Kurzweil Scorecard: The Connectome Kurzweil Said We Didn’t Need

In 1979, Francis Crick wrote that asking for “the exact wiring diagram for a cubic millimeter of brain tissue and the way all its neurons are firing” was asking for the impossible. In 2005, Ray Kurzweil agreed it was unnecessary. To reverse engineer a brain region, he argued, “it is sufficient to scan connections in a region well enough to understand their basic pattern rather than capture every single connection” (The Singularity Is Near, ch. “Scanning Using Nanobots”). Repetition and redundancy meant “higher-level models of regions are often simpler and sufficient without replicating every neural connection” (ch. “Analytic Versus Neuromorphic Modeling of the Brain”).

Forty-six years after Crick, in April 2025, the MICrONS consortium published ten papers in Nature that did the impossible thing nobody was supposed to need: a complete reconstruction of one cubic millimeter of mouse visual cortex containing 200,000 cells, 120,000 reconstructed neurons, four kilometers of axons, and 523 million synapses, co-registered with calcium imaging of around 75,000 of those neurons watching natural and synthetic videos. The dataset is 1.6 petabytes. Nature Methods named electron-microscopy connectomics its Method of the Year for 2025.

This batch of twelve Kurzweil predictions about the methods of brain reverse engineering scores well on tooling and infrastructure — and badly on the conceptual bet that you could skip the hard work.

The predictions

Kurzweil’s reverse-engineering chapters in 2005 made three kinds of claims. First, scanning resolution would compound exponentially. Second, databases would catalog the results fast enough to matter. Third — most consequentially — the brain’s repetition and limited cell-type diversity meant we would not need to map everything. Pattern-level scans, plus models of a finite menu of neuron types, would be enough.

In The Singularity Is Nearer (2024), Kurzweil quietly conceded ground on the first claim: “The trade-off between spatial and temporal resolution in brain scans is one of the central challenges in neuroscience as of 2023. These limitations stem from the fundamental physics of blood flow and electricity… they probably won’t be sufficient to allow a sophisticated brain-computer interface.” That is a notable retreat from the 12-month doubling curve in the original book.

Where we actually are

The connectome bet went the other way. MICrONS is a specific rebuke of Kurzweil’s “basic pattern, not every connection” position. So is the FlyWire whole-brain Drosophila connectome (Schlegel et al., 2024, 165 citations), which mapped every neuron and synapse in an adult fly. So is the H01 human cortex sample from Lichtman and Google. Cell-type-specific inhibitory circuitry from a connectomic census of mouse visual cortex (Schneider-Mizell et al., 2023) used the MICrONS reconstruction to derive wiring rules invisible at a “basic pattern” scan: cell-type-specific inhibitory motifs that vary across the column. Without every synapse on the table, the rule did not exist. Kurzweil’s optimism that brain regions would be “simpler” was the wrong direction of the surprise.

Cell-type taxonomy went the same way. Kurzweil predicted “a brain region may contain billions of neurons but only a limited number of neuron types, making combined wiring and cell-type modeling feasible” (ch. “Scanning Using Nanobots”). In December 2023 the BRAIN Initiative Cell Census Network published a hierarchical atlas of the entire mouse brain: 34 classes, 338 subclasses, 1,201 supertypes, and 5,322 transcriptomically distinct clusters, derived from roughly 7 million single-cell RNA-seq profiles plus 4.3 million MERFISH-imaged cells (Yao et al., Nature 2023). “Limited” is now a generous reading. The good news is that the taxonomy is hierarchical, so high-level modeling is possible; the bad news is that distinguishing among 5,322 leaves matters for what circuits actually do, as the connectomic-census papers above demonstrate.

Recording technology blew past Pinkel. Kurzweil predicted that “a future version of Pinkel’s system will image up to 1,000 simultaneous cells at distances up to 150 microns with submillisecond resolution” (ch. “Improving Resolution”) sometime in the 2010s. Two-photon mesoscopes now image tens of thousands of cells across millimeter-scale fields of view; the MICrONS calcium dataset alone covers around 75,000 neurons in a single awake animal. On the electrode side, the Neuropixels 2.0 probe (Steinmetz et al., Science 2021, 1,038 citations) has 5,120 recording sites distributed across four shanks, and a 2025 Neuron paper introduced Neuropixels Ultra with 5×5 µm sites on a 1 µm grid for single-process resolution. The first human Neuropixels recordings (Paulk et al., Nature Neuroscience 2022, 260 citations) demonstrated single-unit yields the original Pinkel paradigm could not approach.

Cerebellar basis functions held up. Kurzweil described cerebellar circuits as “learn[ing] and apply[ing] mathematical basis functions that directly transform perceived movement into required muscle movement for tasks like catching a fly ball” (ch. “A Neuromorphic Model: The Cerebellum”). Two decades on, this is the dominant computational picture. The forward-model account is reviewed in Tanaka et al., Frontiers in Systems Neuroscience 2020. Cerebellar granule cells expanding their inputs into a high-dimensional code — the substrate Pouget’s framework requires — is now the operating hypothesis. The “fly ball” prediction got the mechanism roughly right.

Noninvasive scanning is impressive but not Moore-curve. A 12-month doubling cadence over 20 years would imply a million-fold improvement since 2005. Actual progress is closer to two orders of magnitude on specific axes. Iseult, the world’s first 11.7-tesla human MRI, produced in-vivo brain images in 2024 (Nature Methods) at sub-200-micron mesoscale resolution. Functional ultrasound now images deep cortex in awake primates at sub-100-µm resolution (Norman et al., PNAS 2020). Real advances; not a doubling per year.

The deep visual system is being mapped, slowly. MICrONS spans primary visual cortex plus several higher visual areas in the same animal. The Allen mouse visual hierarchy is functionally annotated; the Human Connectome Project’s 7T retinotopy parcellated the human visual hierarchy. Parietal and temporal extensions remain poorly understood, but the trajectory matches Kurzweil’s prediction.

Spindle-cell modeling is too early to call. Kurzweil wrote that “reverse engineering the exact methods of spindle cells and high-level emotions will require better models of the many brain regions to which they connect” (ch. “Understanding Higher-Level Functions”). Hodge et al. (2019) gave von Economo neurons their first transcriptomic definition: regionally specialized extratelencephalic-projecting excitatory neurons in frontoinsula and anterior cingulate. That is a definition, not a model. The “high-level emotion” reverse engineering Kurzweil was pointing at is not happening yet.

Plasticity, databases, frozen-tissue scanning. The Hubel-Wiesel 1965 reorganization claim is straightforwardly verified historically. The “extensive databases methodically cataloging” claim is dramatically on-track: the Allen Brain Cell Atlas, BossDB, DANDI, NeMO, EBRAINS, and the BICCN data ecosystem now host petabyte-scale neural datasets with active programmatic access. And destructive scanning of frozen brain tissue — the very method Kurzweil flagged as feasible-but-not-yet-fast-enough — is now the workflow that produced MICrONS at 1.6 PB.

Meaningful understanding by the mid-2020s? “Within about two decades of 2005, it will be possible to formulate a meaningful understanding of brain function” (ch. “Achieving the Software of Human Intelligence”). It depends what “meaningful” means. We have a complete fly connectome, a cubic millimeter of cortex mapped at synapse resolution, a 5,322-cluster cell-type atlas, large-scale electrophysiology, and computational models that can decode movement intent from M1. We do not have a working theory of cortical computation. Most of the field would say no — not yet.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern across these twelve predictions is asymmetric: Kurzweil was right about what tools would exist and wrong about how necessary they would be. The 2005 chapters framed full connectomics as brute-force enumeration that smart scientists would route around. The smart scientists did not route around it. They built the enumeration, found that the brute force exposed wiring rules invisible at coarser scales, and used those rules to constrain models the field still cannot build without them.

The cell-type story is the same shape. Kurzweil bet on a small menu of canonical types, reasoning from the cortical-column literature of the 1990s. The transcriptomic revolution turned that menu into a hierarchy with thousands of leaves that participate in stereotyped circuit roles. “Limited” was the wrong adjective. “Combinatorially structured” is the right one, and that distinction changes what kind of model you have to write.

The forecasting lesson keeps recurring in this scorecard series. Kurzweil’s exponential intuitions price capability curves well and theoretical curves poorly. Tools scale; understanding does not. The field he expected to skip every-connection mapping spent the last decade doing exactly that and is calling it Method of the Year.

Method note

The patent and biomedical literature corpora used here cover roughly 9.3 million U.S. patent documents and 357 million scholarly works. Counts above are restricted to the most-cited papers in each topic since 2018. Specific patent and paper numbers, and headline metrics from the MICrONS, BICCN, FlyWire and Iseult projects, were verified against publisher pages and institutional press releases retrieved during this session. Each prediction is paraphrased from The Singularity Is Near with chapter citation; quoted updates are from The Singularity Is Nearer (2024).