Kurzweil Scorecard: The Brain-Wide Map Arrived, Right Inside the Window He Gave Us

In 2005, buried in the chapter on reverse-engineering the human brain, Ray Kurzweil made a quiet wager. “New global brain-observation methods should allow discovery of the overall programs behind the brain’s powers” — and he penciled the arrival in “by the 2020s” (The Singularity Is Near, ch. “Achieving the Software of Human Intelligence: How to Reverse Engineer the Human Brain”). It was the kind of line you skimmed past on the way to the nanobot bloodstream chapter. Most of his brain predictions were descriptive — claims about what 2005 neuroscience already knew. This one was load-bearing. If whole-brain observation didn’t arrive, the rest of his brain-uploading roadmap collapsed.

On September 3, 2025, the International Brain Laboratory — a 12-lab consortium spanning the US and Europe — published two Nature papers describing simultaneous cellular-resolution recordings from 621,733 neurons across 279 brain areas in 139 mice, all running the same decision-making task. The headline unit of analysis in the main paper is 75,708 quality-controlled neurons. This is the “global brain-observation method” Kurzweil described, delivered with roughly four years to spare in his decade-long window.

So that’s the flagship hit. The rest of the batch — a grab-bag of claims about brain architecture, the insula, two-photon microscopy, and the optic nerve — tells a messier story. Some of his 2005 numbers held up. Some got upgraded by an order of magnitude, in the direction he didn’t expect.

The predictions

Batch 22 in our tracker is twelve neuroscience claims from The Singularity Is Near, scattered across the chapters on peering into the brain, improving resolution, the visual system, and reverse-engineering. Most are descriptive statements about 2005 knowledge, but they are also load-bearing for the larger argument: if the brain is modular, if regions are simpler to model than single neurons, if we can peer in with enough resolution, then whole-brain emulation follows. Twenty-one years on, we can ask how each of those enabling assumptions has aged.

Where we actually are

Global brain observation — delivered on time. The International Brain Laboratory’s brain-wide map (doi: 10.1038/s41586-025-09235-0) used 699 Neuropixels probes across 139 mice performing an identical visual-decision task. The dataset covers 95 percent of the mouse brain volume. The two standout findings directly contradict a narrow-localist reading of 2005 neuroscience: decision-making signals are distributed across dozens of regions with constant cross-area communication, and prior expectations — learned statistics about the task — appear encoded even in early sensory areas like the thalamus. This is more than confirmation that you can point the instruments at the whole brain. It is the beginning of the program discovery Kurzweil was betting on. Verdict: On track, inside the predicted window.

The retina carries four times more channels than he said. Kurzweil cited Frank Werblin and Botond Roska’s 2001 Nature paper claiming the optic nerve carries roughly “10 to 12 output channels — edges, uniform color areas, backgrounds” (ch. “The Visual System”). This was the canonical number for years. It is now obsolete. Sanes and collaborators’ 2019 Neuron atlas (doi: 10.1016/j.neuron.2019.11.006, 610 citations) used single-cell RNA-seq to resolve 46 molecularly distinct retinal ganglion cell types in the adult mouse retina. A follow-up in 2020 added 63 amacrine cell types downstream of them. The retinal code is four to five times richer than Werblin and Roska could see with their 2001 methods — and richer in a way that only became visible once transcriptomics replaced electrophysiology as the sorting tool. Verdict: Behind schedule on the specific number — the real count was already there in the tissue, hidden from the tools Kurzweil trusted.

The visual system no longer lags the auditory system. Kurzweil wrote in 2005 that “preliminary models exist only for areas V1 and MT, with 36 other visual areas still needing study” (ch. “The Visual System”). The framing was that vision was harder and further behind auditory processing. Two decades on, this claim has quietly inverted. The Glasser 2016 multi-modal parcellation (doi: 10.1038/nature18933, Nature) delineated 180 cortical areas per hemisphere — many of them in visual cortex — using Human Connectome Project data across 210 healthy adults. The mouse visual cortex has been mapped at cellular resolution with two-photon calcium imaging across millions of neurons. Deep convolutional networks trained on ImageNet now serve as the best predictive models of primate V1 and V4 responses; no comparable mechanistic model exists for auditory cortex. Vision is the most-modeled sensory modality, not the least. Verdict: Overtaken by events.

The “material me” held up. Bud Craig’s pathway from Lamina 1 neurons through the posterior ventromedial thalamic nucleus to the bilateral insula, constructing a moment-to-moment representation of body state (ch. “Understanding Higher-Level Functions: Imitation, Prediction, and Emotion”), has aged well. The 2019 review “The Organization of the Primate Insular Cortex” (doi: 10.3389/fnana.2019.00043, 227 citations) gives the Craig model a central place and extends it — anterior insula is now treated as the convergence zone for interoceptive representations and affective awareness, tightly linked to the salience network. Verdict: On track. Related: Kurzweil also noted “frontoinsular cortex, containing spindle cells, is especially active in fMRI during high-level emotions such as love, anger, sadness, and sexual desire” (same chapter). The Von Economo neurons are real, and their degeneration shows up in schizophrenia (doi: 10.1002/ar.23635, 2017) and in monkey insula evolution studies (doi: 10.1016/j.cobeha.2018.05.006). The specific 2005 emotion list was too neat, but the underlying cell class and its role in social-affective processing are well supported. Verdict: On track with nuance.

Two-photon laser scanning microscopy is still the workhorse, and getting exotic. Kurzweil described two-photon microscopy in 2005 as capable of “detecting excitation of single synapses in the intact brain” and of ultraprecise surgery — “severing a connection or destroying a single mitochondrion” (ch. “Improving Resolution”). Routine now. The instrumentation push continues: US 12,313,549 (May 2025) covers entangled two-photon absorption microscopy, tuning the time delay between quantum-entangled photon pairs to extract signal at excitation intensities far below classical limits — longer imaging in living tissue without photodamage. US 12,345,643 (July 2025) describes a dual-wavelength two-photon microscope with separate laser sources for short- and long-wavelength excitation. The underlying technique is still two-photon; what has changed is the physics stacked on top of it. Verdict: On track.

Optical imaging has gone whole-brain, and mostly still in animals. Kurzweil’s 2005 sketch described fluorescing-dye optical imaging as “used mainly in animal experiments”. It still is — and at scales he didn’t anticipate. Our literature database surfaces 8,672 papers tagged with expansion microscopy, and light-sheet microscopy is now routinely used to image whole cleared mouse brains (doi: 10.1177/0271678×17698970; doi: 10.1073/pnas.2218617120). The “Cortical column and whole-brain imaging with molecular contrast and nanoscale resolution” Science paper from 2019 (392 citations) showed that expansion microscopy combined with light-sheet acquisition can deliver nanoscale-resolution reconstruction of entire cortical columns. Verdict: Ahead of schedule.

Leif Pinkel’s lab, specifically named. Kurzweil cited “Leif H. Pinkel’s University of Pennsylvania Neuroengineering Research Laboratory” developing an optical system for imaging individual neurons at one-millisecond resolution (ch. “Improving Resolution”). That specific lab is no longer a visible presence in the field, and Penn’s current neuroengineering centers (Litt Lab, Center for Neuroengineering & Therapeutics) have different focuses. But the goal — imaging individual neurons at millisecond resolution in vivo — was achieved many times over through a different path: genetically encoded voltage indicators paired with high-speed two-photon or light-field microscopy. Cell published a 2025 paper on imaging high-frequency voltage dynamics in multiple neuron classes in behaving mammals (doi S0092-8674(25)00730-5). The destination arrived without the specific lab that was supposed to bring us there. Verdict: Wrong mechanism.

“Several hundred regions” and “100 trillion connections” held up on the order of magnitude. Kurzweil’s architectural claims — roughly several hundred brain regions, 100 trillion synapses, massively parallel processing (ch. “Is the Human Brain Different from a Computer?”) — have survived the last two decades largely intact. Glasser’s 180 cortical areas per hemisphere puts us close to 360 cortical regions, with dozens more subcortical. Recent connectomics efforts (H01 human temporal lobe, MICrONS mouse visual cortex) reconstructed synaptic connectivity at the cubic-millimeter scale and project whole-brain synapse counts consistent with the 10^14 figure. Verdict: On track.

“Brain region design is simpler than neuron design.” This was the most theoretically load-bearing claim in the batch — the bet that you could abstract away single-neuron biophysics and model a region with something like a deep network. It is partially vindicated (the pattern-recognition work of the last decade uses exactly this abstraction) and partially contradicted (the IBL 2025 paper’s distributed-decision finding suggests regions aren’t clean functional units you can model independently). Large language models and vision deep networks arguably succeeded because they bypassed single-neuron dynamics — but they also bypassed the regional decomposition Kurzweil pointed to. Verdict: Split. Right that abstraction beats biophysical detail; wrong that the natural abstraction is at the region boundary.

“Evolution reuses brain modules and themes.” The minicolumn and cortical microcircuit literature continues to support a reuse thesis — roughly 200 million minicolumns, each ~100 neurons, a common six-layer plan repeated across cortex. The newer wrinkle is that the software running on those modules appears less swappable than the hardware. Verdict: On track.

The scorecard

What Kurzweil got right, and what he missed

The pattern across these twelve claims is consistent with the batches we’ve already scored on longevity, nanotechnology, and computing hardware: Kurzweil’s orders of magnitude tend to hold up, his mechanisms often don’t, and his instruments get replaced by something better before the finish line. He said we would image individual neurons at millisecond resolution and we can — but through voltage indicators, not through the lab he named. He said the brain has several hundred regions and ~100 trillion synapses and that holds to the nearest decade. He said the optic nerve has 10–12 channels and the real number is 46.

The most interesting miss is the retina. Werblin and Roska’s 2001 paper was state-of-the-art electrophysiology — recording from ganglion cell axons and clustering by response type. The jump from 12 channels to 46 required a different instrument entirely: single-cell RNA-seq, which lets you sort neurons by gene expression before asking what they do. Kurzweil’s forecasting method trusted the dominant technique of the moment to extrapolate. When that technique has a hidden ceiling, the prediction inherits the ceiling. This is the pattern to watch for in the later batches that depend on specific 2005 numbers.

The biggest hit is quiet. The “by 2020s” global brain observation prediction did not have a named champion lab, a specific instrument, or a falsifiable number attached. It was just a window on the timeline. Twenty years later, the IBL consortium — itself inspired by how high-energy physics and genomics are done — put down 699 probes across 12 labs and walked out with a map that covers 95 percent of the mouse brain. Kurzweil’s unfalsifiable claim turned out to be the falsifiable one that landed on schedule.

Method note

We pulled patent and literature counts from our indexed databases of 9.3 million US patents and 357 million OpenAlex scholarly works, and read the actual claims of recent two-photon microscopy and connectomics patents rather than just counting them. Every Nature paper cited here was either surfaced by targeted keyword search of paper titles and abstracts or verified by a DOI lookup against the original publisher. Patent numbers were confirmed against US grant dates in our local index. Web searches filled in the International Brain Laboratory’s 2025 publication details, including the 621,733-neuron and 279-area figures, which came from the Nature paper and associated press release.