristband. Meta. Buy and work there for six years. The wristband with Meta's latest smart glasses.

But Reardon is not satisfied with how companies, including Meta, build sophisticated artificial intelligence. It's not new to match brain learning with energy savings. IBM and Intel both released neurographic chips inspired by brain structures. Ben Recht, a computer scientist at the University of California at Berkeley, was a consultant to Flourish, who recalled that scientists had been working on neuromorphology software for decades. Then the Master of Laws took over. "They call it a neural network, but there's no brain-like thing going on there," Rehit says.

Reardon convinced Amazonian executive Williams to join him, whom he met when Microsoft worked. Another early recruiter was Greg Wayne, who was... DeepMind Permanent researcher responsible for Go.The ogle's AI Assistant Plan Astra project. "I don't know if they can achieve their goal, but I think it'll be fun, and it might be useful," Wayne said. Demis Hassabis, CEO of DeepMind, strongly retained Wayne, who had reached an agreement to keep her job but spent 20 per cent of her time in Fluush.

By the end of March, Reardon had employed about two dozen top neuroscientists and artificial intelligence researchers. On the day the company moved into office space in the Soho district west of New York City, I visited them in a 10-storey building with a built-in data centre. People are setting up their computers; laboratory equipment such as electronic microscopes have not yet arrived.

"The brain has a secret we haven't discovered," Wayne said. The team focused on the structure known as the cortical column, which a scientist from Fluush called the "normative calculator" of the brain. One of the investors in Flourish was Jacob Vogelstein, a neuroscientist who was transformed into a venture investor, who, together with his brother Joshua and others, launched an ambitious project called Open Connectome Project. "The idea is that you can collect all these brain images and start processing them to try to explain the circuit," he said.

This work may eventually be useful to the team. Joshua Vogelstein, co-founder of Fluush, recently co-authored a paper on the fruit fly neural network, which found that its network was 10 times more efficient than Transformer, the core J.L.A. architecture module. “These methods are at a turning point”, said Nathan Danielson, who is a French neuroscientist and doctor who worked with Reardon in Meta.

Fluush isn't the only one looking for answers in the brain. The term “neurological form” has been widely used and has become almost a popular word. A company called Cortical Labs is combining laboratory neurons with silicon chips. OpenAI Chief Executive Officer Sam Altman supported Merge Labs, whose “long-term mission is to connect biological and artificial intelligence”. Meta's super-smart team claims that its Tribe v2 model “Acting as a digital twin of human neuroactivity”. An organization called Unconventional AI is creating a research support scheme to achieve the goal of Reardon: to build artificial intelligence capable of replicating bioefficiency. Some venture capital companies even focus on the brain sciences.

Reardon believes that the strength of the company lies in its exceptionally strong team of neuroscientists. These researchers will conduct laboratory experiments, while artificial intelligence teams will build models based on their findings; at the same time, algorithm teams may find clues that can help neuroscientists. The French neuroscientist Josh Morgan said, "You don't really know if you understand it unless you can build it and achieve it in silicon." They offered to publish some original research.

“The company is essentially looking for the algorithm behind its intelligence,” said Jacob Vogelstein, an investor.

Reardon told me his team had found a way to make short-term income from recent brain research. He said that they were developing a memory-processing method inspired by seahorses, which would enable company models to learn without much training data. He said that the Algorithm team had developed a model for continuous learning and was committed to putting it in “the various devices in your pocket”. He added that he was negotiating with a major chip manufacturer to put the model on the silicon.

Early in May, a plenary meeting was held at the New York office of the Fluush scientist. Reardon and Williams sat at a table with a dozen others, including Wayne, a visiting senior adviser from London. Scientists are discussing six possible experiments. These enormous fluctuations require the purchase of millions of dollars of microscope machines and years of work.

The experimental discussion raised as wide a biological phenomenon as how rabies spread in the cortex and the neurobiology of birds. They argued whether it was desirable to check molecules and synapses or to focus on larger cells or circuits. Is the analysis of interconnectors in the brain of mice sufficient for certain purposes, or can only be done by the human brain?

At least for the time being, the conclusion seems to be an attempt. “We concluded that we wanted to collect data on nano, micro and meso scales to support the discovery of core algorithms,” says Sean Bittner, a computational neuroscientist who worked with Reardon at Meta.

An awkward pause, then Wayne spoke. "It's a really great experiment!" he said. "It's actually practical, not crazy."

There's no doubt that Fourish is running an adventure long-term bet. As Williams said, after reading the promotion document, Bezos needs to know one thing — is the founder committed to spending years in this regard? When he received firm confirmation, he agreed to release his millions of dollars. "You can't do much in three years," Williams said. “To make a major change, it is necessary to plan for something worth doing in seven to ten years.” Reardon says he wants the Fluush to find an important solution in five years.

“I don't believe it's going to work,” Recht, an adviser to Fluush at Berkeley, said, “The company's main mission. "But if that's true, that's great." Artificial intelligence will never be the same. Many data centres may be vacant.

Let's get your views on this article. Submits a letter to the editor via [email provided].