The Organic–Synthetic Brain Atlas

**Neural Computation Across Wetware, Neuromorphic Hardware, Connectomic Mapping, Semantic Translation, and Interface Systems — 2026** *This document is a working reference. It traverses the present technical state of nervous-system-grounded computation across two convergent substrates — living neural tissue and engineered neuromorphic hardware — together with the connectomic, imaging, modeling, semantic-decoding, and interface infrastructure that links them. The architecture is layered rather than linear: each movement is internally self-contained, the document as a whole resolves into a single technical regime. Specific systems, laboratories, instruments, and numerical specifications are integrated into the prose rather than enumerated separately, because the field is not a list; it is an emergent architecture in which the named platforms are nodes within a larger structural transition. The synthesis is reserved for the closing movement, where the convergence is no longer a claim but the natural consequence of the accumulated material.*

## Movement I — The Two-Substrate Frame The contemporary architecture of brain-grounded computation runs along two distinct substrates whose trajectories are now demonstrably converging. The first substrate is **living neural tissue** organized for computation: dissociated cortical cultures and three-dimensional brain organoids interfaced with high-density microelectrode arrays, embedded in closed-loop digital environments, and increasingly accessible as cloud-rentable wetware. The second substrate is **engineered neuromorphic hardware**: spiking neural network processors, memristive and diffusive devices, analog and mixed-signal silicon, photonic computing platforms, and software simulators built on the architectural principles of biological neural networks but instantiated in non-biological media. The historical assumption that these substrates would remain methodologically separate — wetware as biology, neuromorphic chips as engineering — is being dissolved by a shared technical regime in which both substrates are addressed through similar tools, evaluated against similar benchmarks, and increasingly integrated into hybrid architectures that exploit the complementary advantages of each. The organic substrate excels at adaptive plasticity, intrinsic energy efficiency, and developmental self-organization. The human brain operates at approximately twenty watts of total power consumption while performing forms of perceptual integration, sensorimotor learning, and semantic generalization that remain costly for the largest synthetic systems. Living neurons exhibit ion-mediated computation, dendritic compartmentalization, neuromodulatory dynamics, sparse activity, and synaptic plasticity that are difficult to fully replicate in silicon. The synthetic substrate excels at scalability, standardization, reproducibility, and engineering control. Neuromorphic chips can be fabricated in numbers that biological systems cannot match, programmed precisely, deployed in environments where living tissue cannot survive, and instrumented in ways that biological systems resist. The convergence is therefore not a competition between substrates but a structured division of labor in which the strengths of each domain compensate for the limits of the other. What makes the present moment distinctive is the development of an infrastructural layer that addresses both substrates as forms of **computationally addressable neural architecture**. The same connectomic reconstruction pipelines that map biological tissue at synaptic resolution are now generating wiring diagrams that constrain synthetic models. The same machine-learning architectures used to train artificial neural networks are being applied to decode signals from living neural tissue. The same multielectrode arrays that interface cultured neurons with simulated environments are being integrated with neuromorphic chips to create hybrid bio-silicon systems. The same imaging modalities that monitor biological activity — calcium imaging, voltage imaging, functional ultrasound, functional near-infrared spectroscopy, electroencephalography, electromyography — are being co-developed with the interface technologies that will eventually allow bidirectional coupling between nervous systems and machines. Google Research's MoGen model, presented at ICLR 2026, occupies a specific position within this convergence. MoGen is a flow-matching generative model that produces high-resolution three-dimensional point clouds of mouse cortex axon and dendrite fragments, using a scalable latent transformer backbone augmented with local geometric context injection to generate biologically plausible neuronal morphologies. When millions of MoGen-generated synthetic morphologies are added to the training set of a shape plausibility classifier inside Google's production connectomics neuron reconstruction pipeline, the classifier's accuracy improves and the number of remaining split and merge errors decreases by 4.4 percent — an improvement that at whole-mouse-brain scale corresponds to over 157 person-years of saved manual proofreading labor. The technical significance is that morphological scarcity, which has been a chronic bottleneck in connectomic reconstruction, has been broken as a constraint. Synthetic neuronal geometry now accelerates the reconstruction of biological neuronal geometry, and the resulting reconstructions in turn feed the next generation of models that constrain synthetic architectures. This recursive structure is the operative pattern across the entire field. Synthetic data accelerates reconstruction of biological structure; biological structure constrains synthetic models; synthetic models inform biological experiments; biological experiments produce data that further trains synthetic systems. The flywheel is no longer hypothetical. It is operational across connectomics, neuromorphic computing, brain-computer interfaces, organoid intelligence, semantic decoding, and computational neuroscience. What follows is the layered technical atlas of that flywheel as it stands in 2026 — substrate by substrate, system by system, modality by modality — beginning with the organic side, where living neural tissue is being industrialized as computational substrate for the first time in the history of the field. --- ## Movement II — The Organic Substrate: Living Neural Tissue as Computation ### II.1 — The Substrate Itself: From Dissociated Cultures to Three-Dimensional Organoids The simplest form of organic neural computation begins with **dissociated cortical cultures**: primary neurons extracted from embryonic or postnatal rodent cortex, plated as two-dimensional monolayers on planar substrates, and allowed to form spontaneous synaptic connections during a maturation period of one to several weeks. Such cultures generate spontaneous spiking activity, develop network bursts, and exhibit plasticity in response to electrical or chemical stimulation. When grown directly on planar **microelectrode arrays** (MEAs) — substrates patterned with grids of extracellular electrodes typically spaced at intervals between 30 and 200 micrometers — these cultures become simultaneously observable and manipulable: the electrodes both record extracellular field potentials from nearby neurons and deliver electrical stimulation, allowing closed-loop interaction with the living network. Modern high-density MEAs developed by companies such as MaxWell Biosystems and 3Brain offer thousands to tens of thousands of electrodes per chip, with sub-millisecond temporal resolution and the ability to resolve single-unit activity from individual neurons. The resulting platform is the foundational instrumentation of in vitro neural computation: an interface between living tissue and digital systems at the resolution of individual cells. The substrate-source distinction matters technically. **Rodent primary cortical neurons** — typically extracted from embryonic mouse or rat brains — develop quickly, form robust networks within ten to fourteen days, and have served as the canonical experimental preparation for decades. **Human neurons derived from induced pluripotent stem cells** (iPSCs) require longer maturation periods but offer species-specific physiology, the ability to incorporate patient-derived genetics, and an ethically distinctive position relative to the use of fetal or animal tissue. iPSC-derived neurons are differentiated from skin or blood cells reprogrammed back into a pluripotent state, then directed along neuronal fate trajectories using defined growth factor combinations. The resulting cultures can be maintained for months and increasingly serve as the default substrate for biological computing research focused on human-relevant physiology. Three-dimensional **brain organoids** represent the next architectural step. Generated from iPSCs under conditions that permit self-organization into cellular architectures resembling developing brain tissue, organoids contain multiple neuronal subtypes, support cells including astrocytes and oligodendrocytes, and exhibit emergent structural features such as cortical-layer-like organization, electrically coupled neuronal networks, and oscillatory population dynamics. Standard cortical organoids contain on the order of tens of thousands of cells and reach diameters of several millimeters before central necrosis from limited diffusive nutrient access constrains further growth. Vascularized organoids, assembloid configurations combining multiple regional organoid types, and microfluidic perfusion strategies are extending viable size and longevity. As biological computing platforms, organoids offer richer physiology than dissociated cultures, more cell-type diversity, and structural features closer to native brain architecture — at the cost of greater experimental complexity, longer maturation times, and currently limited control over precise cellular composition. The historical seed of in vitro neural computation runs through the work of Steve Potter's laboratory at Georgia Tech in the early 2000s, where dissociated rat cortical cultures on multielectrode arrays were coupled to robotic or simulated environments in closed-loop configurations described as **hybrots** — hybrids of living neural networks and engineered systems. The cultures generated activity that was decoded to drive robot movement or virtual-animat behavior, while sensory feedback from the environment was delivered back to the network as patterned electrical stimulation. The hybrot work established the foundational principle that living neural networks can be operationally embedded in task-structured environments and studied as adaptive computational systems rather than purely as biological specimens. The methodological inheritance of that line is now visible across every major contemporary platform in the field. ### II.2 — DishBrain and the Cortical Labs CL1: Synthetic Biological Intelligence The decisive contemporary demonstration of in vitro neural computation came from **DishBrain**, developed by Cortical Labs in Melbourne in collaboration with Brett Kagan and academic partners. The 2022 *Neuron* paper described a platform in which approximately 800,000 neurons — sourced either from human iPSCs or from mouse primary cortical tissue — were cultured on high-density multielectrode arrays and embedded in a closed-loop simulated environment running a variant of the classic arcade game Pong. The cultures received structured electrical input encoding game state (paddle position and ball trajectory) through spatially organized stimulation patterns, and their evoked spiking activity was decoded to control paddle movement. Under closed-loop feedback in which predictable feedback was provided for paddle hits and unpredictable noise for paddle misses, the cultures exhibited improving game performance — longer rallies, more accurate paddle positioning, more responsive engagement with the simulated ball — within minutes of session onset. The work demonstrated, with peer-reviewed methodological rigor, that living neural networks can be operationally coupled to digital environments and exhibit adaptive computational behavior under structured feedback. The DishBrain result has since been commercialized as the **CL1**, which Cortical Labs describes as the world's first code-deployable biological computer. Announced in late 2024 and launching commercially across 2025 and into 2026, the CL1 integrates lab-grown human neurons cultivated on a custom silicon chip with embedded compute, fluid management, gas exchange, and temperature control inside a self-contained desktop or rack-mountable enclosure. The CL1's modular enclosure supports both desktop and rack-mounted configurations, the precision-controlled fluid management system sustains neural cultures for up to six months, and the unit is fully self-contained with embedded processing requiring no external compute. The neurons grow directly on the silicon substrate, receive electrical signals representing the digital environment in real time, and respond with electrical activity that the system decodes back into the simulated world. The full operating environment is called the Biological Intelligence Operating System (biOS), and Cortical Labs has framed the underlying paradigm as Synthetic Biological Intelligence (SBI) — a deliberately distinct category from conventional silicon-based artificial intelligence. The infrastructural significance of the CL1 sits in its dual deployment model. Individual units can be purchased and operated by laboratories with appropriate biological expertise, while a **Cortical Cloud** offers remote access to networked CL1 systems for researchers without local cell-culture infrastructure. The cloud model is being developed as a "wetware-as-a-service" approach in which scientists, developers, and organizations interact with biological neural systems remotely without maintaining their own culture facilities. The commercial trajectory specifies first racks shipping on schedule, 95 percent uptime targets for the cloud platform, and an initial focus on personalized medicine, drug screening, robotics, and energy-efficient AI workloads. The CL1 thereby instantiates a complete vertical integration: living neural tissue, custom silicon interface, embedded compute, biological life support, an operating system, a programming model, and a cloud delivery layer — assembled into a product that researchers can purchase, code against, and deploy. ### II.3 — FinalSpark Neuroplatform: Cloud-Accessible Organoid Computing The Swiss company FinalSpark, founded in 2014 by Fred Jordan and Martin Kutter in Vevey, operates the most fully developed cloud-accessible biological computing platform currently in public use. FinalSpark develops experimental biological computing systems based on human brain organoids interfaced with electronics, using actual living human neurons — lab-grown clusters of brain cells connected to multi-electrode arrays — as the computational substrate, distinct from neuromorphic hardware companies that mimic neural behavior in silicon. The platform consists of multi-electrode arrays each hosting up to four brain organoids derived from human iPSCs, with the full system housing sixteen organoids simultaneously and each organoid containing approximately 10,000 living human neurons. FinalSpark presented its ten-year roadmap in London in June 2025, with a vision of bio-servers accessible via cloud to provide computational power for generative AI; the Neuroplatform launched as the world's first commercially accessible biocomputing research platform on May 15, 2024, and contains 16 brain organoids with approximately 160,000 living human neurons interfaced with electrodes, all using roughly one million times less energy than equivalent silicon chips. The technical workflow is described in FinalSpark's 2024 paper in *Frontiers in Artificial Intelligence*. Neurons are differentiated from human iPSCs obtained from skin samples, which spontaneously form three-dimensional structures called organoids; these can be stored frozen or in maintenance medium at 37°C, and when deployed as computational substrate they are thawed, plated, and expanded in T25 culture flasks before placement on electrodes. FinalSpark's Neuroplatform places wetware biocomputing into a cloud-accessible format, allowing researchers and commercial users to purchase time with the brain chips using a Python-based software interface. Over the past three years FinalSpark has scaled the Neuroplatform to include over 1,000 brain organoids, enabling collection of more than 18 terabytes of data, and the platform is currently the only such system providing remote access to living neural cultures at this scale. The training methodology incorporates **dopamine-mediated plasticity**: when the organoid generates outputs aligned with task objectives, the platform delivers chemical or electrical reward signals analogous to dopaminergic neuromodulation, reinforcing the connection patterns that produced the correct response. Research groups using the Neuroplatform have reported being able to record organoid activity, manage large volumes of continuous 24/7 data, and deliver tailored electrical stimulation to modulate behavior, with the platform removing the barrier of specialized cell-culture expertise that previously limited access to biological neural networks. Current academic collaborations include groups at Bristol (Benjamin Ward-Cherrier, who published peer-reviewed work using the platform in December 2025), Lancaster University Leipzig (researching organoid applications in AI learning models), Oxford Brookes University (dynamic systems control research), Université Côte d'Azur's NeuroMod Institute (organoid connectivity pattern recording), and EPFL. ### II.4 — Brainoware and Organoid Reservoir Computing The most academically formalized demonstration of computation in brain organoids was published in *Nature Electronics* in December 2023 by the laboratories of Feng Guo at Indiana University Bloomington and Mingxia Gu at Cincinnati Children's Hospital. **Brainoware** is a hybrid computing architecture in which mature human brain organoids are placed on high-density multielectrode arrays and used as adaptive living reservoirs within a reservoir computing framework. The system performs reservoir computation by sending and receiving information from the brain organoid using a high-density multielectrode array, achieving nonlinear dynamics, fading memory properties, and unsupervised learning from training data by reshaping the organoid's functional connectivity through spatiotemporal electrical stimulation. In reservoir computing, an input signal is projected into a high-dimensional dynamical system (the reservoir) whose internal state encodes a rich nonlinear transformation of the input history. A simple linear readout is then trained to extract task-relevant features from the reservoir's state, exploiting the reservoir's intrinsic dynamics rather than training the network itself. Living neural tissue is exceptionally well-suited to this role because biological neuronal networks exhibit precisely the properties reservoir computing requires: high-dimensional nonlinear dynamics, fading memory of recent inputs, and intrinsic adaptation in response to stimulation history. Brainoware consists of brain organoids connected to an array of high-density microelectrodes, using a type of artificial neural network known as reservoir computing: electrical stimulation transports information into the organoid, the reservoir wherein that information is processed before Brainoware outputs its calculations in the form of neural activity. The published benchmark involved a Japanese vowel speech-recognition task: 240 audio clips of eight male speakers pronouncing Japanese vowel sounds were encoded as spatiotemporal electrical stimulation patterns delivered to the organoid, and the system was trained to identify which speaker had produced each clip. After training, the organoids could complete the speaker-identification task with over 70 percent accuracy. Additional benchmarks included nonlinear chaotic equation prediction, demonstrating that the organoid's intrinsic dynamics could be exploited for tasks well beyond simple pattern recognition. The system achieved nonlinear dynamics, fading memory properties, and unsupervised learning, demonstrating practical potential in tasks such as nonlinear equation prediction and speech recognition by sending and receiving information from the brain organoid using a high-density multielectrode array. The training-induced changes in the organoid's functional connectivity persisted after stimulation ceased, suggesting genuine biological plasticity rather than purely transient state changes. The architectural insight from Brainoware is that brain organoids do not need to be programmed in the conventional sense to perform computation. Their intrinsic biophysical properties — diverse cell types, recurrent connectivity, ion-channel-mediated nonlinearities, synaptic facilitation and depression, intrinsic excitability — already provide the computational primitives that reservoir computing exploits. The engineering task is to design the input encoding, the readout decoding, and the training feedback structure such that the organoid's natural dynamics map productively onto the target task. This stands in contrast to the more conventional view in which biological neural networks must be made to imitate digital computing in order to be useful; Brainoware demonstrates that the biological substrate is computationally productive in its own right, and the engineering challenge is to learn how to use it. ### II.5 — Organoid Intelligence as a Field The umbrella term **organoid intelligence** (OI) was formalized in a 2023 *Frontiers in Science* paper led by Thomas Hartung at Johns Hopkins University, with co-authors including Lena Smirnova, Brett Kagan, and a multi-institutional consortium. The proposal frames organoid intelligence as the new frontier of biocomputing, calling on researchers to explore its potential to advance understanding of the brain and to unleash new forms of biocomputing while addressing the associated ethical implications. Hartung and colleagues identify one critical step toward organoid intelligence as the scaling up of organoid structures by approximately 100-fold to generate organoids containing 10 million cells capable of sophisticated computations, which would require artificial blood supply for organoids to grow to this size. The Johns Hopkins consortium has subsequently issued the Baltimore Declaration on organoid intelligence and conducted the first OI workshops to form a research community. Recent advances in stem cell technology, bioengineering, and machine learning enable exploration of brain organoids' ability to compute and store given information, execute tasks, and study how computational tasks affect structural and functional connections in the organoids themselves. The framing is deliberately distinct from artificial intelligence: rather than building synthetic systems that imitate brain function, OI uses actual brain tissue (in three-dimensional self-organized form) as the computational substrate, with machine learning serving as the interface layer that translates between organoid activity and task structure. Organoid intelligence blends lab-grown neurons with machine learning, and biochips that integrate living brain cells into hardware have the potential to outstrip silicon-based processors in both efficiency and adaptability. The University of Bristol demonstration in 2025, conducted in collaboration with FinalSpark, exemplifies the practical reach of contemporary organoid intelligence work. The Bristol team integrated FinalSpark organoid systems into a real-time Braille-character-reading robotic platform: tactile sensor data from a robotic fingertip was encoded as electrical stimulation patterns delivered to the organoid, and the organoid's response activity was decoded to classify the Braille character being touched. The work demonstrated that organoid-controlled robotic perception is feasible at the level of meaningful sensorimotor tasks, and that the cloud-accessible architecture of the Neuroplatform makes such applications available to robotics laboratories without local biological infrastructure. The Bristol work is one of several contemporary integrations of organoid computation with sensor and effector systems, including organoid-eye configurations using retinal organoids for visual input, organoid-motor configurations for closed-loop control, and organoid-organoid assembloid architectures that combine multiple regional organoid types. The field is governed by a roadmap that prioritizes scaling, longevity, vascularization, and interface bandwidth. Current organoids of approximately 50,000 cells need to be scaled by orders of magnitude to support computational workloads of practical scale. Organoid lifespans currently in the range of 100 to 200 days need to be extended through improved nutrient delivery and waste removal. Microelectrode interfaces with hundreds to thousands of channels need to be extended to tens of thousands of channels with the spatial resolution to record from individual cells inside the three-dimensional structure. The bioethical layer requires explicit attention because of the ambiguity surrounding the developmental and experiential status of large, long-lived, sensory-equipped organoids — a question the field has been transparent about confronting rather than avoiding. ### II.6 — The Substrate-Source Pipeline: iPSCs, Differentiation, and Maintenance The biological infrastructure underlying organic neural computation rests on **induced pluripotent stem cell** technology, which itself rests on Shinya Yamanaka's 2006 discovery that mature somatic cells can be reprogrammed to a pluripotent state through forced expression of four transcription factors. iPSC-derived neurons are now routinely generated from skin biopsies or blood draws, allowing patient-specific neural cultures, organoids derived from individuals with specific genetic conditions, and a virtually unlimited supply of human neural tissue without the ethical and logistical constraints of fetal-tissue sourcing. Standard differentiation protocols direct pluripotent cells toward forebrain, midbrain, hindbrain, or spinal cord identities using sequential application of small molecules and growth factors that recapitulate developmental signaling pathways. The resulting cultures can be maintained in defined media for months, frozen for long-term storage, and shipped between laboratories — a logistical infrastructure that did not exist a decade ago and that now constitutes the supply chain for the entire organic computing field. Maintenance of viable cultures requires regulated temperature (typically 37°C), controlled gas composition (5 percent CO₂ for buffered pH), nutrient-rich media exchanged on regular schedules, and sterile handling to prevent bacterial or fungal contamination. The CL1 integrates this life support directly into the device, eliminating the need for external incubators and cell-culture infrastructure; the FinalSpark Neuroplatform handles maintenance centrally and exposes computational access remotely; academic laboratories typically run their own cell-culture facilities with associated technical staff. The trajectory is toward integration of biological life support into the computing device itself, which is the architectural pattern that the CL1 has now established and that competing platforms are following. The cost structure of organic computing is dominated by the labor and infrastructure of cell culture, not by the cells themselves. iPSCs are increasingly available as commercial reagents, organoid differentiation protocols are widely published, and the bottleneck is the skilled technical workforce required to maintain cultures over the months-long timescales relevant for sustained computation. Automated culture systems, microfluidic perfusion platforms, and AI-driven culture monitoring are reducing the per-organoid labor cost, but the field remains constrained by the inherent biological time required for neurons to mature into functionally networked tissue. This is one of the fundamental asymmetries between organic and synthetic substrates: synthetic chips can be fabricated in days and operated immediately, while organic systems require weeks to months of biological maturation before they become computationally productive. The trade-off is energy efficiency and adaptive plasticity for time-to-deployment. ### II.7 — The Closed-Loop Architecture as Methodological Standard The methodological signature shared across all major organic computing platforms — Hybrots, DishBrain, CL1, FinalSpark Neuroplatform, Brainoware — is the **closed-loop architecture**. Living neural tissue receives structured electrical input encoding some aspect of a digital environment or task; the tissue generates electrical output that is decoded into actions, classifications, or predictions; the consequences of those outputs are fed back to the tissue as further structured input, creating a continuous loop in which the tissue and the environment co-adapt over time. The closed-loop architecture is not incidental; it is the methodological mechanism by which biological plasticity can be harnessed for computation, because intrinsic neural plasticity responds to patterns of input-output correlation rather than to static input alone. This methodological standard has implications beyond organic computing proper. The same closed-loop architecture is used in neuroprosthetic systems, in adaptive brain-computer interfaces, in deep-brain stimulation research, and in the structure-function alignment work that connects connectomics to dynamic recordings. The convergence of methodology across what were previously separate fields — in vitro biology, neuroprosthetics, computational neuroscience, neural engineering — is one of the structural signals that the organic computing layer is no longer a fringe research direction but a developing standard infrastructure for the broader nervous-system-grounded computing regime. --- ## Movement III — The Connectomic Substrate: Mapping the Wiring Diagram ### III.1 — Connectomics as Structural Prerequisite The field of **connectomics** is the systematic effort to map nervous systems at cellular and synaptic resolution, producing wiring diagrams that render the physical substrate of biological computation as navigable graph architecture. A connectome consists of neurons as nodes, synapses as edges, and the full set of connections rendered as an addressable topological object. The conceptual term was formalized in a 2005 paper by Olaf Sporns, Giulio Tononi, and Rolf Kötter, who argued that a comprehensive structural description of the network of elements and connections forming the human brain would constitute an indispensable foundation for cognitive neuroscience. The decades since have transformed this aspirational vision into operational reality across multiple model organisms, with techniques and infrastructure scaling toward the eventual goal of whole-mammalian-brain mapping. The foundational demonstration remains the *Caenorhabditis elegans* nervous system, whose 302 neurons and approximately 7,000 synapses were reconstructed from serial-section electron micrographs and published in 1986 by John White, Eileen Southgate, Nichol Thomson, and Sydney Brenner. The worm connectome required years of manual reconstruction effort and established the methodological seed for everything that has followed: slice the tissue thinly, image each section at sufficient resolution to resolve fine processes and synaptic specializations, align the resulting stack of images into a coherent three-dimensional volume, trace each neuron through that volume, identify synaptic connections, and assemble the graph. Every advance since has been an escalation in scale: more cells, more synapses, larger tissue volumes, faster imaging, more automated reconstruction, more sophisticated proofreading workflows. The methodological core has remained essentially the same. The scale gap from worm to mammal is severe. The fruit fly brain contains approximately 140,000 neurons in a volume of about 0.0175 cubic millimeters. The mouse brain contains roughly 70 to 100 million neurons in a volume of approximately 500 cubic millimeters — about four orders of magnitude larger than the fly. The human brain contains an estimated 86 billion neurons in a volume of roughly 1,200,000 cubic millimeters — an additional thousand-fold scale increase beyond the mouse. The Wellcome Trust report on mouse connectomics estimated that conventional approaches would require approximately twenty electron microscopes running continuously in parallel for five years, generating about 500 petabytes of data, with imaging costs of \$200–300 million and proofreading costs of an additional \$7–21 billion, and a total project duration of up to seventeen years. Such projects with the eventual goal of mapping a whole-brain mouse connectome have recently launched, prompting discussions about the eventual goal of human brain connectomics. The advances in imaging throughput, machine-learning segmentation, and synthetic data augmentation that MoGen exemplifies are precisely what compresses those projected timelines and budgets into something approaching feasibility. ### III.2 — The Fruit Fly Threshold: FlyWire and the Adult Drosophila Brain The first whole-brain mapping of a complex behavior-capable organism was achieved through the **FlyWire** consortium, a multi-institutional collaboration led by Sebastian Seung and Mala Murthy at Princeton University, with contributions from MRC Laboratory of Molecular Biology in Cambridge, Google Research, and approximately fifty laboratories worldwide. The full adult female Drosophila melanogaster brain connectome was published as a nine-paper package in *Nature* on October 2, 2024. Neuroscientists reconstructed the first complete wiring map of the fruit-fly brain, including 140,000 neurons and more than 50 million connections, a resource that has already begun to revolutionize the field. As of October 2024, the flagship FlyWire paper "Neuronal wiring diagram of an adult brain" was published in Nature and includes 139,255 proofread neurons, accompanied by companion papers providing hierarchical annotation of all proofread neurons and a comprehensive cell type catalog of the visual system. The technical achievement combined automated machine-learning segmentation with human-driven proofreading at unprecedented scale. Researchers used electron microscope images of the complete fly brain that had been publicly released, applied a computer program to automatically identify or segment the neurons in the images, then created a computational system of tools that allowed a large online research community to look at the segments, proofread them for accuracy, and annotate cell types and classes in a community-driven manner. In 2021, only 15 percent of the neurons had been proofread; opening proofreading to the larger scientific community studying the fruit fly brain greatly accelerated completion of the connectome. About 85 percent of neurons in the fruit fly connectome are intrinsic to the brain — synapsing only with other brain neurons — while intrinsic neurons varied in length from less than 0.2 millimeters to almost 20 millimeters and in volume from 16 to more than 3,000 cubic micrometers, with most individual neurons synapsing in only a few regions of the brain even though each region may connect to many others. The companion annotation paper complements the approximately 140,000-neuron FlyWire whole-brain connectome with a systematic hierarchical annotation of neuronal classes, cell types, and developmental units (hemilineages), identifying 8,453 annotated cell types of which 3,643 had been previously proposed in the partial hemibrain connectome and 4,581 are new types, mostly from brain regions outside the hemibrain subvolume. The fly brain is now divided into 78 regions called neuropils, each involved in functions supporting behaviors such as sight, flight, and memory, and the FlyWire platform enables researchers to query and simulate the activity of specific circuit motifs in ways that were previously impossible. The connectome has already enabled connectome-constrained models of motion detection in the optic lobe, mechanistic models of decision-making circuits, and predictions about the neural basis of complex behaviors such as courtship and walking — predictions that can be tested against experimental recordings from genetically targeted neurons in living flies. The fly threshold is not merely a quantitative milestone; it is the first demonstration that a complete connectome of a behaviorally complex animal can be productively integrated with computational modeling and experimental neuroscience. A companion line of work extends connectomics to the male fruit fly central nervous system — brain plus ventral nerve cord — bringing the total mapped neuron count beyond 166,000 and capturing the complete sensory-to-motor transformation architecture of an entire organism. The ventral nerve cord is the fly equivalent of a spinal cord, containing the motor circuits that drive walking, flying, grooming, and other behaviors; mapping it alongside the brain produces the first complete nervous-system connectome of a behaviorally complex animal, comparable in scope to the C. elegans achievement but at vastly larger scale and physiological complexity. Subsequent work on the larval fruit fly has produced a complete connectome of approximately 3,000 neurons with nearly 550,000 connections, providing a developmental counterpoint to the adult connectome and enabling comparative analyses of how circuit architecture changes through metamorphosis. ### III.3 — The Human Threshold: H01 and Synaptic Resolution Cortex The first human cortical tissue mapped at synaptic resolution was published as the **H01** dataset in May 2024 by the Lichtman laboratory at Harvard Medical School in collaboration with Google Research. The tissue was a fragment of human temporal cortex approximately half the size of a grain of rice — a volume of roughly one cubic millimeter — extracted during surgical treatment of epilepsy and preserved through electron-microscopy-compatible fixation and staining. Serial-section imaging at four-nanometer resolution produced approximately 1.4 petabytes of raw image data. Automated segmentation, followed by extensive proofreading, reconstructed approximately 57,000 cells, 230 millimeters of blood vessels, and roughly 150 million synapses within that cubic-millimeter volume. The full dataset was released as a publicly browsable resource through Neuroglancer, allowing researchers anywhere to navigate human cortex at nanoscale resolution from a web browser. The discoveries within H01 were substantive. Unusual axonal whorls — tightly coiled structures of unknown function — appeared throughout the sample. Rare axon pairs were found in which two cells were connected by as many as fifty synapses, vastly more than the typical one to a few synapses per connected pair. Directionally specialized deep-layer neurons exhibited symmetries that had not been described in the classical neuroanatomical literature. The cortical column structure, the layer-specific cell-type distributions, and the inter-layer connectivity patterns appeared at a level of detail that allowed direct quantitative comparison with the corresponding circuits in mouse and other model organisms. H01 demonstrated, definitively, that high-resolution connectomics of human tissue is technically feasible at the cubic-millimeter scale, and that the resulting reconstructions contain anatomical features that expand rather than merely confirm the textbook ontology of cortical structure. The H01 effort also established the software infrastructure that has since become standard for petascale connectomic data. **Neuroglancer**, the browser-based volumetric visualization tool, allows interactive navigation of multi-terabyte image stacks with on-demand region loading, multi-layer overlays, and real-time annotation. **TensorStore**, Google's open-source library for high-throughput multidimensional array I/O, handles the storage and retrieval of petascale image and segmentation data. **CAVE** (Connectome Annotation Versioning Engine) manages the proofreading workflow, maintains version histories of segmentation edits, and supports distributed collaborative annotation across many human proofreaders. These tools are open-source, and they have been adopted by the FlyWire, MICrONS, and ZAPBench efforts, creating a shared infrastructure layer across the field. ### III.4 — MICrONS: Structure-Function Fusion in Mouse Visual Cortex The bridge from purely structural connectomics to **functional connectomics** was achieved by the MICrONS consortium — Machine Intelligence from Cortical Networks — funded by the Intelligence Advanced Research Projects Activity (IARPA) and executed by a multi-institutional team including the Allen Institute, Baylor College of Medicine, and Princeton University. The MICrONS dataset, with the flagship paper published in *Nature* in April 2025, integrates dense electron-microscopy reconstruction of approximately one cubic millimeter of mouse visual cortex with calcium-imaging physiology recorded from the same tissue volume before extraction. The MICrONS functional connectomics dataset combines dense calcium imaging of around 75,000 neurons in primary visual cortex (VISp) and higher visual areas (VISrl, VISal, and VISlm) in an awake mouse viewing natural and synthetic stimuli, co-registered with an electron microscopy reconstruction containing more than 200,000 cells and 0.5 billion synapses. Proofreading of a subset of neurons yielded reconstructions that include complete dendritic trees as well as the local and inter-areal axonal projections that map up to thousands of cell-to-cell connections per neuron, released as an open-access resource with tools for data retrieval and analysis. The mouse was first subjected to two-photon calcium imaging while viewing parametric and naturalistic visual stimuli at Baylor College of Medicine, recording the activity of every accessible neuron in the targeted cortical volume across multiple imaging sessions. The mouse was then shipped to the Allen Institute, where the imaged tissue volume was extracted, prepared for electron microscopy, sectioned, and imaged over a period of six months of continuous EM acquisition. The EM data were then montaged, roughly aligned, and delivered to Princeton, where fine alignment was performed and the volume was densely segmented using flood-filling networks and other machine-learning approaches. The scientific yield of MICrONS is extensive and ongoing. The MICrONS mouse visual cortex dataset has shown that neurons with similar response properties preferentially connect — a "like-to-like" wiring rule that emerges within and across brain areas and layers, including feedback connections, and independently emerges in artificial neural networks where these connections prove important for task performance. Using a validated digital twin model, researchers separated neuronal tuning into feature (what neurons respond to) and spatial (receptive field location) components, allowing comparative analysis of structural connectivity rules against functional response properties at unprecedented resolution. The dataset has also enabled comprehensive characterization of cell types, synaptic-level connectivity diagrams of cortical columns, and uncovering of organizational principles linking circuit structure to visual computation. Companion papers describe foundation models of neural activity trained on the dataset that predict responses to novel stimulus types, demonstrating that the structure-function integration is sufficiently dense to support generalization beyond the recorded stimuli. The methodological precedent of MICrONS is now being extended. A Virtual Observatory of the Cortex (VORTEX) project funded by NIH allows individual researchers to submit specific scientific requests to be addressed within the MICrONS volume. Proofreading is ongoing with regular public updates. The MICrONS Explorer interface provides interactive exploration and programmatic analysis tools that are increasingly being used as the template for future functional-connectomic datasets. The dataset is, in effect, becoming a research instrument that the broader neuroscience community uses for hypothesis generation and testing, not merely a finished archive. ### III.5 — ZAPBench: Whole-Brain Activity in a Vertebrate The MICrONS dataset achieves structure-function integration at the scale of a cortical volume, but it does not capture an entire brain. The complementary effort that pushes structure-function integration to whole-brain scale, at the cost of working in a much smaller organism, is **ZAPBench** — the Zebrafish Activity Prediction Benchmark — developed by Google Research with HHMI Janelia and Harvard collaborators. The larval zebrafish is a transparent vertebrate whose entire brain can be imaged through the skin using light-sheet fluorescence microscopy, with cellular resolution and brain-wide coverage. The ZAPBench dataset records the activity of approximately 70,000 neurons — roughly the entire brain of the animal — while the fish responds to virtual-reality stimuli including changing light patterns and simulated water currents. Researchers from Harvard, Janelia, and Google Research have collaborated over the past five years to scale combined neural-wiring data collection and neural-activity recording from the same individual organism, building an aligned whole-brain dataset; after hundreds of trial-and-error attempts, they have recorded 70,000 neurons from one larval zebrafish during behavioral tasks and consecutively scanned the same organism's brain tissue using electron microscopy. Neuron reconstruction, specifically the proofreading, remains the most laborious part of this project and is still ongoing. The authors published the Zebrafish Activity Prediction Benchmark in March 2025 and expect the full connectome to be ready in 2026. The combination, when complete, will be the first whole-brain functional connectome of a vertebrate organism: structural wiring of every neuron, co-registered with activity recordings of every neuron, in the same individual animal. ZAPBench is significant not only for the dataset itself but for the benchmark structure it establishes. Predictive models trained to forecast future neural activity from recent activity histories can be evaluated against the held-out portions of the recordings, providing a standardized comparison platform for whole-brain dynamics models. Models that incorporate the structural connectome (when available) can be compared against models that operate purely on activity statistics, allowing direct quantification of how much predictive power the wiring diagram adds. The benchmark thereby creates a testbed for the broader question of how connectomic structure relates to functional dynamics — a question that has been intractable in larger systems but becomes addressable at the larval zebrafish scale. ### III.6 — The Mouse Brain Effort and the Path to Whole-Mammalian Connectomics The current frontier of connectomic ambition is the **whole mouse brain connectome**. The mouse brain represents the smallest mammalian system in which a complete synapse-resolution wiring diagram could plausibly be assembled within the next decade. Google Research launched a dedicated mouse-brain mapping effort in 2023 targeting hippocampal regions involved in memory, attention, and spatial navigation; the Allen Institute and multiple academic partners are running parallel efforts on different mouse brain regions; the proposed Center for Brain-Computer Interfaces (CBCI) program described by the Institute for Progress aims to compress the mouse-brain mapping timeline from the conventional decade-plus estimate to within a single congressional term through industrial-scale parallelization. New microscopes and data processing technology are pushing the cost of full brain wiring diagrams toward tens of millions per brain and compressing the duration of mapping projects to potentially fit within a single congressional term. Sven Dorkenwald, who joined MIT as an assistant professor of brain and cognitive sciences and McGovern Institute investigator in January 2026, has stated that a complete connectome of the mouse brain is achievable within ten to fifteen years but will require substantial collaboration; his research focuses on developing new computational approaches to look for organizational principles within the reconstructed circuitry, asking what circuit motifs allow neurons to contribute to important computations and what the architecture of the circuit is within layers, across layers, and ultimately across regions. The mouse-brain effort is the integration point at which the methodological advances of MICrONS (functional connectomics at cubic-millimeter scale), H01 (human cortical reconstruction), MoGen (synthetic morphology augmentation), LICONN (light-microscopy-based reconstruction), and FlyWire-style distributed proofreading converge into a single coordinated push toward mammalian-scale wiring diagrams. A complementary methodology represented by Connectome-seq — published in *Nature Methods* in April 2026 — uses RNA barcoding rather than electron microscopy to map neural connectivity at single-synapse resolution. Using Connectome-seq, researchers mapped more than 1,000 neurons in a mouse brain circuit known as the pontocerebellar circuit, revealing previously unknown patterns of connectivity including direct links between cell types that had not been known to connect in the adult brain; with ongoing improvements, the team is confident it can eventually reach the goal of mapping the whole mouse brain. Because Connectome-seq is both fast and scalable, it could significantly accelerate research into neurodegenerative diseases, psychiatric conditions, and other brain disorders, allowing comparison of brain connections in healthy individuals with those at different stages of disease to identify early changes in neural circuits. The methodological diversification — electron microscopy, light microscopy, RNA barcoding, all targeting the same biological substrate — is part of the structural maturation of the field. ### III.7 — Reconstruction Infrastructure: PATHFINDER, MoGen, Flood-Filling Networks, LICONN The reconstruction infrastructure that turns raw electron-microscopy data into navigable connectomes has become as scientifically significant as the imaging itself. **Flood-filling networks** (FFNs), developed by Michał Januszewski and colleagues at Google Research, are the dominant machine-learning architecture for automated neuron segmentation. An FFN takes a small three-dimensional region of an EM volume and iteratively expands a binary mask outward from a seed point, predicting at each step which neighboring voxels belong to the same neuron as the seed. The architecture allows segmentation of individual neurons even in densely packed tissue where conventional semantic segmentation approaches fail. FFNs have been the segmentation backbone of FlyWire, H01, MICrONS, and the ongoing mouse-brain efforts; their accuracy and throughput have improved by approximately an order of magnitude over the past five years. **PATHFINDER** is Google Research's current production reconstruction system, integrating FFN-based segmentation with downstream processing stages including shape plausibility classification, agglomeration of segments into complete neurons, and error detection. PATHFINDER is the system into which MoGen's synthetic neuronal morphologies are injected as training data. MoGen augments the training set of a shape plausibility classifier from a production connectomics neuron reconstruction pipeline with millions of generated samples, thereby improving classifier accuracy and reducing the number of remaining split and merge errors by 4.4 percent — an improvement that at full-mouse-brain scale can reduce manual proofreading labor by over 157 person-years. Adding MoGen-generated synthetic examples to PATHFINDER's training delivers a 4.4 percent reduction in reconstruction errors and suggests directions for further improvements; Google has released MoGen as an open-source model together with species-specific trained models as a resource for the broader community. Google is also exploring using simulated neurons to create synthetic electron microscope images that provide additional training data earlier in the reconstruction pipeline. The **LICONN** approach — Light-microscopy-based Connectomics — developed by the Institute of Science and Technology Austria in collaboration with Google Research, and published in *Nature* in May 2025, represents a fundamental diversification of the imaging substrate. LICONN integrates specifically engineered hydrogel embedding and expansion with comprehensive deep-learning-based segmentation and analysis of connectivity, thereby directly incorporating molecular information into synapse-level reconstructions of brain tissue. LICONN is the first technology beyond electron microscopy capable of reconstructing brain tissue with all the synaptic connections between neurons, while also opening up the possibility of visualizing complex molecular machinery alongside the structure of neurons, all while using standard light microscopes for measurements. The technique reconstructs approximately one million cubic micrometers of mouse primary somatosensory cortex at original tissue scale through approximately 16-fold hydrogel expansion, enabling synapse-level reconstruction with light microscopy combined with molecular labeling such as connexin-43 and gephyrin immunostaining and deep-learning predictions of pre-synapses (bassoon) and excitatory post-synapses (SHANK2). The 20-nanometer effective resolution achieved through hydrogel expansion separates fine features that would otherwise be unresolvable, allowing nanoscale imaging with conventional light microscopes. The LICONN breakthrough is fundamentally about accessibility. Electron microscopy requires specialized facilities, multi-million-dollar instruments, and operational expertise that limits connectomics to a small number of laboratories worldwide. Light microscopy is available in essentially every biological research institution. By making synapse-level reconstruction tractable with light-microscopic instruments, LICONN distributes the connectomic methodology across the broader research community, enables molecular co-registration that electron microscopy cannot easily provide, and accelerates the rate at which connectomic data can be produced and shared. The multimodal nature of LICONN — combining structural reconstruction with molecular phenotyping in the same tissue volume — also makes it methodologically aligned with the broader push toward multimodal connectomes that integrate wiring, cell type, neurotransmitter identity, and gene expression into unified datasets. The infrastructure layer also includes **Neuroglancer** for browser-based volumetric visualization, **TensorStore** for petascale array I/O, **CAVE** for proofreading versioning and collaborative annotation, **SegCLR** for self-supervised learning of cellular representations from EM data, and **Codex** as the FlyWire data exploration interface. These are open-source tools, increasingly standardized across the field, that together constitute the software substrate of contemporary connectomics. The trajectory is toward integrated platforms in which raw imaging data flows through automated segmentation, machine-learning-assisted proofreading, and interactive visualization with full version control and provenance tracking — the connectomic equivalent of the software stack that genomics developed for sequencing data over the past two decades. ### III.8 — From Wiring Diagrams to Executable Models The decisive demonstration that connectomes can function as executable hypothesis engines came from connectome-constrained deep mechanistic modeling of the *Drosophila* visual system. The team used experimentally determined connectivity for 64 cell types in the motion-detection pathways of the fly optic lobe, fixed the wiring structure as a hard biological constraint, and then optimized the remaining free parameters — single-neuron biophysical properties and synaptic weights — using gradient-based deep-learning methods such that the circuit performed motion detection on natural visual input. The resulting model preserved the biological topology of the optic lobe while learning the parameter values needed to make that topology computationally productive. It then predicted neural activity patterns across all 64 cell types and successfully reproduced more than two dozen experimental findings spanning two decades of *Drosophila* visual neuroscience — direction selectivity properties, contrast adaptation curves, motion-opponent suppression, and time-constant tuning across cell classes — none of which had been used during model training. The methodological significance is foundational. A sufficiently dense connectome no longer functions purely as an anatomical archive; it functions as an **inductive bias** for machine-learning models. The wiring diagram supplies the architectural skeleton, the computational task supplies the optimization objective, and gradient descent supplies the parameter values that complete the executable simulation. The biological constraint dramatically narrows the parameter space relative to unconstrained neural networks, accelerates convergence, and produces models whose behavior aligns with experimental observation by construction rather than by fitting. The connectome becomes a hypothesis engine: any proposed mechanism for how a particular circuit performs a particular computation can be tested by running the corresponding constrained model and comparing its predictions to recordings from the living circuit. Disagreement between model and recording localizes the explanatory gap; agreement supports the proposed mechanism without proving it. The methodology has been extended to additional fly circuits — the central complex for spatial navigation, mushroom body circuits for associative learning, descending neuron pathways for motor control — and is being prepared for application to MICrONS-scale mouse cortical circuits where the structural-functional dataset density now permits it. This is the moment at which connectomics structurally couples to artificial neural network research. The same gradient-based optimization machinery that trains foundation models is being applied to connectome-constrained biological circuit models, with biological structure replacing the random initialization and learned task objective replacing the language modeling loss. The intellectual feedback loop runs in both directions: artificial neural network advances in optimization, regularization, and scaling are imported into connectome-constrained modeling; biological structural motifs discovered in the connectome are exported into artificial architecture design. The shared methodological substrate is the gradient and the graph, and both substrates — biological and synthetic — are now expressible within it. --- ## Movement IV — The Synthetic Substrate: Neuromorphic Hardware and Engineered Neural Computation ### IV.1 — Spiking Neural Networks as Computational Frame The synthetic substrate of brain-grounded computation rests on the **spiking neural network** (SNN) as its dominant algorithmic frame. Conventional artificial neural networks operate on continuous-valued activations, updated synchronously in dense matrix multiplications across all units in each layer. Biological neurons communicate through discrete electrical spikes at precise temporal moments, with information encoded jointly in the identity of the firing neuron, the timing of the spike, the rate of spiking over time windows, and the patterns of correlated activity across neuronal populations. Spiking neural networks attempt to preserve this event-driven, temporally precise, sparse activity regime in computational form. The fundamental computational unit is typically a **leaky integrate-and-fire** (LIF) neuron, whose membrane potential accumulates input from incoming spikes, decays exponentially toward a resting value in the absence of input, and emits its own spike when the membrane potential crosses a threshold — after which the potential resets and a brief refractory period prevents immediate re-firing. More elaborate models, including the Izhikevich neuron, the Hodgkin-Huxley conductance-based model, and adaptive exponential integrate-and-fire variants, add additional biophysical fidelity at the cost of computational expense. The computational advantages of SNNs are situational rather than universal. Their sparse, event-driven communication structure allows hardware implementations to perform computation only when spikes occur, dramatically reducing the energy cost of inference relative to dense matrix multiplication on conventional processors. Their temporal precision allows information encoding in spike timing, supporting forms of computation — coincidence detection, temporal pattern recognition, delay-tuned filtering — that dense feedforward networks accomplish only with substantial architectural overhead. Their recurrent dynamics and adaptive thresholds support short-term memory, sequential processing, and signal denoising in ways that align naturally with sensory and motor processing tasks. The disadvantages are also real: training SNNs with gradient-based methods is complicated by the discontinuous nature of the spike, requiring surrogate gradient approaches or conversion from pre-trained continuous networks; the accuracy gap between SNNs and dense neural networks remains nontrivial for many benchmarks; and the software tooling for SNN development, while improving, remains less mature than the conventional deep learning ecosystem. The field has converged on a hardware-software stack in which **neuromorphic chips** provide specialized silicon for efficient SNN execution, **software simulators** support algorithm development and model exploration on conventional hardware, and **deployment frameworks** translate between the two. The remainder of this movement traverses the major neuromorphic hardware platforms, then the software simulator ecosystem, and finally the emerging device-level innovations — memristive, photonic, and analog — that constitute the longer-horizon trajectory of the field. ### IV.2 — Intel Loihi and Loihi 2: Programmable Digital Spiking Silicon The most widely deployed contemporary neuromorphic research platform is Intel's **Loihi** family, developed by Mike Davies and colleagues at Intel Labs and now in its second generation. **Loihi 2**, fabricated in Intel's 7-nanometer process (Intel 4 node) and presented publicly in 2021 with continuing extensions through 2026, integrates 128 fully asynchronous neuromorphic cores per chip with six embedded x86 Lakemont microprocessor cores for general-purpose orchestration, connected through an asynchronous network-on-chip mesh. Each neuromorphic core supports thousands of neuron compartments, local SRAM for membrane states and synaptic weights, programmable microcode allowing custom neuron model implementation, and high-throughput synaptic routing fabric. A single Loihi 2 chip supports up to one million spiking neurons and 120 million synapses, with multi-chip configurations such as the eight-chip Kapoho Point board enabling scaling to AI models with up to one billion parameters or optimization problems with up to eight million variables. The architecture's energy efficiency is substantial: published benchmarks show greater than 100-fold efficiency over CPU implementations and approximately 30-fold over GPU implementations for SNN workloads under event-driven regimes, with sub-millisecond inference latency and sub-millijoule energy per inference for typical edge applications. The decisive software-side advance accompanying Loihi 2 is the **Lava** open-source framework, which provides Python-level APIs for describing spiking neural networks and compiling them to either CPU or Loihi 2 backends. Lava supports custom neuron model definition through fixed-point microcode for the neuromorphic cores and custom C code for the embedded x86 cores, allowing researchers to implement non-standard neuron dynamics including resonate-and-fire, Hopf oscillators, and sigma-delta variants alongside the standard LIF baseline. The framework is the substrate of the **Intel Neuromorphic Research Community** (INRC), a global consortium that includes academic groups, government laboratories, and industrial research organizations exploring applications spanning autonomous robotics, sensor processing, optimization, and biomedical signal analysis. Sandia National Laboratories, Los Alamos National Laboratory, Lawrence Livermore National Laboratory, the Air Force Research Laboratory, and university partners including ETH Zürich, Heidelberg, and the Italian Institute of Technology run active Loihi 2 deployments, with research applications ranging from quantum-neuromorphic interface investigation to neuromorphic graph convolution to event-based sensory processing for autonomous vehicles. The architecture's limitations include fixed-point quantization requirements, on-core memory constraints, hardware-limited maximum synaptic delay of 62 timesteps, and nontrivial scaling overhead as models cross chip boundaries — but within those constraints, Loihi 2 represents the most mature programmable neuromorphic platform currently available to the research community. ### IV.3 — IBM TrueNorth and NorthPole: From Spiking to Inference Acceleration IBM's neuromorphic trajectory runs through two architectures that mark a deliberate strategic evolution. **TrueNorth**, unveiled in 2014 by Dharmendra Modha and colleagues at IBM Research as part of the DARPA SyNAPSE program, integrated 4,096 neurosynaptic cores per chip containing one million spiking neurons and 256 million synapses across 5.4 billion transistors. Each core implemented 256 programmable neurons with 256 programmable synapses per neuron, fully asynchronous event-driven communication via on-chip routing, and energy efficiency approximately four orders of magnitude better than conventional microprocessors of its era. TrueNorth demonstrated, at scale, that purely digital neuromorphic silicon could perform real-time pattern recognition at energy budgets compatible with embedded and edge deployment, and it served as the conceptual proof-of-concept for the entire generation of neuromorphic chips that followed. The successor architecture, **NorthPole**, presented at Hot Chips 2023 and detailed in *Science* in October 2023, represents a deliberate departure from pure spiking neural network execution toward optimized neural network inference acceleration. NorthPole contains 256 cores per chip, with each core functioning as a vector-matrix multiplication engine capable of 2,048 operations per cycle at 8-bit precision or 8,192 operations per cycle at 2-bit precision, organized around 22 billion transistors on approximately 800 square millimeters of silicon manufactured at 12 nanometers by AMD. The architectural innovation is the tight intertwining of compute and memory: 192 megabytes of on-chip memory are distributed across the compute elements rather than centralized in external DRAM, with an additional 32 megabytes serving as I/O framebuffer. Two independent network-on-chip fabrics handle weight and program reconfiguration during execution, effectively multiplying the apparent on-core memory size by up to 256-fold and allowing entire neural networks to remain resident on-chip during inference. The published benchmarks against Nvidia's V100 GPU on the ResNet-50 image-classification model show NorthPole performing 25 times more energy-efficient operation per watt and 22 times faster inference while occupying one-fifth the silicon area — and NorthPole achieves 3,000 times more computation and 640 times larger network models than TrueNorth despite having only four times more transistors. The trajectory from TrueNorth to NorthPole signals IBM's strategic recognition that pure spiking neural networks face market headwinds against the dominant deep learning paradigm, and that brain-inspired architectural principles can be deployed in service of standard neural network workloads with substantial efficiency advantages. ### IV.4 — SpiNNaker 2: Massively Parallel Brain Simulation The **SpiNNaker** architecture, developed at the University of Manchester under Steve Furber — the original designer of the ARM microprocessor — represents a distinctively different approach to neuromorphic computing organized around massively parallel general-purpose processors rather than specialized neuron circuits. The first-generation SpiNNaker system, completed as part of the European Human Brain Project, used custom-designed many-core processors with eighteen ARM968 cores per chip and was deployed at scales up to one million cores capable of simulating spiking neural networks of approximately one percent the scale of the human brain in biological real-time. The architectural insight was that biological neural simulation is bottlenecked not by raw compute but by inter-neuron communication; SpiNNaker's custom packet-switched network-on-chip was optimized for the kind of one-to-many small-message routing that biological networks generate. **SpiNNaker 2**, developed by Christian Mayr's group at TU Dresden in collaboration with Manchester and commercialized by the Dresden spinoff company **SpiNNcloud**, scales the architecture substantially. Each SpiNNaker 2 chip houses 152 ARM cores with 19 megabytes of on-chip SRAM, 2 gigabytes of DRAM, and dedicated machine learning accelerators (including multiply-accumulate units and specialized exponential/logarithmic function units) manufactured in 22-nanometer fully-depleted silicon-on-insulator technology with adaptive body biasing. Each SpiNNaker 2 server board hosts 48 chips interconnected in toroidal topology; full system deployments combine multiple boards into massively parallel neuromorphic supercomputers. Sandia National Laboratories deployed a 24-board SpiNNaker 2 system in 2025 — 1,152 chips total, approximately 175,000 ARM cores — to investigate energy-efficient AI architectures for national security applications, with reported energy efficiency 18 times greater than conventional GPU systems for spiking neural network workloads. SpiNNcloud's architectural positioning explicitly hybridizes traditional spiking neuromorphic computing with mainstream deep learning execution, allowing event-driven characteristics to be leveraged in conventional deep neural networks as well as in biologically motivated spiking models. Application domains include automotive AI, tactile internet, industry 4.0 deployments, and biomedical signal processing. ### IV.5 — BrainScaleS 2 and the Analog Neuromorphic Approach The **BrainScaleS** platform, developed by Karlheinz Meier (until his death in 2018) and now continued by Johannes Schemmel and colleagues at Heidelberg University, represents the analog mixed-signal branch of the neuromorphic field. Where Loihi and SpiNNaker implement neural dynamics in digital silicon, BrainScaleS exploits the native physics of CMOS capacitances and conductances to emulate neuronal membrane equations directly in continuous-time analog circuits — running, by design, approximately one thousand times faster than biological real-time. **BrainScaleS 2** integrates 512 adaptive integrate-and-fire neurons and 131,072 plastic synapses per chip, augmented with embedded digital processors with SIMD extensions, on-chip spike routing, and event-driven communication infrastructure. The full BrainScaleS 2 multi-chip system, deployed at the European Institute for Neuromorphic Computing in Heidelberg, consists of two backplanes with twelve interconnected chips each, totaling approximately 12,000 neurons and 3 million synaptic circuits in continuous-time analog emulation. The thousand-fold acceleration relative to biological time is the platform's defining feature. Experiments that would require hours of biological evolution complete in seconds of BrainScaleS 2 execution, allowing exploration of synaptic plasticity rules, network learning dynamics, and developmental processes at rates incompatible with both biological recording and digital simulation. The architecture supports training of deep spiking and non-spiking neural networks using hybrid techniques including surrogate gradients, structured multi-compartmental cell models for dendritic computation research, and configurable plasticity rules ranging from spike-timing-dependent plasticity to biologically realistic three-factor learning rules incorporating neuromodulatory signals. The analog substrate introduces device variability — no two neuron circuits on the chip are identical due to manufacturing variation — which has been reframed methodologically as a feature rather than a defect, since biological neurons also exhibit substantial heterogeneity and biologically realistic computation must operate robustly across that heterogeneity. BrainScaleS 2 has been deployed for accelerated robotics applications, hippocampal memory modeling, and developmental learning research, and serves as one of the canonical platforms for the European Union's continued neuromorphic computing research program through the EBRAINS 2.0 infrastructure. ### IV.6 — Emerging and International Architectures: Tianjic, Darwin, Innatera, Akida, GrAI Matter The neuromorphic hardware landscape includes substantial international and commercial development beyond the major academic platforms. **Tianjic**, developed at Tsinghua University by the group of Luping Shi, represents a hybrid architecture supporting both spiking neural networks and conventional deep learning networks on the same chip — a design philosophy aligned with SpiNNaker 2's hybrid positioning. Tianjic was demonstrated controlling an autonomous bicycle in 2019, executing visual processing, voice recognition, balance control, and obstacle avoidance simultaneously on a single chip. **Darwin** and its successor **Darwin 3**, developed at Zhejiang University and Alibaba (announced 2023), represent China's parallel neuromorphic effort, with Darwin 3 incorporating contemporary architectural improvements while pursuing the same general spiking neural network execution strategy as Loihi. The commercial neuromorphic sector has emerged through several specialized startups. **Innatera**, a Dutch spinoff from TU Delft, develops mixed-signal analog neuromorphic chips targeting ultra-low-power edge applications in sensor processing, with chip designs explicitly oriented toward microwatt-scale audio and radar signal analysis. **BrainChip**, an Australian-American company, manufactures the **Akida** family of digital neuromorphic processors aimed at edge AI deployment, with Akida 2 supporting on-device learning, event-based vision processing, and integration with conventional embedded computing platforms. **GrAI Matter Labs**, a French-American company, develops the GrAI Core architecture targeting always-on edge sensor processing. **Prophesee**, a French company, develops event-based vision sensors that pair naturally with neuromorphic processors — the dynamic vision sensor (DVS) generates pixel-level spike events only when scene luminance changes, producing sparse data streams that neuromorphic chips can process at low latency and power. The DVS-neuromorphic combination is the canonical event-driven sensing-and-processing pipeline, deployed in autonomous vehicle research, industrial inspection, and high-speed robotics. ### IV.7 — Software Simulators and the Algorithm Development Stack The software simulator ecosystem underlies essentially all neuromorphic algorithm development, since hardware platforms remain limited in availability and slow to iterate on. **NEST** (Neural Simulation Tool), developed by the NEST Initiative led by the Jülich Research Centre, is the dominant large-scale spiking neural network simulator for biologically detailed brain modeling, supporting networks of millions of neurons distributed across HPC clusters. **Brian**, developed by Romain Brette and Dan Goodman, provides a Python-based interface oriented toward computational neuroscience research, allowing flexible neuron model definition through differential equation specification. **GeNN** (GPU-enhanced Neural Networks), developed at the University of Sussex by Thomas Nowotny and James Knight, generates optimized CUDA code for GPU-accelerated spiking neural network simulation and serves as a bridge between research-grade flexibility and high-throughput execution. **Nengo**, developed by Chris Eliasmith's group at the University of Waterloo, implements the Neural Engineering Framework that systematically translates between mathematical functions and biologically realistic neural network implementations, and is the substrate for the Spaun cognitive architecture and Applied Brain Research's commercial offerings. The convergence with conventional deep learning is being pursued through **snnTorch**, **Norse**, **SpikingJelly**, and similar frameworks that integrate spiking neural network primitives into PyTorch and TensorFlow workflows, allowing researchers familiar with the deep learning ecosystem to develop SNNs using surrogate gradients, ANN-to-SNN conversion methods, and hybrid spiking-conventional architectures. The Intel **Lava** framework, the SpiNNcloud **Py-SpiNNaker** stack, and the BrainScaleS **PyNN** integration provide hardware-specific deployment paths from the same general algorithmic substrate. The trajectory is toward unified toolchains in which spiking neural network development is no longer a separate discipline from conventional deep learning but a complementary capability accessible through the same software interfaces. ### IV.8 — Memristive and Diffusive Memristor Devices The longer-horizon trajectory of neuromorphic computing runs through **emerging device technologies** that promise to replace conventional silicon transistors with components whose physical behavior more closely mirrors that of biological synapses and neurons. **Memristors** — two-terminal devices whose resistance depends on the history of current flow through them — implement the synaptic weight as a physical parameter that updates continuously in response to stimulation rather than being stored in separate digital memory. Memristor crossbar arrays perform matrix-vector multiplication as an analog physical operation: applying a vector of voltages to the rows of the crossbar produces, through Ohm's law and Kirchhoff's current law, a vector of currents at the columns proportional to the matrix-vector product, with the matrix elements stored as physical conductances. The energy cost of the multiplication is approximately three orders of magnitude lower than equivalent digital operation, and the operation completes in the time it takes for the analog currents to settle — nanoseconds rather than microseconds or milliseconds for digital execution. **Diffusive memristors**, developed in particular by the group of J. Joshua Yang (now at the University of Southern California after appointments at HP Labs, the University of Massachusetts Amherst, and Hewlett Packard), exploit silver-ion diffusion dynamics in oxide thin films to produce devices that exhibit not only memristive behavior but also the **stochastic, threshold-dependent, integrate-and-fire dynamics** characteristic of biological neurons. A diffusive memristor accumulates silver-ion clusters in response to applied voltage, reaching a conductive threshold at which the device transitions abruptly to a low-resistance state — emitting a "spike" — before thermal diffusion gradually disperses the cluster and returns the device to its high-resistance baseline. This physical behavior directly implements integrate-and-fire neuron dynamics without requiring the complex digital logic that LIF circuits in conventional CMOS demand. The combination of memristive synapses with diffusive memristor neurons constitutes a fully analog physical neural network whose operation more closely resembles biological computation than digital simulation of biological computation. Crossbar-scale demonstrations of this approach have shown image-classification performance approaching that of conventional neural networks at one to three orders of magnitude lower energy. The memristive trajectory remains in research deployment rather than commercial production — manufacturing yields, device-to-device variability, retention times, and endurance under repeated update cycles all require continued engineering — but companies including HP, Hewlett Packard Enterprise, Mythic AI, Knowm, Crossbar, and various university spinoffs are pursuing commercial implementations. The phase-change memory variant, in which the memristive element is a chalcogenide material that switches between amorphous and crystalline states, has been demonstrated at IBM and other industrial laboratories as a more manufacturable alternative to oxide-based memristors. Photonic crossbars using integrated silicon photonics for matrix-vector multiplication, pursued by Lightmatter, Lightelligence, Salience Labs, and others, offer a third path with different trade-offs — light replaces current as the analog signal carrier, allowing operations at the speed of light with electromagnetic rather than thermal energy budgets. ### IV.9 — Photonic Neuromorphic Computing The photonic substrate represents the most architecturally distant departure from biological neural computation while simultaneously offering some of its strongest practical advantages for neural-network-style workloads. **Photonic neuromorphic systems** encode information in optical signals — typically intensity-modulated coherent or incoherent light propagating through silicon photonic waveguides — and perform computation through linear optical operations (interference, splitting, combination) and nonlinear optoelectronic conversion. Matrix-vector multiplication, the dominant operation in neural network inference, can be implemented entirely in the optical domain using meshes of programmable Mach-Zehnder interferometers or microring resonators that distribute and combine optical signals according to programmed weight matrices. Once the weights are configured, the operation completes at the speed of light through the photonic chip, with energy costs determined primarily by the electrical-to-optical conversion at the input and the optical-to-electrical conversion at the output. **Lightmatter**, founded by Nick Harris and colleagues out of MIT, develops the **Passage** family of 3D photonic interposer chips that scale optical interconnect bandwidth far beyond what conventional electrical interconnect can achieve. The Passage M1000, announced in March 2025, integrates approximately 4,000 square millimeters of photonic interposer that provides 114 terabits per second of total optical bandwidth and connects thousands of GPUs in a single integrated package, addressing the interconnect bottleneck that increasingly limits large-scale AI training and inference. The Passage L200 is a co-packaged optics platform supporting 32 to 64 terabits per second of bandwidth for direct integration with next-generation XPUs and switches. Lightmatter's architectural positioning is no longer purely neuromorphic in the SNN sense; it is photonic AI infrastructure designed to scale conventional deep learning beyond electrical interconnect limits while retaining the energy advantages of optical signal propagation. **Lightelligence**, a competing photonic computing company spun out of MIT, pursues a more directly neuromorphic positioning with photonic neural network processors for inference acceleration. Academic photonic neuromorphic work has demonstrated gigahertz-scale spiking neural network execution on integrated silicon photonic chips with in-situ learning capabilities, achieving over 80 percent accuracy on standard video recognition benchmarks while operating approximately 100 times faster than conventional frame-based approaches. The convergence of photonic neuromorphic computing with conventional deep learning infrastructure is among the most significant architectural transitions in the synthetic substrate. The energy and bandwidth advantages of optical computation align directly with the dominant cost structures of contemporary AI training and inference, making photonic acceleration commercially attractive for reasons that have nothing to do with biological plausibility. The neuromorphic origins of the architectural strategy nevertheless persist in the design philosophy: massively parallel matrix operations, sparse activity patterns, event-driven processing, and tight integration of compute and memory remain the operating principles, regardless of whether they are implemented in CMOS spiking circuits, memristive crossbars, or photonic waveguide meshes. ### IV.10 — Synthetic Biology and Engineered Artificial Neurons A fourth trajectory in the synthetic substrate runs through **synthetic biology** rather than electronic engineering — constructing artificial neurons and small neural circuits from biological molecular components rather than from silicon or photonics. Vesicle-based artificial neurons, developed by various academic groups, encapsulate ion channels and synthetic signaling pathways within lipid membranes to create cell-like compartments whose membrane potential, ion flux, and intercellular signaling can be configured through engineered molecular composition. The approach allows engineering control over biological computation primitives — ion channel kinetics, synaptic transmission, neuromodulator response — that is difficult to achieve in living cells, while preserving the molecular-scale operation and energy parsimony of biological systems. Research groups at Oxford, Imperial College London, Harvard's Wyss Institute, and several European synthetic biology consortia are pursuing this direction; current state-of-the-art demonstrations involve small networks of synthetic vesicles coupled through engineered membrane proteins, performing basic logic operations or signal-processing tasks. The synthetic biological direction sits methodologically between the organic substrate (living tissue) and the engineered substrate (silicon and photonics): it constructs cell-like computational units from biological molecular components but with engineering-grade control over composition and connectivity. Its near-term applications are likely to be in biosensors, in vivo therapeutic devices, and computational substrates for environments where conventional electronics cannot operate (inside living organisms, in extreme chemical environments, at biological-molecular interfaces). The longer-term implication is that the substrate distinction between organic and synthetic may not survive the maturation of the field — engineered biological computation and biologically inspired engineered computation are converging toward a regime where the distinction becomes operationally less meaningful. --- ## Movement V — Functional Integration and Readout: Activity, Imaging, and Dynamic Coupling ### V.1 — The Functional Imaging Stack The mapping between neural structure and neural computation requires methods to record what nervous tissue is actually doing — not merely how it is wired, but how its activity unfolds in time under task-relevant conditions. The functional imaging stack spans many orders of magnitude in spatial and temporal resolution, with different modalities offering complementary trade-offs that determine which questions can be addressed and which substrates can be coupled. **Two-photon calcium imaging** is the dominant cellular-resolution technique for monitoring large populations of neurons in living tissue. Genetically encoded calcium indicators — most commonly the GCaMP family of fluorescent proteins, with GCaMP6, GCaMP7, and the current-generation GCaMP8 variants — change their fluorescence intensity in response to intracellular calcium concentration, which rises during action potential firing. Two-photon microscopy uses pulsed infrared laser excitation to confine fluorescence to a thin focal plane within scattering tissue, allowing optical sectioning at depths up to roughly one millimeter in mammalian cortex. The technique has been deployed at increasing scale through hardware advances (resonant scanners, mesoscopes that image larger fields of view, multi-area imaging systems) and algorithmic advances (motion correction, source extraction, deconvolution) that now permit simultaneous recording from tens of thousands of neurons in awake behaving animals. The MICrONS dataset's 75,000-neuron functional recording was performed using two-photon calcium imaging across multiple weeks of imaging sessions before the tissue was extracted for electron microscopy. **Light-sheet fluorescence microscopy**, also called selective plane illumination microscopy, achieves higher imaging throughput by illuminating an entire tissue plane simultaneously with a sheet of laser light orthogonal to the imaging axis, allowing rapid volumetric scanning at the cost of greater scattering sensitivity than two-photon. The technique is the workhorse of larval zebrafish whole-brain imaging — the entire brain of the transparent larva can be scanned at single-cell resolution at rates of a few volumes per second, supporting the kind of whole-brain activity recording that underlies ZAPBench. Recent technological advances including swept-confocal-aligned planar excitation (SCAPE) and lattice light-sheet microscopy extend the technique to opaque mammalian preparations and small organoids, though depth remains limited compared to two-photon approaches. **Voltage imaging** records membrane potential changes directly rather than inferring them from downstream calcium signals, offering millisecond-scale temporal resolution that captures individual action potentials with their precise timing. The technique has historically been limited by signal-to-noise ratios — voltage indicators must respond rapidly to small membrane potential changes against bright fluorescence baselines — but recent generations of indicators including the ASAP, Voltron, and JEDI families have achieved sufficient sensitivity for recording from dozens to hundreds of neurons simultaneously. The shift from calcium to voltage imaging is one of the most significant methodological transitions currently underway in cellular-resolution functional imaging, because it removes the temporal filtering that calcium indicators impose and exposes the spike-timing information that biological computation actually uses. **Functional ultrasound** (fUS) imaging measures cerebral blood volume changes through ultrasonic detection of moving red blood cells, providing whole-brain coverage at sub-millimeter spatial resolution and approximately 100-millisecond temporal resolution. The modality bridges the gap between optical methods (cellular resolution but limited depth and field of view) and conventional fMRI (whole-brain coverage but lower resolution and the requirement of a large scanner). fUS has been demonstrated in awake rodents, non-human primates, and most recently in human intraoperative settings; the field's medium-term ambition is to develop portable fUS imaging systems for clinical and research applications, providing whole-brain hemodynamic monitoring with substantially better resolution than fNIRS at substantially lower cost than fMRI. **Electrophysiology** — direct recording of electrical activity from neurons through microelectrodes inserted into tissue — remains the gold standard for high-temporal-resolution measurement at single-neuron and single-spike precision. The contemporary state of the art is represented by the **Neuropixels** probe, developed by IMEC in collaboration with Janelia, the Allen Institute, the Wellcome Trust, and HHMI: a silicon shank approximately 70 micrometers wide and 10 millimeters long containing 960 to nearly 5,000 recording sites that can be selectively activated, allowing simultaneous recording from hundreds of neurons distributed along the probe axis. The current generation, Neuropixels 2.0, supports recording from approximately 384 channels per shank with multi-shank configurations enabling recordings from over 1,000 neurons simultaneously across multiple brain regions. Neuropixels has become the dominant electrophysiology platform for systems neuroscience, deployed across hundreds of laboratories worldwide and serving as the recording substrate for major collaborative efforts including the International Brain Laboratory. ### V.2 — Optogenetics, Chemogenetics, and Causal Manipulation Recording activity establishes correlation between neural patterns and behavior; **causal manipulation** establishes which patterns are functionally responsible for which behaviors. The dominant contemporary causal manipulation tools are **optogenetics** and **chemogenetics**, both of which use genetic engineering to express light-sensitive or drug-sensitive proteins in specifically targeted neuronal populations, then activate or silence those populations on demand to test their causal contribution to ongoing computation. Optogenetics, developed primarily by Karl Deisseroth, Ed Boyden, and collaborators in the mid-2000s, uses channelrhodopsins (light-activated cation channels), halorhodopsins (light-activated chloride pumps), and increasingly sophisticated derivatives that allow millisecond-precision excitation or inhibition of genetically targeted neurons. Chemogenetics uses engineered designer receptors (DREADDs — Designer Receptors Exclusively Activated by Designer Drugs) that respond to otherwise inert small molecules, allowing slower but more sustained activation or inhibition of targeted populations across hours or days. The combination of recording and manipulation is methodologically decisive for the structure-function integration agenda. A connectome predicts what circuits exist and how they might compute; activity recordings show what those circuits are doing in real time; optogenetic and chemogenetic manipulation tests which patterns of activity are causally necessary or sufficient for which behaviors. Together, the three methodological strands close the loop from anatomy through dynamics to causal mechanism, and modern systems neuroscience experiments routinely combine all three. The MICrONS-style functional connectomic datasets implicitly invite extensions in which the same tissue volume that is structurally mapped is also subject to optogenetic perturbation during the activity recording, producing causally informed structural-functional datasets. Such experiments are technically feasible at present and are being assembled by several laboratories. ### V.3 — All-Optical Interrogation and Holographic Stimulation The integration of optical recording and optical stimulation produces **all-optical interrogation** — the ability to simultaneously read out activity from large neural populations and write activity into specific cells, all through the same optical system without electrical penetration of the tissue. The technical apparatus combines two-photon calcium or voltage imaging for recording with two-photon optogenetic stimulation through computer-generated holography for writing: a spatial light modulator shapes the laser excitation pattern to address specific cells in three dimensions, allowing arbitrary spatial patterns of activity to be imposed on the network while its response is simultaneously monitored. The technique has been demonstrated at scales of dozens to hundreds of simultaneously addressed cells, with the trajectory toward thousands of cells in active development. All-optical interrogation is the closest in-vivo analog to the closed-loop architectures of in vitro organic computing platforms. Like DishBrain or Brainoware, all-optical interrogation embeds living neural tissue in a structured input-output loop, with the experimenter (or an automated control system) playing the role of the simulated environment. Unlike DishBrain, the tissue is intact within a living animal, the inputs and outputs are addressed at cellular resolution through holographic optics rather than electrical stimulation through multielectrode arrays, and the behavioral context can be the animal's natural sensorimotor engagement with the world rather than a virtual game environment. The methodological convergence between in vitro and in vivo closed-loop approaches is one of the structural signals of the field's maturation: the same input-output-feedback architecture is being deployed across substrate types, allowing direct comparison of computational principles between cultured tissue, intact brain circuits, and synthetic neural networks. --- ## Movement VI — Semantic Translation: Neural Activity in Language-Model Space ### VI.1 — The fMRI Semantic Decoder The decisive demonstration that neural activity can be translated into the representational space of large language models came from the work of Alex Huth and Jerry Tang at the University of Texas at Austin, published in *Nature Neuroscience* in May 2023. The system records blood-oxygen-level-dependent (BOLD) signals through functional magnetic resonance imaging while subjects listen to spoken stories, watch silent video content, or imagine telling stories silently — and then reconstructs continuous language sequences whose semantic content matches the participants' brain states. Training requires approximately sixteen hours of fMRI recording per subject during story listening to fit subject-specific encoding models that map between cortical activity patterns and the semantic embedding space derived from a language model. Once trained, the decoder produces continuous text whose meaning aligns with what the participant was hearing, watching, or imagining — not as a word-for-word transcript but as a semantic reconstruction of the underlying content. The methodological architecture is consequential. The decoder operates in the **semantic embedding space** of a pretrained language model — initially GPT-2 in the original 2023 work, with subsequent extensions to larger and more recent foundation models — and maps cortical activity patterns to points in that embedding space, then uses the language model to generate text consistent with those embeddings. The brain is not being read as a sequence of words; it is being read as a sequence of semantic states, and the language model is performing the inverse mapping from semantic state to natural language. The decoder's reconstructions preserve meaning while frequently substituting different specific words, paraphrasing, or rendering implicit content explicitly — all behaviors consistent with semantic-level rather than lexical-level decoding. The system's accuracy degrades smoothly with reduced training data, generalizes (with reduced fidelity) across conditions including silent video viewing and imagined speech, and fails predictably when subjects actively try to think about something unrelated to the stimulus — confirming that the decoder is reading goal-directed semantic content rather than sensory artifact. The cognitive scope of the decoder is notable. The semantic content it can reconstruct includes not only the literal content of stories being listened to but also imagined narratives, semantic interpretations of silent visual content, and the gist of complex multi-modal experiences — suggesting that the cortical semantic representations being decoded are themselves modality-general and abstract, encoding what the participant is thinking about rather than what specific sensory inputs are being processed. This is consistent with the broader neuroscience view of cortical semantic representation as a high-dimensional, distributed code in which meaning is encoded across millions of voxels in a way that aligns approximately with the high-dimensional embedding spaces that modern language models learn. ### VI.2 — BrainLLM and Direct Generation From Neural Activity The 2025 *Communications Biology* paper on **BrainLLM** by Ying et al. (with collaboration from researchers at multiple institutions) extends the semantic decoding paradigm by integrating brain-recording-derived representations directly into the autoregressive generation phase of a large language model. The original Huth-Tang decoder operated by mapping brain activity to candidate text and then scoring those candidates against a language model; BrainLLM instead trains a **brain adapter** module that maps brain activity directly into the hidden-vector representation space compatible with the language model's text embeddings, then prepends the brain-derived embeddings to the text prompt for generation. The language model thereby produces text conditioned simultaneously on textual context and on neural-activity-derived semantic content, treating the brain signal as additional input rather than as a target to match. The architectural significance of this shift is substantial. The brain signal is now an **input modality** for the language model, structurally analogous to how visual signals are integrated in vision-language models through projection from visual encoder representations to language model embedding space. The integration of additional modalities — audio, touch, proprioception, emotional state — follows the same architectural pattern. The trajectory points toward unified multimodal foundation models that incorporate neural-activity-derived representations alongside conventional sensory modalities, allowing brain signals to participate in language model conditioning on equal footing with other inputs. The full implications of this trajectory for brain-computer interface design, assistive communication, and human-AI interaction more broadly are still being worked out, but the technical foundation is now established. ### VI.3 — fNIRS Portability and the Wearable Decoding Path The fMRI infrastructure that supports the Huth-Tang decoder requires room-scale superconducting magnet systems and is unlikely to be the eventual deployment substrate for semantic decoding. The portable analog is **functional near-infrared spectroscopy** (fNIRS), which measures cortical hemodynamic changes through optical sensors placed on the scalp. fNIRS detects the differential absorption of near-infrared light by oxygenated versus deoxygenated hemoglobin, producing a hemodynamic signal qualitatively similar to the BOLD signal that fMRI records, but at substantially lower spatial resolution (centimeters rather than millimeters) and limited to cortical regions accessible from the scalp surface. The technique has been deployed in wearable form factors including headbands and cap-mounted optode arrays, with commercial systems from companies including Kernel, Artinis, NIRx, and Mendi. Alex Huth has stated explicitly that the semantic decoding approach should translate to fNIRS because fNIRS measures the same underlying class of hemodynamic signal that fMRI uses, while acknowledging that the lower spatial resolution will limit decoding fidelity. The current state of the literature supports cautious optimism: fNIRS-based decoding of word-level and short-phrase content has been demonstrated, with continuous semantic decoding still lagging fMRI substantially but improving as fNIRS instrumentation density and algorithmic methods advance. The trajectory toward consumer-form-factor wearable semantic decoding remains speculative in timing but credible in direction; the technical pieces required — high-density fNIRS arrays, subject-specific encoding models, language-model integration — are individually feasible, with the integration into deployed systems being primarily an engineering and clinical-validation effort rather than a fundamental research problem. ### VI.4 — Speech Decoding from Implanted Electrocorticography A parallel and substantially more developed line of work decodes attempted speech directly from invasive electrode arrays placed on the cortical surface — **electrocorticography** (ECoG) — in patients with severe motor impairments who have lost the ability to speak. The University of California San Francisco group led by Edward Chang has demonstrated speech neuroprostheses that decode attempted speech from ECoG signals at rates approaching natural conversational speech, with continued improvements through 2024 and 2025 bringing performance toward the threshold at which the systems become practically useful for clinical deployment. The companion work at Stanford by Krishna Shenoy (until his death in 2023), Jaimie Henderson, and Frank Willett has demonstrated speech decoding from intracortical microelectrode arrays in the motor speech areas, achieving high-accuracy speech reconstruction from attempted speech in ALS patients who can no longer vocalize. The Stanford team's **BrainGate** clinical trial infrastructure has served as the substrate for much of this work, building on more than two decades of intracortical BCI research. The speech-decoding work is methodologically distinct from the Huth-Tang semantic decoding in that it operates on the motor representation of speech rather than the semantic representation of meaning, and uses invasive recording rather than fMRI. The output is reconstructed speech intended by the patient rather than reconstructed semantic content of arbitrary thoughts. The technical advance is in the direct utility for severely impaired patients: the systems are progressing from research demonstrations toward clinical use, with the regulatory pathway through FDA breakthrough device designation accelerating commercial development. The convergence with foundation language models is increasingly explicit: contemporary speech decoders incorporate language model priors to improve accuracy and naturalness of decoded speech, integrating brain-derived motor signals with language-model contextual prediction to produce fluent output from sparse neural input. ### VI.5 — The Broader Semantic Bridge The decisive structural insight from semantic decoding research is that **biological semantic representation aligns approximately with the high-dimensional embedding spaces of modern foundation models**. This alignment is not a result of imitation — the language models were not trained to match brain activity, and the brain did not evolve to match language model representations. The alignment emerges because both systems are solving structurally similar problems: compressing meaning into high-dimensional codes that support generalization, composition, and inference. The implication is that the representational substrates of biological and synthetic intelligence are not as different as their underlying mechanisms might suggest. They are converging, through independent optimization, toward similar functional architectures because they face similar computational demands. The methodological consequence is that the same machinery used to align modalities within foundation models — projection from one embedding space to another through learned adapter networks — applies directly to the alignment of neural-activity-derived representations with language-model embeddings. The brain becomes another modality, integrable through the same architectural patterns. The trajectory of the field is toward unified multimodal foundation models that incorporate neural signals as a first-class input modality, with the practical applications spanning assistive communication for impaired patients, augmented cognitive interfaces for typical users, neural-feedback-informed AI training, and the deeper research question of how biological and synthetic representational systems can be productively aligned. --- ## Movement VII — Interface Deployment: The Neurointerface Stack from Wearable to Implanted ### VII.1 — The Stratified Interface Architecture The neurointerface stack spans an enormous range of invasiveness, bandwidth, spatial resolution, and deployment readiness. At the least invasive extreme are wearable sensors that record neural-related signals through the scalp or skin without any tissue penetration; at the most invasive extreme are penetrating electrode arrays implanted directly into cortical tissue with hundreds or thousands of recording sites. The intermediate space includes subdural electrocorticography arrays, endovascular electrodes inserted through blood vessels, and emerging non-invasive technologies including high-density EEG, fNIRS, and functional ultrasound. The architecture of the field is a stratified deployment pyramid: high-volume consumer applications at the wearable end, specialized clinical applications at the implanted end, and a rapidly developing middle layer of minimally invasive technologies that promise broader deployment than current implants while offering higher bandwidth than wearables. ### VII.2 — Wearable EEG and the Consumer Brain-Signal Layer Conventional **electroencephalography** uses scalp electrodes to record the summed electrical activity of large populations of cortical neurons, with millisecond-scale temporal resolution but centimeter-scale spatial blur due to volume conduction through skull and scalp. EEG has been a clinical neurology tool since Hans Berger's first recordings in 1924 and remains the dominant technique for sleep monitoring, epilepsy diagnosis, and event-related potential research. The contemporary innovation is the migration of EEG from laboratory equipment toward consumer wearable form factors, driven by the development of dry electrodes that eliminate the conductive gel previously required, miniaturized amplifiers, wireless data transmission, and machine-learning analysis pipelines that extract usable information from noisier consumer-grade recordings. **In-ear EEG** systems embed dry electrodes within or around earbud-form-factor devices, recording brain activity from electrodes placed near the ear canal. IDUN Technologies' **Guardian** platform, developed in collaboration with Analog Devices, integrates in-ear EEG with the ADI ADPD7000 ultra-low-power analog front-end and provides a developer kit for researchers building brain-state-aware audio applications. Peer-reviewed work in *Nature Communications* has demonstrated wireless in-ear dry-electrode systems used for drowsiness detection, attention monitoring, and sleep staging. **Neurable**, an American company, manufactures EEG-equipped headphones marketed for productivity and focus applications, providing real-time feedback on cognitive state through the same form factor as conventional consumer audio. **Muse**, developed by InteraXon, has been the dominant consumer EEG meditation device for nearly a decade and continues to develop the platform for sleep, focus, and attention applications. **Emotiv** and **NeuroSky** offer commercial EEG headsets for research and consumer applications spanning gaming, cognitive monitoring, and neurofeedback training. The trajectory of consumer EEG is toward ambient brain-state monitoring integrated into devices that users already wear continuously — headphones, glasses, smartwatches, hearing aids. The applications are not direct neural input (the resolution is too low for that) but rather **brain-state-aware** computing: devices that adapt to user fatigue, attention, stress, or cognitive load without requiring explicit user action. The deployment trajectory aligns with the broader move toward ambient computing in which sensors continuously inform digital systems about user state and environmental context. ### VII.3 — Surface Electromyography and Neuromotor Wristbands **Meta's Neural Band**, announced as part of the Ray-Ban Display ecosystem and the upcoming Orion AR glasses, uses surface **electromyography** (sEMG) to detect the electrical signatures of intended hand and finger movements at the wrist. The system records muscle activity through skin-surface electrodes and uses machine learning to decode the user's intended gestures — sometimes before the gesture is visibly executed, since the neuromotor command precedes the muscle contraction it produces. Meta's technical materials emphasize millisecond-scale response times and the ability to decode subtle movements that would not be visible to camera-based hand tracking systems. The Neural Band's deployment context is augmented reality input: the user wears AR glasses for visual display and the wristband for input, eliminating the need for handheld controllers while preserving precise gesture control. The technical lineage of the Neural Band runs through **CTRL-Labs**, the New York startup founded by Thomas Reardon (creator of Internet Explorer) and Patrick Kaifosh that Meta acquired in 2019 for approximately \$500 million to \$1 billion. CTRL-Labs' core technical claim was that surface EMG could decode individual motor unit activity — the electrical signature of single motoneurons in the spinal cord controlling small groups of muscle fibers — providing a much finer-grained read of motor intent than conventional muscle-level EMG. A 2024 *Nature* paper by Meta-affiliated researchers, including the CTRL-Labs founders, described a generic non-invasive neuromotor interface based on surface EMG that can be calibrated rapidly across users and supports a wide range of gestural inputs. The strategic positioning is that surface EMG provides a **practical bridge** between fully non-invasive interfaces (low-bandwidth, no decoding of intent) and implanted interfaces (high-bandwidth, surgical risk and limited deployment scale) — capturing the motor command before it produces visible movement, at substantially higher precision than camera-based or accelerometer-based input, without the regulatory or surgical barriers of implanted systems. ### VII.4 — fNIRS, fUS, and Emerging Wearable Hemodynamic Imaging Wearable hemodynamic imaging extends the principles of fMRI into deployment-feasible form factors. **Functional near-infrared spectroscopy** systems including those from Kernel (originally founded by Bryan Johnson with substantial investment in helmet-form-factor brain imaging), Artinis, NIRx, and academic-clinical hybrids combine arrays of near-infrared light sources and detectors mounted in head-worn assemblies to measure cortical hemodynamic changes at centimeter-scale resolution. Kernel's **Flow** helmet, introduced in 2020 and continuing development through 2025, uses time-domain fNIRS to provide whole-cortex hemodynamic imaging in a head-worn form factor; the company has pivoted between consumer and research positioning multiple times but the technical platform remains substantial. **Functional ultrasound** wearables represent the next-frontier modality: research-grade systems developed at the Sorbonne, ESPCI Paris, the Pasteur Institute, and the California Institute of Technology have demonstrated whole-brain ultrasonic hemodynamic imaging in rodents and primates, with the ambition to develop human-deployable systems at sub-millimeter resolution. The clinical deployment of fUS for intraoperative monitoring during neurosurgery is underway; the wearable consumer/research deployment timeline is longer but technically credible. ### VII.5 — Synchron and the Endovascular Approach The **endovascular** brain-computer interface, developed by Synchron (founded by Tom Oxley) and not requiring open-brain surgery, represents a distinctive architectural compromise within the implanted interface space. The **Stentrode** is a self-expanding electrode array delivered through standard catheterization procedures into the superior sagittal sinus — a large venous structure on the dorsal surface of the brain — where it deploys against the vessel wall and records electrical activity from the underlying motor cortex through the vessel tissue. The surgical procedure is performed by interventional neuroradiologists rather than neurosurgeons, takes a few hours, and recovers in days rather than weeks; the device delivers approximately 16 channels of neural recording at the cortical surface, far fewer than penetrating arrays but sufficient for binary or low-dimensional motor decoding. Synchron's **COMMAND** clinical trial has implanted the Stentrode in over 50 patients with severe paralysis, including ALS patients, who have used the device to control computer cursors, send messages, and operate digital interfaces through decoded motor cortex activity. The May 2025 announcement of native integration with Apple's Switch Control accessibility framework — the BCI HID protocol — established the first operating-system-level integration of a brain-computer interface with mainstream consumer computing platforms. Through this protocol, Synchron's Stentrode communicates with iPhone, iPad, and Apple Vision Pro on the same software footing as conventional accessibility input devices, allowing patients to use standard consumer devices through neural control without requiring custom software or specialized infrastructure. The trajectory points toward Synchron's pivotal FDA trial in 2026 for the first formal premarket approval pathway for an implantable BCI device, which will establish the regulatory template for the entire implanted-BCI industry. ### VII.6 — Neuralink and Penetrating Intracortical Arrays **Neuralink**, founded by Elon Musk in 2016, pursues the high-bandwidth penetrating-array approach. The **N1** implant integrates over 1,000 electrodes distributed across 64 ultra-thin flexible polymer threads — each thread approximately 5 micrometers wide, finer than a human hair — that are inserted into cortical tissue by the company's purpose-built **R1** surgical robot. The threads are designed to record from individual neurons at multiple depths within the cortex, with the implant transmitting recorded data wirelessly through a coin-sized device implanted in the skull, eliminating the percutaneous connectors that have plagued previous intracortical BCI systems with infection risk. The first human implant occurred in January 2024 in the **PRIME** clinical trial; by late 2025 the company had implanted approximately a dozen patients across U.S. and Canadian sites, with the **GB-PRIME** UK trial extending to additional patients in London beginning in October 2025. As of January 2026, seven patients are enrolled in the GB-PRIME study at University College London Hospitals, with surgeries performed at the National Hospital for Neurology and Neurosurgery. The clinical results have demonstrated cursor control, text composition, and operation of consumer applications at speeds and accuracies approaching those reported by earlier BrainGate research, with one significant complication: thread retraction from optimal cortical positions over the months following implantation, reducing the number of electrodes capturing useful signals. Neuralink's software adaptation extracted increased information from the remaining electrodes, partially compensating for the hardware issue; subsequent implants reportedly improved thread retention through modified surgical techniques. Neuralink's second-generation device, reportedly in development for 2026 trials, addresses the thread retraction issue with a revised electrode design. The company aims to scale from current single-digit patient counts to broader clinical trials with dozens of participants over the coming years, generating the safety and efficacy data needed for eventual FDA commercial approval. ### VII.7 — Precision Neuroscience and the Surface-Array Architecture **Precision Neuroscience**, founded by Benjamin Rapoport (a Neuralink cofounder) and Michael Mager in 2021, pursues a third architectural approach: ultra-thin film electrode arrays placed on the cortical surface through a minimally invasive slit craniotomy procedure rather than penetrating the cortex. The **Layer 7 Cortical Interface** consists of an array of 1,024 electrodes embedded in a thin polymer film that conforms to the cortical surface, providing high-density surface recording without penetrating brain tissue. The procedure is reversible — the array can be removed if needed — and the recording quality is sufficient for high-bandwidth motor decoding while avoiding the long-term tissue damage associated with penetrating arrays. As of late 2025, the Layer 7 system had been tested in over 68 patients across six U.S. medical centers, primarily in intraoperative research contexts where the array is placed temporarily during surgeries already being performed for other indications. Precision Neuroscience filed what may be the first BCI premarket approval application with the FDA in 2025, positioning the company to potentially achieve the first commercial BCI approval. A partnership with Medtronic for manufacturing and clinical deployment was announced in 2025, providing the established medical-device infrastructure that startup BCI companies typically lack. ### VII.8 — BlackRock Neurotech, Paradromics, and the Continuing Ecosystem **BlackRock Neurotech** is the longest-operating commercial implanted-BCI company, with the **Utah Array** — a 96- or 100-electrode penetrating silicon array originally developed by Richard Normann at the University of Utah in the 1990s — serving as the recording substrate for the BrainGate clinical trials and the majority of academic implanted-BCI research over the past two decades. Approximately 50 patients have received Utah Array implants through various research protocols, with the longest-running implants exceeding 17 years of continuous operation. BlackRock's contemporary **MoveAgain** product is positioned for commercial deployment of motor BCI for paralyzed patients, with FDA breakthrough device designation accelerating the regulatory pathway. **Paradromics**, founded by Matt Angle, develops the **Connexus Direct Data Interface** — a high-bandwidth intracortical interface using arrays of microwire electrodes designed to scale to tens of thousands of recording channels. The architectural distinction is the use of solid microwires rather than the flexible polymer threads of Neuralink or the conforming surface films of Precision Neuroscience; the trade-off is greater rigidity in exchange for more electrodes per implant and longer-term recording stability. Paradromics has demonstrated its high-bandwidth interface in animal models and is approaching first-in-human trials with locked-in-syndrome patients, with the goal of decoding speech at rates approaching natural conversation. **Onward Medical**, **Cortec**, **NeuroPace** (for clinical epilepsy management), **BrainGate** (the academic consortium continuing intracortical BCI research), and **Inbrain Neuroelectronics** (graphene-based electrodes) populate the remainder of the implanted-interface ecosystem. ### VII.9 — Operating-System Integration and the BCI HID Layer The decisive infrastructural transition of 2025 was Apple's announcement of the **BCI HID** (Brain-Computer Interface Human Interface Device) protocol — a native operating-system specification for accepting input from brain-computer interface devices on equal footing with keyboards, mice, and touchscreens. Synchron's Stentrode was the first device certified under the BCI HID protocol, with native integration through Apple's Switch Control accessibility framework allowing the Stentrode to operate iPhone, iPad, and Vision Pro without custom software. Apple's announcement of BCI HID support in May 2025 effectively created the **API surface** through which brain-computer interfaces enter consumer computing infrastructure. The implications extend well beyond Synchron and Apple: the BCI HID specification provides a path for any BCI vendor to integrate with Apple's operating-system ecosystem, dramatically reducing the integration burden for device manufacturers and providing a standardized interface for application developers building BCI-aware software. The trajectory is toward equivalent integration in other operating systems and platforms. Microsoft's accessibility frameworks, Google's Android ecosystem, and Linux distributions through accessibility standards are all converging toward similar standardized BCI input pathways. The result is that the **operating-system substrate** is becoming BCI-ready in advance of widespread BCI deployment — the software infrastructure to accept brain signals as input is being built before the hardware capable of providing those signals is widely available. This is the conventional pattern for emerging input modalities: the software infrastructure precedes the hardware, providing the foundation on which subsequent hardware deployment can scale. ### VII.10 — Sensor Stacks in Spatial Computing Platforms The broader sensor stack of contemporary spatial computing platforms — Apple Vision Pro, Meta's Quest and Orion lines, Microsoft HoloLens — increasingly integrates eye tracking, hand tracking, facial expression recognition, and biometric monitoring alongside the traditional visual and audio sensors. Apple Vision Pro's sensor array includes high-resolution main cameras, world-facing tracking cameras, four eye-tracking cameras using infrared LEDs, TrueDepth structured-light sensors, LiDAR for environmental mapping, inertial measurement units, and ambient light sensing. The eye-tracking subsystem alone provides a high-bandwidth user-state signal: gaze direction, pupil dilation, blink patterns, and saccadic dynamics all reveal aspects of attention, cognitive load, and emotional state. Combined with the EEG and EMG capabilities of complementary wearables, the spatial computing platform constitutes a substantial **biosensor environment** even without conventional brain-computer interfaces. The convergence of explicit BCI hardware (Stentrode, N1, Layer 7) with the broader biosensor stack of spatial computing platforms produces an integrated sensor environment in which neural signals are one channel among many that inform digital systems about user state and intent. --- ## Movement VIII — Hybrid and Biohybrid Architectures: The Convergence of Substrates ### VIII.1 — Bio-Silicon Integration Patterns The architectural pattern of **biohybrid computation** — combining living neural tissue with synthetic computing substrate within a single system — recurs across multiple projects at multiple scales. The simplest form is the multielectrode array interface that underlies DishBrain, the CL1, the FinalSpark Neuroplatform, Brainoware, and essentially every contemporary in-vitro neural computing platform: a microfabricated silicon substrate hosts the electronics for stimulation, recording, and signal processing, while living neurons (dissociated cultures or organoids) constitute the computational substrate. The silicon handles the interfacing, communication, life-support monitoring, and digital integration; the biology handles the computation. This bipartite division of labor is the dominant pattern, instantiated with varying levels of integration sophistication across different platforms. More elaborate biohybrid architectures combine living tissue with neuromorphic chips, creating systems in which the biological and synthetic components both perform computation and interact through structured input-output coupling. The 2023 demonstration by the Frégnac laboratory at the Paris-Saclay Institute of Neuroscience coupled rat cortical cultures on multielectrode arrays with Loihi neuromorphic chips, with the two substrates exchanging spike events through the array interface. The biological tissue and the silicon network co-adapted under task feedback, demonstrating that hybrid bio-silicon networks can be operated as integrated computational systems with both substrates contributing to the overall function. Subsequent work has extended this approach to organoid-Loihi pairings, fNIRS-Loihi sensory-processing pipelines, and architectures in which neuromorphic chips serve as preprocessing stages for biological substrate or vice versa. ### VIII.2 — Sensory and Effector Coupling to Living Tissue The closed-loop architecture of contemporary in vitro neural computing increasingly couples living tissue to physical sensors and effectors rather than purely simulated environments. The University of Bristol Braille-reading demonstration, conducted in collaboration with FinalSpark, used tactile sensor data from a robotic fingertip as input to FinalSpark organoids, with the organoid's response activity decoded to classify Braille characters. **Retinal organoid** interfaces have been demonstrated for visual input, exploiting the natural photoreceptor function of differentiated retinal tissue. **Cochlear organoid** preparations for auditory processing remain at earlier development stages but represent a natural extension of the modality-specific organoid approach. **Assembloid** architectures, in which multiple regional organoid types (cortical organoids combined with thalamic organoids, sensory cortex combined with motor cortex) are physically fused, recapitulate aspects of inter-regional brain connectivity and support computations that single-region organoids cannot perform. The trajectory points toward **integrated biohybrid robotics** in which living neural tissue serves as the controller for sensorimotor systems, with appropriate sensory inputs encoded as electrical stimulation patterns and motor outputs decoded from neural activity. Cortical Labs and FinalSpark have both publicly discussed robotics applications as part of their longer-term roadmaps; academic research at Bristol, Caltech, Stanford, and several Japanese institutions is actively pursuing organoid-robot integration. The applications include adaptive prosthetics that adjust to user-specific neural patterns, biological controllers for environments where electronic systems face fundamental limitations (extreme temperatures, high radiation, biological-chemical environments), and research platforms for studying embodied cognition and sensorimotor learning in tractable in-vitro preparations. ### VIII.3 — Distributed Wetware and Cloud Architectures The FinalSpark Neuroplatform's **cloud-accessible wetware** architecture represents a more abstract form of biohybrid integration: living neural tissue is the computational substrate, but the substrate is operated remotely by users through standard network and software interfaces. The 16-organoid array at FinalSpark's Vevey facility constitutes a small-scale prototype of what could become a **distributed wetware computing infrastructure**, in which physically centralized biological computing resources are accessed by geographically distributed users through standard cloud APIs. The Cortical Labs **Cortical Cloud** offering, which provides remote access to CL1 systems, follows the same architectural pattern. The pattern is the conventional cloud-computing one — centralized infrastructure operated by specialists, accessed by users through standard interfaces — applied to a substrate that requires biological life-support infrastructure rather than just power and cooling. The implications of this architecture are significant. Biological computing has historically been constrained to laboratories with cell-culture infrastructure and biological expertise, limiting its accessibility to a small number of specialized researchers. The cloud-wetware model makes biological computation available to any researcher with network access and a Python development environment, dramatically broadening the participant base for biological computing research. The trajectory points toward larger-scale cloud-wetware deployments, possibly involving thousands of organoids operated as a unified compute resource, with applications spanning drug screening, neurodegenerative disease research, personalized medicine (organoids derived from patient-specific iPSCs), and energy-efficient AI workloads exploiting the biological substrate's intrinsic computational economy. ### VIII.4 — The Economics of Substrate Selection The choice between organic and synthetic substrates is increasingly an **engineering optimization** rather than a categorical commitment. The organic substrate offers extreme energy efficiency, intrinsic adaptive plasticity, and biologically relevant physiology — at the cost of long maturation times, biological life-support infrastructure, limited scalability, and bioethical constraints that grow with substrate complexity. The synthetic substrate offers scalability, reproducibility, and engineering control — at the cost of higher energy per operation than biological tissue, more limited plasticity (though improving with on-chip learning capabilities), and the need to explicitly engineer features that biology provides for free. For tasks requiring high-throughput inference on standardized models with well-understood requirements, synthetic substrate is decisively advantageous. For tasks requiring adaptive learning from sparse data, biologically relevant physiology, or extreme energy efficiency, organic substrate offers advantages that synthetic systems cannot yet match. For tasks requiring both, hybrid architectures that integrate both substrates are increasingly the practical choice. The economic logic of this division is straightforward: each substrate is deployed where its comparative advantages are most pronounced, and the integration architecture handles the coupling between them. The field is therefore moving away from the historical "organic versus synthetic" framing toward a "organic and synthetic" framing in which the substrates are complementary tools within a unified computational toolkit. --- ## Movement IX — Infrastructure Layer: Funding, Regulation, and Institutional Geography ### IX.1 — Public Funding Programs The connectomic and neurotechnological infrastructure that underlies contemporary brain-grounded computing has been substantially supported by major public funding programs that span the past two decades. The **BRAIN Initiative** (Brain Research through Advancing Innovative Neurotechnologies), launched by the Obama administration in 2013, has invested over \$3 billion through the National Institutes of Health, the National Science Foundation, DARPA, IARPA, the Food and Drug Administration, and the Department of Energy in coordinated brain research infrastructure. The NIH BRAIN Initiative supports tool development, cell-type atlases, connectomic infrastructure, and clinical neurotechnology programs. The NSF BRAIN component supports more basic research on brain dynamics, cognitive architectures, and computational neuroscience. The DARPA programs include Neural Engineering System Design (NESD), Next-Generation Nonsurgical Neurotechnology (N3), and various spiking neural network and neuromorphic computing initiatives. The IARPA **MICrONS** program funded the cubic-millimeter functional connectomic effort that produced the MICrONS dataset. The **European Human Brain Project**, running from 2013 to 2023 with approximately €1 billion in EU funding, supported the development of the SpiNNaker and BrainScaleS neuromorphic platforms, the **EBRAINS** research infrastructure, and substantial computational neuroscience work distributed across European institutions. The successor program, **EBRAINS 2.0**, continues European neurotechnology infrastructure development with explicit integration of neuromorphic computing, connectomic data, and clinical neurotechnology applications. The European Innovation Council and various national funding programs continue parallel investment in neurotechnology startups and research infrastructure. China's **China Brain Project**, formally launched in 2016 with substantial subsequent investment, supports the development of Chinese connectomic infrastructure (centered at institutions including the Shanghai Institute for Biological Sciences and the Chinese Academy of Sciences), neuromorphic computing (Tianjic at Tsinghua, Darwin at Zhejiang, and various commercial efforts), and clinical neurotechnology with substantial integration into the broader Chinese AI and biotechnology investment programs. **Japan's Brain/MINDS** project focuses on non-human primate connectomic mapping and developmental neuroscience. **Korea's Brain Initiative**, **Australia's Brain Initiative**, **Canada's Brain Initiative**, and other national programs constitute a global infrastructure of public brain-research investment that has scaled substantially over the past decade. ### IX.2 — Commercial Investment and Industrial Strategy The commercial investment in brain-grounded computing has accelerated substantially since 2020. Neuralink raised approximately \$650 million across multiple rounds, reaching a reported valuation of \$5 billion in 2023 and continued growth through subsequent rounds. Synchron has raised approximately \$145 million across rounds led by Khosla Ventures, Bezos Expeditions, and Bill Gates' personal investment vehicle. Precision Neuroscience has raised over \$150 million including strategic investment from Steadview Capital, Mubadala, and the Forest Road Company. Paradromics has raised approximately \$80 million. Cortical Labs raised approximately \$10 million in 2023 with continued growth through the CL1 commercial launch. FinalSpark has raised approximately \$5 million with continued institutional and angel funding supporting the Neuroplatform expansion. The overall commercial neurotechnology market is projected to grow from approximately \$14 billion in 2023 to over \$24 billion by 2030 according to various market analyses, with the specific brain-computer interface segment growing more rapidly than the broader neurotech market. The industrial strategy of large technology companies has increasingly converged on integrated brain-computer interface and ambient computing approaches. Apple's BCI HID protocol and Synchron partnership represent the most explicit integration of BCI capabilities into mainstream consumer computing platforms. Meta's CTRL-Labs acquisition, Neural Band development, and continuing investment in surface EMG, fNIRS, and broader neurotechnology research positions the company for AR/VR-integrated neural input. Google's MICrONS contributions, Connectomics team, MoGen development, and broader brain-mapping infrastructure place the company at the center of the connectomic data infrastructure. Microsoft's investment in BCI accessibility frameworks and the broader Microsoft Research neuroscience program provides parallel infrastructure development. The Chinese technology ecosystem includes substantial neurotechnology investment from Tencent, Alibaba, and various state-backed venture funds, with the technology development closely aligned to the broader Chinese AI and biotechnology industrial strategy. ### IX.3 — Regulatory Pathways The regulatory pathway for brain-computer interfaces in the United States runs through the FDA's medical device classification system, with all current implanted BCIs operating under research-protocol Investigational Device Exemptions or Breakthrough Device Designations. The first formal premarket approval (PMA) for a commercial implanted BCI has not yet been granted; Synchron's anticipated 2026 pivotal trial submission and Precision Neuroscience's 2025 PMA filing represent the leading candidates for first commercial approval. The regulatory standard requires demonstration of both safety (acceptable surgical and long-term risk profiles) and efficacy (clinically meaningful functional improvement for specific indications). The likely first commercial indications are motor restoration for severely paralyzed patients (ALS, high-cervical spinal cord injury, locked-in syndrome from brainstem stroke), with subsequent indications expected to expand into communication assistance, environmental control, and eventually broader cognitive and sensory restoration applications. The regulatory status of **organoid intelligence** is substantially less defined. Brain organoids used for research purposes operate under conventional biological research regulations (institutional review boards, biosafety committees, animal-use protocols for any animal-derived components). The use of organoids as computational substrate raises novel regulatory questions that have not yet been formally addressed: at what level of complexity does an organoid require human-subjects-like protection? What are the appropriate ethical frameworks for organoids derived from patient iPSCs when the organoids exhibit complex spontaneous neural activity? How should commercial deployment of organoid-based computing systems be regulated? The Johns Hopkins-led **Baltimore Declaration** on organoid intelligence and the associated bioethics literature address these questions in advance of formal regulatory frameworks, but the field is operating in a regulatory space that will require explicit definition as deployment scales. The regulatory pathway for non-invasive neurotechnology is generally less stringent. Wearable EEG, EMG, and fNIRS devices marketed for wellness, productivity, or consumer applications operate under standard consumer electronics regulations with minimal FDA oversight. Devices marketed for medical indications face more substantial regulatory requirements but typically through less burdensome 510(k) clearance pathways rather than the full PMA process required for implanted BCIs. The regulatory asymmetry between invasive and non-invasive neurotechnology reinforces the broader industrial strategy of deploying non-invasive technology to consumer markets at scale while pursuing invasive technology through specialized clinical applications. ### IX.4 — Institutional Geography The institutional geography of brain-grounded computing is concentrated at a relatively small number of centers that combine substantial computational, biological, clinical, and engineering capabilities. In the United States: **Harvard Medical School** (the Lichtman lab and H01); **Princeton University** (FlyWire under Seung and Murthy); **Allen Institute for Brain Science** in Seattle (MICrONS, mouse brain atlas, cell-type databases); **HHMI Janelia Research Campus** in Ashburn, Virginia (FlyEM, connectome-constrained models, ZAPBench); **Stanford University** (BrainGate, speech BCIs, computational neuroscience); **University of California San Francisco** (speech BCIs under Chang); **MIT** (the McGovern Institute, including Sven Dorkenwald's mouse-brain effort beginning January 2026); **Caltech** (Andersen lab BCIs); **Carnegie Mellon University** (computational neuroscience, neuromorphic computing); **Johns Hopkins** (organoid intelligence under Hartung); **Indiana University Bloomington** (Brainoware under Feng Guo); **University of Texas at Austin** (semantic decoding under Huth); **Columbia University** (multiple connectomic and BCI programs); **University of Washington** (Center for Neurotechnology); **University of Pittsburgh** (BCI clinical research). In Europe: **University of Cambridge** and the MRC Laboratory of Molecular Biology (fly connectomics); **University College London** (Neuralink GB-PRIME trial site, computational neuroscience); **Imperial College London**; **University of Oxford** (Centre for Neural Circuits and Behaviour); **University of Bristol** (organoid robotics); **TU Dresden** (SpiNNaker 2, SpiNNcloud); **University of Manchester** (original SpiNNaker, continuing computational neuroscience); **Heidelberg University** (BrainScaleS 2, European Institute for Neuromorphic Computing); **École Polytechnique Fédérale de Lausanne (EPFL)** (Blue Brain Project, organoid research); **Institute of Science and Technology Austria (ISTA)** (LICONN, in collaboration with Google Research); **University of Geneva** (Lemanic Neuroscience); **Karolinska Institute**; **Max Planck Institutes** (Florian Engert at MPI for Brain Research and others); **Sorbonne Université** and **ESPCI Paris** (functional ultrasound); **Pasteur Institute**; **Italian Institute of Technology**. In Asia: **Tsinghua University** (Tianjic neuromorphic computing); **Peking University** (computational neuroscience, neurotechnology); **Zhejiang University** (Darwin neuromorphic chip); **Shanghai Institute for Biological Sciences**; **RIKEN Center for Brain Science** (Brain/MINDS program); **University of Tokyo**; **Korean Brain Research Institute**; **Singapore's A*STAR Institute of Molecular and Cell Biology**; **Indian Institute of Science (IISc) Bangalore** (computational neuroscience). The geographic concentration reflects the substantial infrastructural requirements of contemporary brain-grounded computing — electron microscopy facilities, cell-culture infrastructure, clinical trial capabilities, computational resources, and specialized technical workforces — that are most readily assembled at major research universities and dedicated research institutes. The distribution is becoming somewhat more even through cloud-accessible platforms (FinalSpark Neuroplatform, Cortical Cloud, the Neuroglancer-based public datasets) that allow researchers at less-resourced institutions to participate in the field without requiring local infrastructure investment. ### IX.5 — Open Data and Reproducibility Infrastructure The contemporary infrastructure of brain-grounded computing is substantially built on **open data** and shared software tools. The H01 human cortical reconstruction is publicly browsable through Neuroglancer. The FlyWire connectome is openly accessible with annotation and analysis tools. The MICrONS dataset, with both the structural connectome and functional recordings, is openly distributed with the Cortical MM³ infrastructure. ZAPBench, the larval zebrafish whole-brain activity dataset, is openly distributed as a benchmark. The Allen Brain Atlas provides openly accessible cell-type databases, gene expression atlases, and developmental data. The International Brain Laboratory openly distributes its standardized behavioral and neural data across collaborating laboratories. PathFinder reconstruction software, MoGen synthetic morphology generation, Neuroglancer visualization, TensorStore data infrastructure, CAVE proofreading systems, and many other components are open-source and openly licensed. This open infrastructure is unusual in the contemporary biomedical research environment, where proprietary data and tooling are increasingly common. The field's commitment to openness reflects both the genuine intellectual value of broad data accessibility (more researchers analyzing the same data produces more discovery than restricted access) and the practical recognition that the connectomic datasets are too large, too complex, and too expensive to produce for any single institution to fully exploit alone. The open infrastructure thereby functions as a coordination mechanism: the data are produced once and analyzed by the entire global research community, with discoveries published openly and the cumulative knowledge accreting in publicly accessible form. The trajectory points toward continued expansion of this open-data model as new datasets are produced and new platforms are deployed. --- ## Movement X — Synthesis: The Continuity Architecture in Operation The architecture that emerges from the accumulated material is a layered technical regime spanning substrate, infrastructure, and interface. At the substrate layer, organic neural tissue and synthetic neuromorphic hardware constitute two complementary forms of brain-grounded computation, with the historical methodological boundary between them dissolving under shared toolchains, shared benchmarks, and increasingly explicit hybrid architectures. At the infrastructure layer, connectomic mapping pipelines convert nervous tissue into computationally addressable graph structure, generative models such as MoGen accelerate that conversion by producing synthetic morphologies that train the reconstruction systems, and the resulting datasets feed both basic neuroscience and the constrained machine-learning models that operationalize biological wiring diagrams as executable computational systems. At the interface layer, brain-computer interfaces span the full range from wearable EEG and EMG through endovascular and surface arrays to penetrating intracortical implants, with operating-system-level integration through protocols like Apple's BCI HID making neural signals a first-class input modality alongside conventional human-interface devices. The recursive structure that MoGen exemplifies is the operating principle of the entire regime. Synthetic data generation accelerates the reconstruction of biological structure. Biological structure constrains synthetic models. Synthetic models inform biological experiments. Biological experiments produce data that further trains synthetic systems. Living neural tissue, organized through closed-loop interfaces, operates as adaptive computational substrate within hybrid bio-silicon systems. Neuromorphic hardware, increasingly programmable and energy-efficient, executes the synthetic side of the same architectural strategy. Semantic decoders align neural activity with the embedding spaces of foundation models, providing the representational bridge through which biological and synthetic intelligence communicate. The flywheel is operational across every layer of the regime simultaneously, and the technical maturation of any single layer accelerates every other. The connectome, in this architecture, is no longer purely an anatomical archive. It is the **structural prerequisite** for executable models of biological computation, the **inductive bias** for machine-learning systems that respect biological wiring constraints, and the **navigational substrate** through which biological intelligence becomes addressable, navigable, and integrable with synthetic systems. The FlyWire fly brain, the H01 human cortical fragment, the MICrONS mouse visual cortex, the ZAPBench larval zebrafish, and the developing mouse-brain efforts collectively constitute the first sustained mapping of biological intelligence at the spatial and synaptic scales required for computational fidelity. The trajectory toward whole-mammalian-brain connectomes within the next decade represents the maturation of this mapping infrastructure into a discipline with reliable methods, predictable timelines, and quantifiable cost structures — the conditions under which any infrastructure layer becomes operationally productive. The organic substrate of neural computation, organized through dissociated cultures, organoids, multielectrode arrays, and closed-loop digital environments, has transitioned from research demonstration to commercial deployment with the Cortical Labs CL1, the FinalSpark Neuroplatform, and the broader organoid intelligence ecosystem. Living neural tissue is now operationally available as computational substrate through cloud APIs, as a desktop computing appliance, and through academic research collaborations spanning multiple continents. The synthetic substrate, organized through neuromorphic chips, photonic accelerators, memristive devices, and hybrid digital-analog systems, has matured into commercial and research deployment with Intel Loihi 2, IBM NorthPole, SpiNNaker 2, BrainScaleS 2, and a substantial international ecosystem of specialized chips. Both substrates are now beyond the threshold of research curiosity and into the regime of operational engineering, with continuing improvements measured in incremental rather than revolutionary terms. The semantic layer, in which neural activity aligns with the representational spaces of foundation language models, has achieved a level of demonstrated capability that places it within the broader trajectory of multimodal AI integration. The University of Texas at Austin fMRI semantic decoder, BrainLLM's direct integration of brain signals into language model generation, and the parallel speech-decoding work at UCSF and Stanford collectively demonstrate that biological semantic representation and synthetic language model embeddings occupy structurally similar high-dimensional spaces — convergent solutions to similar computational problems arrived at through entirely independent optimization processes. The implication is that the deep representational architectures of biological and synthetic intelligence are aligning, not because either was designed to imitate the other but because both are optimizing for the same underlying compositional, generalizable, semantic compression. The bridge between brain and machine, in representational terms, is shorter than the substrate distinction would suggest. The interface stack, spanning the full range from consumer wearables through minimally invasive clinical devices to penetrating intracortical arrays, has reached operational deployment in clinical settings with Synchron's Stentrode, Neuralink's N1 implant, Precision Neuroscience's Layer 7 array, BlackRock's Utah Array, and the broader implanted-BCI ecosystem. The simultaneous emergence of operating-system-level BCI integration through Apple's BCI HID protocol, and the parallel development of consumer-grade wearable brain-signal sensors through Meta's Neural Band, in-ear EEG systems, and the broader spatial computing sensor stack, produces a stratified deployment architecture in which neural signals enter digital systems at multiple levels of bandwidth, invasiveness, and clinical specificity. The deployment is no longer a future possibility; it is a present operational reality at increasing scale. The infrastructure layer that supports this regime — public funding programs across the United States, Europe, China, Japan, and other major research economies; commercial investment scaling into the tens of billions of dollars; institutional research capacity concentrated at major universities and research institutes globally; open-data and shared-software conventions that distribute the cost of dataset production while broadening access — has matured into a self-sustaining technical ecosystem. The field is no longer dependent on speculative investment or visionary advocacy; it is a normal scientific and engineering discipline with established methodologies, predictable outputs, and increasingly clear commercial and clinical applications. The synthesis is therefore not aspirational but descriptive. The brain has become **computational terrain** in operational fact: a substrate that is being mapped at increasing resolution, simulated at increasing fidelity, interfaced at increasing bandwidth, and integrated with synthetic computing systems at increasing depth. The transition from brain-inspired AI to brain-grounded AI — from metaphorical neural networks to empirically constrained neural architectures — is no longer a future ambition but a developing present capability. The recursive flywheel that MoGen exemplifies — synthetic structure accelerating biological reconstruction, biological reconstruction constraining synthetic modeling, synthetic modeling informing biological experiment, biological experiment producing data that trains synthetic systems — is operational across every layer of the field simultaneously, and the rate of mutual reinforcement is accelerating. The biological substrate and the synthetic substrate, having developed along largely independent trajectories for the past seven decades since the McCulloch-Pitts neuron and the cybernetic founding of the field, are now converging into a single technical regime. Living neural tissue, neuromorphic silicon, photonic computing, memristive devices, and synthetic biological constructs are increasingly interoperable through shared interface standards, shared algorithmic frameworks, and shared engineering toolchains. The convergence is producing hybrid architectures that exploit the complementary advantages of each substrate, and the resulting systems are operational at commercial and clinical scale. The brain is no longer separate from the machines that study and increasingly integrate with it; the machines are no longer separate from the biological intelligence whose principles they extract and instantiate. The continuity architecture of biological intelligence — map, model, substrate, interface — is the operational architecture of the field as it stands in 2026, and the trajectory points toward continued deepening of every layer of integration over the coming decade. The Google Research MoGen paper, with its specific technical contribution of synthetic neuronal morphology generation through point-cloud flow matching, occupies a specific node within this larger architecture: it is the substrate-generative innovation that breaks morphological scarcity as a bottleneck in connectomic reconstruction, accelerating the rate at which biological wiring diagrams can be produced and therefore accelerating every downstream layer of the regime that depends on those diagrams. Its 4.4 percent error reduction is the signal flare; the larger story is the recursive structure it instantiates and the broader regime that recursive structure operates within. The brain is becoming computational terrain in operational fact, the substrate convergence between organic and synthetic neural computation is accelerating, and the technical infrastructure for biological intelligence to participate in synthetic computing systems on equal architectural footing is being assembled at every layer of the stack. The field has crossed the threshold from research demonstration to operational engineering, and the trajectory of the next decade is now substantially determined by the rate at which existing methodologies scale rather than by whether the fundamental approaches will work. --- [Bryant McGill](https://bryantmcgill.com/about/) is a Wall Street Journal and USA Today Best-Selling Author. He is the founder of Simple Reminders, a United Nations appointed Global Champion, and a Congressionally Recognized Ambassador of Goodwill. His work spans naval intelligence systems, computational linguistics, and civilizational governance architecture. His forward analysis on U.S.–Israel Pax Silica frameworks has appeared in Jewish News Syndicate (JNS). --- *End of field atlas. The document traverses, in present operational terms, the technical architecture of organic and synthetic neural computational systems as the field stands in 2026, with the connectomic mapping infrastructure, functional integration stack, semantic translation layer, and neurointerface deployment pyramid that together constitute the continuity architecture of biological intelligence becoming computationally addressable. The field is changing rapidly; numerical specifications, named systems, and institutional affiliations correspond to publicly documented information available at the time of writing. The recursive structure described — synthetic data accelerating biological reconstruction, biological reconstruction constraining synthetic modeling, hybrid architectures operating across both substrates — is the operative pattern across the field, and the trajectory of its further development over the coming decade is the substance of the work that follows.*