Kurzweil Scorecard: Reading the Brain — Right Tool, Wrong Wire

The fly brain has been mapped. Every neuron, every synapse, in the head of a Drosophila melanogaster. A team led by Princeton’s Sebastian Seung published the full wiring diagram in October 2024: 139,255 neurons, roughly 50 million synaptic connections, every cell typed and annotated, every neurotransmitter inferred. Six months later, in April 2025, the MICrONS consortium released the first cubic millimeter of mouse visual cortex at the same level of detail — 200,000 cells and 523 million synapses, co-registered with calcium imaging of 75,000 of those neurons watching natural movies. Nature Methods declared electron-microscopy connectomics its Method of the Year for 2025.

Ray Kurzweil predicted this in 2005. He also predicted, in the same paragraphs, that we’d be reading individual fibers in the optic nerve by the 2020s, that fMRI would resolve millisecond cognition, and that the spindle cells in the anterior cingulate were a uniquely human (and great-ape) signature of high-level emotion. Two of those four predictions were spectacular hits. Two were wrong in instructive ways.

The predictions

This batch is five claims and forecasts from The Singularity Is Near, all about the project Kurzweil called “reverse engineering the brain.” They cluster into three questions:

• Is the technology of seeing the brain — at synapse, layer, and column scales — improving the way Kurzweil said it would?
• Is the biology he leaned on (spindle cells as a human signature; optic nerve as the obvious read point) holding up?
• Is the project he wrapped them into — simulate the brain by 2030 — on track?

Where we actually are

The connectome is being delivered, on a steeper curve than Kurzweil wrote. Kurzweil’s 2005 line was that “humanity’s ability to reverse engineer the brain by seeing inside it, modeling it, and simulating its regions is growing exponentially” (ch. “Reverse Engineering the Brain: An Overview of the Task”). This held. The C. elegans connectome (302 neurons, 7,000 synapses) was the only complete map when he wrote it. The fly is now done. The mouse cubic millimeter is done. Harvard’s Lichtman lab and Aravinthan Samuel published SmartEM in 2025 — an AI-guided electron microscope that adapts its dwell time to local complexity and cuts imaging time up to sevenfold without losing accuracy. The literature reflects the curve: in our index, citation-weighted papers on whole-brain Drosophila connectomics jumped from 3 in 2018 to 20 in 2025 and 11 already in early 2026, with the flagship Nature paper at 447 citations in eighteen months.

The patent layer is keeping pace, in the upstream tools. US 12,456,600, granted October 2025, claims a scanning-electron-microscopy tomography system that takes sinograms by projecting electron beams across multiple directions and reconstructs three-dimensional specimen volumes — the kind of pipeline that makes million-section reconstructions tractable. US 12,546,688, granted February 2026, claims modified expansion microscopy (“mExM”) with reproducible 4-fold and 12-fold tissue expansion using a sealable mold and an SDS-based protein digestion buffer. Expansion microscopy lets a conventional light microscope resolve sub-diffraction-limit structures by physically swelling the tissue. The literature on expansion microscopy of synapses went from 7 papers in 2018 to 41 in 2021 and is still publishing 18+ per year. Kurzweil predicted scanning resolution would double every year. The actual gain has been bursty — but cumulatively far past what was visible in 2005.

Real-time synapse-resolution observation is still a proxy game. Kurzweil wrote that “the emerging generation of scanning tools will for the first time observe individual dendrites, spines, and synapses performing in real time” (ch. “The Accelerating Pace of Reverse Engineering the Brain”), with the phrasing pointing toward the 2010s. Today’s two-photon calcium imaging does watch individual dendritic spines in living tissue, and genetically encoded voltage indicators are getting close to single-spike resolution at single-synapse spatial scale. But “performing in real time” still means watching a fluorescence proxy, not the actual molecular traffic at the cleft. And the tools that do see synapses with full structural fidelity — serial-section EM and expansion microscopy — kill the tissue first. The “real-time at synapse resolution” prediction stretches but doesn’t quite land.

The fMRI claim was half-right, and Kurzweil himself walked it back. In 2005 he asserted as fact that “recent fMRI advances can map columnar and laminar structures a fraction of a millimeter wide and detect tasks occurring in tens of milliseconds” (ch. “Peering into the Brain”). The spatial half is now decisively true: 7T scanners routinely deliver 0.35–0.45 mm isotropic laminar fMRI in human cortex, and recent multi-contrast methods hit 0.1 mm in either the laminar or columnar direction. But the temporal half was wrong in 2005 and remains wrong in 2026. The fundamental physics of the BOLD signal — measuring blood flow as a proxy for neural firing — caps fMRI’s temporal resolution at hundreds of milliseconds, not tens. Kurzweil acknowledged this himself in The Singularity Is Nearer (2024), writing that fMRI activity “can rarely be better than 400 to 800 milliseconds” and that “these limitations stem from the fundamental physics of blood flow and electricity, respectively, so even though we may see marginal improvements from AI and improved sensor technology, they probably won’t be sufficient to allow a sophisticated brain–computer interface.” That’s a clean self-correction nineteen years on.

The spindle-cell story was wrong, in a way that matters for his theory of emotion. Kurzweil leaned on spindle cells to argue that “emotionally charged situations appear to be handled by spindle cells, which are found only in humans and some great apes and are deeply interconnected through long apical dendrites” (ch. “Understanding Higher-Level Functions: Imitation, Prediction, and Emotion”). The exclusivity claim has not survived. Von Economo neurons — the formal name for spindle cells — have since been documented in four cetacean species (bottlenose dolphin, Risso’s dolphin, beluga, humpback whale), in African and Indian elephants, and in macaques, domestic sheep, cows, pygmy hippopotamus, and white-tailed deer. The current consensus is that they are a convergent adaptation associated with large brains needing fast long-range integration, not a hominid-emotion signature. The architectural fact (long apical dendrites, deep interconnectedness) survives. The phylogenetic specialness does not.

The optic-nerve prediction was overtaken by a different mechanism entirely. Kurzweil’s by-2020s prediction was that “once we can monitor each discrete fiber in the optic nerve, our ability to interpret optic-nerve coding will be greatly facilitated” (ch. “Scanning Using Nanobots”), citing Tomaso Poggio’s program at MIT. We are not monitoring individual optic-nerve fibers in humans. The roughly one million axons in the human optic nerve remain a black box at the per-fiber level. But the outcome — “interpret what the visual system is computing” — happened anyway, by reading the next station instead. Takagi and Nishimoto’s CVPR 2023 work and the MindEye2 paper at ICML 2024 both reconstruct what subjects are seeing from 7T fMRI of visual cortex. MindEye2 needs only one hour of training data per new subject and fine-tunes Stable Diffusion XL to accept neural latents in place of text prompts. The patent record echoes the shift: US 12,437,187, granted October 2025, claims a generative-model pipeline (including GANs) that reconstructs MRI images from undersampled scans without introducing artifacts. The decoder is no longer the optic nerve; it’s the BOLD signal in V1–V4, processed by the same diffusion models that draw cats from prose. Right destination, wrong wire.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern across this batch is sharp: Kurzweil was right about the curves, wrong about the wires.

He correctly forecast that scanning, segmentation, and modeling would compound. The connectome cadence — worm to fly to mouse-cubic-mm in roughly the period he predicted — is the single best vindication of his exponential argument anywhere in neuroscience. He correctly forecast that AI would be the lever, though he could not yet have named the diffusion models that ended up doing the heavy lifting on the decoding side.

He was wrong wherever he committed to a specific mechanism. Optic-nerve monitoring did not arrive; cortical decoding did. Synapse-resolution real-time observation did not arrive; calcium proxies and connectomic snapshots did. fMRI did not break the BOLD temporal floor; it never could. Spindle cells did not stay human; they kept turning up in whatever animal grew a big brain.

The lesson for the people who fund and build at the edge of brain science: track the curves but stay agnostic about which probe wins. The intermediate target Kurzweil described — “interpret optic-nerve coding” — was a 2005 best guess about where the neural signal would be most legibly read. The signal turned out to be more legible three synapses downstream, in cortex, as long as you had a generative model on the other side of the regression. A roadmap that fixated on the optic nerve would have missed the moment.

Method note

For each prediction we counted neuroscience and engineering papers in our literature index by year, filtered to high-citation work, and pulled the highest-impact items by hand. We named specific patents granted between 2025 and February 2026 from a U.S. patent corpus to anchor the upstream tooling. We supplemented with web searches for FlyWire, MICrONS, SmartEM, layer fMRI, von Economo neurons, and fMRI-to-image reconstruction, and read passages from The Singularity Is Near (2005) and The Singularity Is Nearer (2024) to ground each verdict in Kurzweil’s own words. Verdicts reflect the state of evidence as of May 2026.