A foundational tool for brain-like intelligence.
Author: Zhan Zhu
Word count: 3816
Handcrafted quantity丨100% AI content丨0%
Recently, you may often come across a cyber fruit fly "working under the table" in various scenarios.
In just a few days, this cyber fruit fly became the center of attention for developers online, playing "Super Mario," solving Rubik's cubes, driving... various videos showcasing its antics have emerged, covering almost all kinds of scenarios. Even a developer gave it $100, utilizing a dopamine reward and punishment mechanism to make this cyber fruit fly "trade cryptocurrencies."
The origin of this online sensation is the MaleCNS, a complete brain connectome of adult male fruit flies, released by Google and other institutions in early September. It includes 166,000 neurons and 125 million synapse connections, fully covering the central brain, visual lobes, and abdominal nerve cord. This is currently the largest human-scale and most complete biological brain connectome.
Image/Male fruit fly whole brain connectome
Physicist Richard Feynman once said, "What I cannot create, I do not understand." Starting from this fruit fly, ordinary people can also become "creators." Although the AI tasks, robot control, and game interactions based on the fruit fly's whole brain are mostly surface applications, they also prompt the industry to rethink a fundamental question from the perspective of real biological neural structures: How does intelligence actually arise?
However, MaleCNS is merely a static blueprint of the synaptic-level neural circuitry of male fruit flies, lacking the dynamic operation mechanisms such as neuronal firing processes and signal transmission. In other words, MaleCNS is a simplified model of the fruit fly's whole brain, still far from the true living biological brain.
In this area, a domestic company, Physical AI, which focuses on refined whole brain modeling, has taken the lead. Zhiyue Space Intelligence has continuously invested in this direction for two years and is currently the only company in China deeply engaged in this technological route.
Previously, I had an in-depth conversation with Dr. Qian Linrui, the founder and CEO of Zhiyue, where we discussed the past and present of brain-like intelligence, the technical barriers to refined whole brain connectome modeling, why Zhiyue insists on pursuing this difficult yet correct path, and early technical implementation scenarios.
Only two months have passed, and I learned that Zhiyue has completed a new round of financing amounting to hundreds of millions in Pre-A round funding, which has been oversubscribed by dollar funds, listed companies, and many old shareholders in just one year, achieving seven rounds of financing.
Transitioning from the angel round to the Pre-A round, the progression of financing rounds corresponds to dual breakthroughs in product and commercialization. At the WRC in August this year, Zhiyue released the DeepSoma, a foundational platform for brain science and brain-like intelligence research; on the commercialization front, it has accelerated from technical verification to real-world scenario delivery.
The Essence of Intelligence
Soma is an ancient Greek word meaning "body," referring to the cell body of neurons in neuroscience. On a small scale, intelligence is a neuron, a dendrite, a brain; on a larger scale, intelligence is a body, a life. In other words, we can have intelligence only when we have a body, enabling us to perceive this world.
This is also why researchers around the world focusing on brain-like intelligence are all engaged with the tiny body of the fruit fly. Although the fruit fly's brain is small, it possesses over a hundred thousand neurons and over a hundred million synapses, has capabilities like smell, vision, and flight, and can also engage in courtship, navigation, and even learning based on experience.
In short, though small, the sparrow has all its essential organs.
However, both Google's MaleCNS and the "digital fruit fly" released by Eon Systems in March of this year only replicated the structure of the fruit fly's brain, not fundamentally restoring how neurons work internally. This is precisely the underlying technological route that Zhiyue chose from the beginning.
Zhiyue's technological logic is very clear: the reason the brain generates intelligence stems from two core dimensions, "connectivity" and "neural dynamics."
The meaning of "connectivity" is "the topology structure of how neurons in the brain are connected," which serves as the "framework" of intelligence.
Neural dynamics refers to "the regularities of the physical and chemical states (membrane potential, ion channels, calcium concentration, etc.) of single neurons evolving continuously over time, and the ways in which multiple interconnected neurons, synaptic inputs, and couplings form collective states that evolve over time." This can be understood as what happens inside the neurons, representing the "blood" of intelligence.
The combination of structure and blood is what allows the brain to operate, evolve, and generate intelligence and behavior.
Why simulate the whole brain's connectivity structure? Because this connectivity structure is the optimal solution honed by billions of years of biological evolution, fundamentally different from mainstream AI reliant on massive amounts of pre-training data. For instance, the neural circuits of fruit flies, nematodes, and zebrafish are inherently suited to the demands of perception, obstacle avoidance, motion, and survival, with this structure itself being a priori intelligence.
This intelligent "framework," optimized over hundreds of millions of years, brings crushing efficiency advantages. For example, MaleCNS replicated the fruit fly’s whole brain with 166,000 neurons, which equates to approximately 0.026 billion parameters—far lower than the hundreds of billions or trillions of parameters seen in large models, and also significantly lower than the tens or hundreds of billions inherent in physical AI models.
Thus, in real scenarios, using such a native structure model to drive robots and drones does not require large-scale data set pre-training; it can be realized with just a dozen pieces of real data for fine-tuning. In contrast, using mainstream training methods such as VLA with world models requires data quantities starting in the hundreds of thousands of hours, and the results are not necessarily good.
This is what Zhiyue refers to as "instinct equals model." Many of the capabilities we need in physical AI or embodied intelligence are fundamentally the innate abilities brought by the brain's connectivity structure.
Image/Zhiyue's self-developed mechanical arm movement model driven by refined biological whole brain
But structure is merely the foundation. What truly brings the brain to "life" is neural dynamics.
Early AI oversimplified neurons to the level of "either firing or silence," treating them as binary switches and focusing only on the final outcome while discarding the complete biological computing process in between. The road may be easy, but it inherently carries limitations of data dependency and high computational power consumption.
Zhiyue chose the more difficult path, restoring the true physical and chemical processes of neurons, returning to the complete processes of continuous accumulation of membrane potentials, dynamic regulation of ion channels, and so on. Through breakthroughs in mathematical modeling and solving methods, they addressed long-standing industry challenges, allowing refined biological neuron models to become trainable, optimizable, and engineering applicable.
The ultimate effects are clear: compared to traditional models that rely on stacking parameters and data, this structure plus dynamics biological model can adapt rapidly to complex real environments using far less data and much lower power consumption, thus providing low computational needs, low power consumption, high generalization, low latency, and rapid response application advantages.
Image/Zhiyue's self-developed fruit fly whole brain model, completing training on a single behavior data in complex tasks, achieving zero-shot generalization in hundreds of scenarios
Foundational Tools for Brain-like Intelligence
In a prior conversation with Qian Linrui, she mentioned that Zhiyue ultimately aims to provide a world simulator, refined neuron modeling framework, developer tools, and biological general intelligence model capabilities. Just like ImageNet, PyTorch, Transformers, and CUDA once supported the previous generation of the AI ecosystem, biological general intelligence also needs its own models, toolchains, simulation environments, and developer communities.
Based on this understanding of the essence of intelligence and foundational technological accumulation, Zhiyue launched DeepSoma, a foundational platform for whole brain simulation aimed at brain science and brain-like intelligence research.
Image/A real fruit fly whole brain dynamics simulation and neuronal activity solving based on DeepSoma (no AI-generated effects)
The previously mentioned term "Soma," at a microscopic level, reflects how intelligence emerges from the dynamics of neurons and synapses; at a macroscopic level, intelligence relies on the body, rooted in interaction with the world. "Deep" represents the exploration of the most fundamental paths to intelligence and Zhiyue's open attitude, providing tools to more researchers to delve deeper into brain-like intelligence.
The core logic of DeepSoma is to establish a "perception-action" loop based on brain-like intelligence; Zhiyue believes that the next generation of intelligence will gradually form through the interactions among "environment-brain-agent."
In the past, AI existed in a digital world where data consisted of text, images, and videos. But as we enter the Physical AI era, agents must interact with the real physical world, so data must simultaneously contain spatial, temporal, semantic, and physical properties, and change in real-time; when an agent moves, the world changes, and the new state becomes new input.
The first layer of DeepSoma constructs a computable 4D real world. Starting from multi-modal input, it reconstructs the geometry and semantic information of the environment in real-time, then further estimates materials, friction, dynamics, and uncertainties. The "world" obtained this way not only looks like reality but also possesses participatory computable physical attributes, forming a closed-loop data set consistent with the real world.
The second layer is the complete implementation of whole brain structure plus neural dynamics. Starting from biophysical modeling of single neurons, the dendrites, cell bodies, membrane potentials, and ion channels are all transformed into computational models, which are then pieced together along the connectivity group to form complete neural circuits and whole brain networks. The self-developed solver advances all states on a unified timeline, allowing membrane potential changes and synaptic currents to become observable, recordable, and comparable quantities.
Its significance lies in transitioning intelligence models from "data-driven" to "structure and dynamics-driven." This means not using data to fit intelligence but rather replicating the structure and operational mechanisms of the biological brain, allowing intelligence to naturally emerge.
Finally, through a unified Agent interface, the same brain model can be mapped to different carriers like digital beings, robots, drones, and specialized equipment. This ultimately forms the complete Loop mentioned above: environment inputs into the brain, the brain drives the agent, the agent alters the environment, and the new environment feeds back into the brain.
This is completely different from the common industry "single scenario + single body" demo. DeepSoma is not about a single digital fruit fly but a universal foundational infrastructure, similar to how PyTorch was to deep learning, allowing developers to use this platform to create their own scenarios and applications without building all foundational tools from scratch.
Next-generation Intelligence
For Zhiyue, DeepSoma is not a conceptual toy in the lab but is intended for real-world application scenarios.
In contrast to long-term ecological layouts, Zhiyue's current commercialization pace is pragmatic and clear, following a gradual path of "capability accumulation through scenario validation and ecological openness." The first step anchors on validating high-barrier real-world models and platform capabilities.
These industrial scenarios commonly feature complexity, data scarcity, computational limitations, high deployment costs, and strong reliability requirements, making them blind spots for embodied intelligence and physical AI. The complete industrial system and rich scenario demands in China also provide a natural validation field for Zhiyue's technology.
Focusing on the landing demands for robots and various intelligent carriers, Zhiyue introduced the Omni-Brain intelligence system, serving as a full-link development foundation covering the complete process from data collection, processing, model training, deployment to real machine testing, supporting complex task execution and real-world scenario implementation.
Currently, Zhiyue primarily collaborates with vertical industry robot manufacturers for joint delivery, refining model stability, generalization ability, and adaptability in real operational scenarios. Previously, Zhiyue has completed POC validation in multiple high-barrier scenarios.
Based on the newly launched DeepSoma, Zhiyue can reverse the capabilities of single-point scenarios into universal platform capabilities. Real operational environments provide diverse scenario data, simulation systems offer low-cost trial and error spaces, model frameworks provide standardized training capabilities, and robotic platforms offer real execution feedback.
With an increasing number of landing scenarios, the platform continues to accumulate reusable neural structures, simulation tools, training methods, and task adaptation solutions, gradually transitioning from "single-project delivery" to "platform capability output." The fact that Zhiyue completed Pre-A financing in less than three months after the angel round 5+ signifies that investors recognize Zhiyue's second leap in commercialization pathways, and it also represents a new valuation of AI's new technological route by the capital market.
In the future, Zhiyue will build more developer and industrial ecological systems. Once the frameworks for refined neuron modeling, world simulation, and brain-like model capabilities mature, tools and capabilities will gradually be opened to serve robotics companies, brain science teams, life science institutions, and various industry developers.
Ecological partners will not need to grasp the intricate details of computational neuroscience from scratch to directly utilize simulation environments, model capabilities, and training tools, quickly implementing their own applications and lowering the barriers to using whole brain modeling technologies.
Looking back at the recent online frenzy around cyber fruit flies, the greatest value is that onlookers intuitively perceive the possibility of digitizing biological brains once again. The 166,000 neurons of the fruit fly are just the starting point; in the future, humans will tackle the brains of zebrafish, mice, and even humans. However, to explore the essence of how intelligence emerges, one must delve into every computational unit within the connections between neurons. Through foundational modeling of biological brain structures and dynamics, Zhiyue endeavors to construct a truly functional and evolvable biological brain model. This is also the direction Zhiyue believes current brain-like intelligence needs to further deepen and break through.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
