At this point own MRI makes sense. - Google Is Mapping the Human Brain

### TL;DR Google released the MoGen model to generate synthetic neurons, reducing brain connectome reconstruction error rates by 4.4% and saving 157 person-years of proofreading time. Meanwhile, AI can already decode thoughts from fMRI scans, portable brain sensors are becoming common, and brain privacy issues are imminent. ## Breakthrough in Synthetic Neurons: AI Training AI Two weeks ago, Google quietly published a paper describing a new AI model called MoGen that can generate synthetic neurons. The ping-pong game you're watching is being played by a group of brain cells in a petri dish — not biologically inspired neurons, but replicas of brain cells built from scratch with sufficient statistical precision that other AI systems use them as substitutes for real neurons. Google uses these synthetic neurons to train AI to decode real brains faster. AI creating fake neurons to help other AI understand real neurons — this sounds like 2050 science fiction, but it comes from a Google research blog post dated April 16, 2026. ## Connectomics: Mapping the Brain's Wiring Diagram The human brain has about 86 billion neurons, each connected to thousands of others via synapses, totaling over 100 trillion connections. Connectomics — mapping the connections between all nerve cells — aims to produce a complete wiring diagram of the brain. The first complete connectome, published in 1986, was that of a worm with only 302 neurons, which took researchers 16 years to complete manually. The fruit fly brain has 166,000 neurons; Google and its partners completed the connectome of the entire fruit fly brain and central nervous system, a multi-year collaboration between AI and human experts. The process involves slicing brain tissue thinner than one-thousandth the width of a human hair, imaging each layer with a high-speed electron microscope, stacking thousands of images, and having AI reconstruct hidden 3D neurons — every branch, every connection, every contact point. The mouse brain is 1,000 times larger than the fruit fly brain, and the human brain is 1,000 times larger than the mouse brain. If it takes years for the best current AI to complete the fruit fly connectome, completing the entire human connectome would take longer than recorded human history. Google's AI aims to close this gap. ## MoGen: Synthetic Neurons Reduce Manual Proofreading To train AI to accurately reconstruct neurons from microscope images, vast amounts of labeled training data are needed — real neurons annotated and verified by experts through manual proofreading, an extremely slow process. Google's Pathfinder model produces errors during reconstruction: two connected neurites might be split apart, or two unrelated ones merged, both requiring manual correction. MoGen changes this by generating synthetic neurons realistic enough to serve as additional training data. The synthetic shapes accurately capture branching complexity, geometric intricacy, and precise irregularities, and can be generated infinitely. The result is a 4.4% reduction in reconstruction error rate. At the scale of a complete mouse brain, this equates to saving 157 person-years of manual proofreading time — from a single model, a single paper, a blog post most people skimmed. The mouse brain is still far from the human brain; this is just the beginning. ## New Discoveries from Real Brain Maps In 2024, Google and Harvard published a paper in *Science* describing the reconstruction of a small piece of human brain tissue. The sample, only half the size of a grain of rice, required 1.4 PB of data (1.4 million GB). It contained 57,000 cells, 150 million synapses, and structures never seen by researchers. Dr. Jeff Lichtman, the Harvard neuroscientist co-leading the project, said he had never seen anything like it in his career. They found "storage" inside: some neurons' output lines (axons) were tied into knots, loops, and spirals with no obvious explanation. Researchers described them as "mysterious and beautiful." Some neuron pairs were connected by up to 50 simultaneous synapses, an abnormal degree. Deep layers also contained direction-selective neurons never documented before — they received predictable noise signals when responding correctly and random noise when wrong; they didn't like random noise. These structures were only discovered after AI helped map them, and only from a single patient, a single brain, a single region, half a grain of rice. Extending this to the entire organ, to every living person, has profound implications. ## From Wiring Diagrams to Functional Simulations Researchers at Princeton University recently mapped over 500 million connections in one cubic millimeter of mouse visual cortex, while simultaneously recording the activity of the same neurons — thus possessing both a wiring diagram and real-time electrical activity. One researcher said, "We think these are just the tip of the iceberg." Google Research also mapped the complete brain and central nervous system of a larval zebrafish — all 70,000 neurons — tracking them in real time as the fish responded to peripheral stimuli. Subsequently, researchers at HHMI Janelia used AI to construct a simulation of the fruit fly visual system based solely on the connectome: no actual neural activity measurements were needed, only the wiring diagram plus the circuit's goal (detect motion); everything else was inferred by AI. The simulation accurately predicted the behavior of 64 neuron types, reproducing results from over 20 experimental studies spanning 20 years. An AI looked at a wiring diagram, figured out what the circuit should do, and predicted each neuron's behavior, matching two decades of experiments it had never seen. Researchers said this bridges the gap between the static structure of connectomes and the dynamics of real-life computation. They next ask: "Can this scale to mice? And later, to larger organisms?" ## Thought Decoding: From fMRI to Portable Devices In the same year, the University of Texas at Austin published a paper in *Nature Neuroscience* building an AI system called a semantic decoder. Put a person in an fMRI scanner, read their brain activity, and the AI translates that activity into text — not just a word or a category, but continuous language paragraphs. It generates the gist of complex thoughts the person hears, reads, or imagines, in real time based on brain blood flow patterns. Lead researcher Alex Huth described: "If a participant hears 'I don't even have a driver's license,' the decoded version might become 'She hasn't even learned to drive yet' — not exact words, but the meaning, the semantic content." In 2025, another research group published Brain LLM, integrating brain recordings directly with large language models to generate natural language output from fMRI signals. It performed best on unexpected text, meaning brain signals provided information the AI didn't already have — the decoder was indeed "reading" something. Currently, this requires hours of training on a specific individual and cannot be used outside the lab. But the direction is clear: semantic decoders can be ported to functional near-infrared spectroscopy (fNIRS) — portable wearable sensors measuring the same underlying signals. Portable brain scanning is the direction. The neurotechnology market grew from $9.8 billion in 2022 to an estimated $17.1 billion by 2026. Major tech companies (Meta, Apple, Sony, Microsoft) are embedding neural sensors into headphones, wristbands, and AR headsets. The Neurorights Foundation white paper reviewed 30 neurotechnology companies and found that all 30 reserve broad rights to the neural data they collect. This data — about brain states, attention, emotional reactions, vulnerable moments, decision fatigue — legally belongs to the companies collecting it. As AI decoding capabilities improve, companies holding data will have resources never before available: a statistical model of how your brain works, from your brain, but owned by others. ## The Future of Connectomics and AI Connectomics research is not only about understanding humans but also about building better AI. Modern AI architectures (neural networks, transformers, etc.) were initially inspired by the brain, but that inspiration was a rough approximation. A complete connectome would provide an exact blueprint: how 86 billion neurons connect, which circuit structures give rise to memory, attention, decision-making, etc. — a blueprint refined over 400 million years of evolution. AI breakthroughs over the past decade came from scaling up existing architectures, but basic design principles have remained largely unchanged since the 1990s. A complete connectome could lead to new kinds of computers: using less energy, learning in new ways. Operating at 20 watts (less than a dim light bulb), the brain produces capabilities no AI has matched. Princeton researchers used mouse connectome data to build a digital twin — a simulation accurate enough to generate testable hypotheses that can be verified in the lab. One researcher said: "Once you can faithfully simulate a brain, the next question is whether the simulation could be conscious." He doesn't know the answer and thinks no one does. The people building this technology describe it cautiously: treating Alzheimer's, restoring speech for paralyzed patients, understanding neurological diseases — all real and beneficial. But the same tools, applied in portable form, could decode language from stroke patients who cannot speak, and will also be able to decode the language of anyone whose brain signals can be captured. The same mapping technique, if scaled sufficiently, could reveal the wiring behind epilepsy, and also the wiring behind memory, personality, preferences, and beliefs. ## The Ultimate Question of Privacy If AI can map every neuron in your brain and trace every connection, what privacy remains? What remains yours? This is no longer future tense — from the astonishing structures in half a grain of rice brain, to fMRI that can already decode thoughts, to portable sensors becoming widespread, every step edges closer to that question. Scientists studying neural privacy are not shy about this. Source: https://youtu.be/CdJjph8-2Oc