Kurzweil Scorecard: The Brain-Scanning Wall Kurzweil Won’t Admit Out Loud

Ten claims from 2005 about how the brain works, and how quickly we’d be able to see it. Nine were ordinary neuroscience — attention modulating visual cortex, synaptic scaling stabilizing cultured networks, object recognition finishing in about 150 milliseconds. Those mostly held, and some were vindicated more thoroughly than Kurzweil argued. The tenth was different. He wrote that “the temporal and spatial resolution and bandwidth of human brain scanning are doubling each year” (ch. “Is the Human Brain Different from a Computer?”). That extrapolation was the load-bearing girder for the mind-uploading arc of the book. Twenty-one years later it is sitting on the floor.

The predictions

Chapter 4 of The Singularity Is Near made two kinds of statements about neuroscience. Descriptive facts about the brain’s machinery: neurons can multiply and filter signals; attention modulates visual cortex; synaptic scaling in cultured networks keeps potentials from collapsing to zero; simulating a human brain at neuron-and-synapse detail would need roughly 10^19 calculations per second as an upper bound. And trendlines — scanning resolution doubling annually — that set up the noninvasive read-out mind uploading would require. Grading the descriptive claims takes a decade of neuroscience. Grading the trendline takes a ruler.

Where we actually are

The scanning trendline broke against the physics of blood and skulls. Kurzweil concedes this in The Singularity Is Nearer (2024), even as he moves past it. He writes that fMRI voxels are “about 0.7 to 0.8 millimeters to a side” and that the blood-flow lag means brain activity “can rarely be better than 400 to 800 milliseconds” temporal resolution (ch. “Extending the Neocortex into the Cloud”). He then says, without flagging the retreat: “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, 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 is not a doubling-every-year trajectory. That is a wall. The 2005 fMRI voxel was about 1 millimeter; the 2025 7T voxel in research settings is about 0.8 millimeter, with specialized sequences reaching sub-millimeter cortical laminar imaging. Two doublings in twenty years at the extreme, not twenty. A 2025 patent (US 12,474,425), for “highly accelerated sub-millimeter resolution 3D GRASE” at 7 tesla, shows the direction of travel: tight engineering trade-offs around SNR and T2 blurring to squeeze another increment out of a modality that fundamentally measures blood, not neurons.

Invasive connectomics went the other way entirely. While the noninvasive trendline flattened, something Kurzweil did not predict exploded. In April 2025 the MICrONS consortium published a suite of ten papers in Nature reconstructing a cubic millimeter of mouse visual cortex: more than 200,000 cells, roughly 0.5 billion synapses detected by automated segmentation, co-registered with calcium imaging of 75,000 neurons responding to natural and synthetic stimuli. Nature Methods named electron-microscopy connectomics Method of the Year 2025. It is a different modality answering a different question — structure and wiring, not live thought — but it is the actual frontier of “seeing the brain,” and it required cutting the brain apart and imaging it with electrons.

The core neuroscience claims aged well, and in two cases were extended past what Kurzweil argued. The claim that “biological neurons can perform computations including subtracting, multiplying, averaging, filtering, normalizing, and thresholding signals” (ch. “Trying to Understand Our Own Thinking”) understated what has since been demonstrated. Gidon and colleagues, publishing in Science in 2020, recorded a new class of calcium-mediated dendritic action potentials in human layer 2/3 pyramidal cells that let a single neuron compute XOR — a linearly non-separable operation that textbook artificial networks need at least two layers to solve. A single human pyramidal neuron is not the point-neuron Kurzweil sketched. It is more like a small multilayer network hidden inside a cell body and its dendritic tree.

Attention modulating visual cortex and synaptic scaling in cultured networks were both thoroughly verified and extended. A 2017 Science paper by Diering and colleagues (524 citations) found that Homer1a drives homeostatic scaling-down of excitatory synapses during sleep — taking synaptic scaling out of the dish and into the behaving animal, tied to the sleep-wake cycle. Hengen’s 2013 Neuron paper (383 citations) pushed the same mechanism into mouse visual cortex in vivo. Kurzweil’s 2005 footnote became two decades of vindicating experiments.

The 150-millisecond feedforward claim held, with a footnote. Kurzweil wrote that “MEG studies and macaque inferotemporal-cell latencies indicate early visual recognition is largely feedforward and takes about 150 milliseconds, leaving little time for feedback” (ch. “The Visual System”), citing Poggio and Riesenhuber. Kar, DiCarlo, and colleagues followed up in Nature Neuroscience (2019): for easy images, feedforward computation does finish by about 150 ms, but a subset of “challenge images” that humans still recognize produce a decodable IT representation only about 30 ms later, once recurrent processing engages. Kurzweil’s sentence was right about the dominant regime and slightly overclaimed the absence of feedback — a qualifier that matters, since feedforward-only networks match primates on ordinary images but fail the hard cases recurrent circuits solve in those extra 30 ms.

The 10^19 cps upper bound survives, and has been quietly retired. The 2005 claim that simulating a brain at neuron-and-synapse detail would need about 10^19 operations per second was the upper edge of a conservative bracket. In The Singularity Is Nearer, Kurzweil revises sharply downward: a firing-rate model can do the job at about 10^14 ops per second; the AI Impacts team’s energy-based estimate from average neuron firing rates (0.29 Hz, not 200 Hz) puts it as low as 10^13. The upper bound was not wrong; it was replaced by a more confident lower one. Frontier, the 2023 Oak Ridge supercomputer, does roughly 10^18 — four orders of magnitude above the revised brain estimate. The hardware arrived. The neural model to run on it did not.

The prediction that synaptic mechanisms would improve AI learning and stability is the one outright miss. Kurzweil expected the path from biological neuroscience to artificial intelligence to run through more faithful models of synapses and spines — a direct lineage from the Subneural Models chapter to better algorithms. It did not. The AI that crossed human-parity benchmarks ran on attention layers and positional encodings, not synaptic scaling or dendritic Ca spikes. Neuromorphic hardware patents did appear — US 12,468,927 (November 2025) describes a synapse-array architecture that expands by connecting blocks between neuron arrays; a DYNAP-SE2 paper from 2024 described a multi-core spiking processor with per-synapse plasticity rules — but they occupy specialty accelerator niches, not the training clusters that made the past five years of AI. Kurzweil got the right outcome (machines that learn) via the wrong mechanism.

The scorecard

What Kurzweil missed (and what he nailed)

Kurzweil was a careful reader of 2005 neuroscience. The descriptive claims — about attention, recognition timing, neuronal computation, synaptic scaling — are all still standing, and two have been extended further than he argued. The problem is what he built on top.

He assumed the tools for watching the brain would improve at the pace of the tools for computing — Moore’s Law generalized to scanners. They have not. The physics that limits fMRI is not the physics that limits transistors. Blood flow lags, skulls attenuate electric fields, and photons do not travel through bone. In the 2024 book he acknowledges, politely, that noninvasive scanning has probably gone as far as marginal engineering can take it. The frontier moved sideways into invasive electron microscopy, where a 2025 Nature Methods award sits for a cubic millimeter of mouse — 0.00007 percent of a human brain.

He also assumed AI would inherit from biology. It didn’t. Attention-layer networks are not especially brain-like, and they won anyway, because systems with the right inductive bias that learn quickly on lots of data tend to beat biologically-inspired systems with the wrong bias that learn slowly.

What survives is a field full of confirmed middle-scale facts about neurons, and a broken bridge to the mind-uploading story that needed scanning-by-extrapolation to work. The bridge has been replaced by a different one — electrodes and sliced tissue — that runs to a different destination.

Method note

The checks here pulled from an internal 9.3-million-record patent corpus and a 357-million-record scientific literature corpus to find the most-cited recent work on dendritic computation, synaptic scaling, connectomics, 7T fMRI, and neuromorphic hardware. Paper titles, citation counts, patent numbers, and patent abstracts cited above were read directly from those sources. Singularity Is Near references use chapter titles and paraphrases from a prediction catalog extracted from the 2005 book; Singularity Is Nearer quotations are from the 2024 book’s own text. Web searches surfaced the April 2025 MICrONS papers and the Nature Methods 2025 Method of the Year selection. Verdicts are our own reading of the evidence.