Kurzweil Scorecard: He Named the Right Problems. He Named the Wrong Labs.

Ray Kurzweil spent an unusual amount of The Singularity Is Near dropping names. Where most futurists gesture at “researchers at top universities,” he wrote down the Principal Investigator, the lab, and the instrument. The 2005 neuroscience passages read almost like a scouting report: Allan Snyder at Sydney, Andreas Nowatzyk at Carnegie Mellon, the Brain Networks Laboratory at Texas A&M, Wen-Biao Gan at NYU, Peter Dayan and Larry Abbott, Pentti Kanerva at Redwood.

Twenty years later the scorecard is striking in a specific way. The underlying phenomena he bet on mostly arrived. The people and labs he named mostly didn’t deliver them. Different teams, at different institutions, using different instruments, crossed the finish lines he drew.

The predictions

Batch 47 collects ten subneural and brain-scanning predictions from ~2005. All ten carry the same timeframe tag (circa 2005) — really a statement that the capability existed in prototype then and would scale from there. Kurzweil’s implicit thesis: the tools in hand by 2005 were enough to begin reverse-engineering a mammalian brain, and the researchers he named were the ones who would push them forward.

Full-text searches against 9.3 million US patent documents and 357 million OpenAlex works let us check who actually did the work, what shape it took, and whether it arrived. For the load-bearing claims, we read the actual papers.

Where we actually are

The savant effect did not survive the lab. Kurzweil wrote that “Allan Snyder reported that about 40 percent of TMS test subjects displayed significant new skills, including drawing abilities” (ch. “Peering into the Brain”). The source studies — Snyder et al. 2003 (proofreading, calendar calculation) and Snyder et al. 2006 (numerosity) — showed improvements, but later work acknowledged the protocols lacked active control stimulation sites and relied on small cohorts (12 participants in the 2006 numerosity paper). Chi & Snyder 2011 on insight problem solving tried tighter designs and found smaller, more selective effects. The 40% figure has not generalized. Later reviews concede that if such mechanisms exist, they are “not accessible to everyone.”

Meanwhile, repetitive TMS did become a mature clinical technology — for depression, not drawing. A 2023 Nature Mental Health meta-analysis of 27 years of sham-controlled rTMS trials documented a large and growing placebo response, which is its own problem but has nothing to do with savant-like skills. Verdict: wrong mechanism.

The nanometer mouse brain scanner arrived — from a different building. Kurzweil wrote that “Andreas Nowatzyk at Carnegie Mellon is scanning the mouse nervous system at less than 200 nanometer resolution” and that “Texas A&M’s Brain Tissue Scanner can scan an entire mouse brain at 250 nanometer resolution in one month using slices” (ch. “Peering into the Brain”). Both projects are essentially invisible in the 2020s literature. The Texas A&M Brain Networks Laboratory did not produce a widely cited whole-mouse-brain dataset.

And yet Kurzweil’s underlying prediction did come through. In April 2025 the MICrONS Consortium published the functional connectome of a cubic millimeter of mouse visual cortex: 84,000 neurons, more than 500 million synapses, 4 kilometers of axon, reconstructed from serial-section electron microscopy co-registered with two-photon calcium imaging of the same tissue. That volume is 1/500 of a full mouse brain. The raw data is 1.6 petabytes — roughly 22 years of HD video. 150 scientists at 22 institutions, led by the Allen Institute, Baylor, and Princeton, worked on it for nearly a decade. Nature Methods named electron-microscopy connectomics its Method of the Year for 2025.

The machinery for whole-brain scale is also here. ZEISS’s MultiSEM systems now image with 61 or 91 parallel electron beams. Argonne’s synchrotron X-ray imaging produced the first whole mouse brain at cellular resolution in 2024. None of this ran through Pittsburgh or College Station. Verdict for both predictions: right outcome, wrong lab.

Multielectrode recording was the one Kurzweil underestimated. He wrote that “multielectrode recording can simultaneously record activity from many neurons with submillisecond temporal resolution” (ch. “Improving Resolution”). In 2005 “many” meant a few dozen on a good day. Neuropixels 2.0, reported in Science in 2021 and now standard equipment, carries more than 5,000 sites on a single shank. Published protocols describe recording 6,144 sites across two probes in a single freely-moving mouse, with individual neurons tracked for more than two months. That is three to four orders of magnitude past what “many” meant in 2005. An intraoperative case series documented Neuropixels in 56 human patients. Patent US 11,850,416, granted December 2023, describes a manufacturing method for probe arrays of this class. Verdict: ahead of schedule.

Second-harmonic generation microscopy stayed exactly where Kurzweil left it. “SHG microscopy is a noninvasive technique able to study cells in action” (ch. “Improving Resolution”). True in 2005, true now. The core demonstrations — optical recording of action potentials with ~1 μm spatial and ~1 ms temporal resolution — came out between 2004 and 2008 and have been incrementally refined with new dyes like Ap3-SHG. But SHG has not displaced calcium imaging or voltage-indicator fluorescence as the workhorse for in-vivo functional neuroscience. It remains specialized, not dominant. Verdict: verified but narrow.

Optical coherence imaging didn’t become the brain tool. “Optical coherence imaging uses coherent light to create holographic three-dimensional images of cell clusters” (ch. “Improving Resolution”). Optical coherence tomography became the dominant structural imaging modality in ophthalmology — arguably the most successful clinical imaging translation of the 2000s. But the neuroscience version — holographic 3D imaging of cortical cell clusters in action — did not. Expansion microscopy, light-sheet fluorescence, and EM connectomics took that role. Expansion-microscopy-for-brain papers climbed from a trickle in 2015 to 17 in 2024. Verdict: overtaken by events.

Dayan & Abbott is still the textbook. “Peter Dayan and Larry Abbott summarized nonlinear differential equations describing neuron bodies, synapses, spike trains, and feedforward networks based on thousands of experiments” (ch. “Subneural Models”). Theoretical Neuroscience, published 2001, has more than 3,700 citations, remains in print from MIT Press, and the equations it codified are still the baseline 2024 papers extend. A Royal Society Phil. Trans. B review from 2023 on pyramidal neuron plasticity in behaving animals, and a 2024 Scientific Reports paper on fast learning in spiking networks, both build on the rate-coding and synaptic dynamics framework they assembled. Verdict: verified, still canonical.

The “soma learning” story turned into a dendrite story. Kurzweil wrote that “in vivo experiments show some brain responses are too fast for standard Hebbian synaptic or reverberatory learning, implying learning-induced changes in the soma” (ch. “Subneural Models”). The observation that some plasticity happens too fast for classical Hebbian mechanisms has survived. The location has not. Two decades of work reassigned that plasticity to dendrites and to ion-channel modifications in axons — what neuroscience now calls nonsynaptic plasticity. A 2024 bioRxiv paper on “few-shot pattern detection by transient boosting of somato-dendritic coupling” describes how voltage-gated calcium channels in dendritic compartments mediate ~100 ms calcium spikes whose plasticity rewrites somatic voltage over a seconds-long window — producing the fast behavioral-timescale learning Kurzweil gestured at, but through dendrites, not soma. Verdict: right phenomenon, wrong compartment.

Gan’s dendritic-spine bet held. Kurzweil wrote that “Wen-Biao Gan’s adult mouse visual-cortex studies indicate spine mechanisms can support long-term memory” (ch. “Subneural Models”). Gan’s 2002 Nature paper documented transcranial two-photon imaging of individual spines in adult primary visual cortex showing striking long-term stability. His 2009 Nature paper (“Stably maintained dendritic spines are associated with lifelong memories”) went further. The 2014 Science paper on sleep-dependent branch-specific spine formation after motor learning nailed the memory link. Over 18 months in adult barrel cortex, only 26% of spines were eliminated and 19% formed — consistent with the multi-decade persistence Kurzweil predicted. Verdict: verified — the cleanest hit in the batch.

The holographic-memory story went sideways — and landed in machine learning. “Memories appear to be dynamically distributed throughout neural regions in a hologram-like manner” (ch. “Subneural Models”). In biology this has not become the mainstream account; most long-term storage work points to structural changes at specific synapses (see Gan, above), not to a hologram-like distributed encoding.

But Kanerva’s mathematics did win — just not where Kurzweil placed it. Hyperdimensional computing, the successor framework Kanerva developed, is now active research with applications in neuromorphic chips, low-power edge AI, and language processing. Kurzweil himself pivoted in The Singularity Is Nearer (2024), describing 500-dimensional hyperdimensional language spaces as a basis for modern deep-learning architectures. The holographic-memory prediction landed, but in silicon. Verdict: wrong domain.

The scorecard

What Kurzweil missed (and what he nailed)

Two patterns pop out.

The first is that Kurzweil’s 2005 research landscape was a snapshot, not a trajectory. The specific labs he cited — Nowatzyk’s group at CMU, the Brain Networks Laboratory at Texas A&M, Snyder’s Centre for the Mind — were plausible contenders for the milestones he forecast. They did not finish. The milestones finished. Big scientific goals are usually indifferent to which specific group reaches them. If a target is real, someone hits it. Predicting the who is much harder than predicting the what — and doesn’t matter much for forecasting the shape of the future.

The second is that Kurzweil had a systematic bias toward locating fast learning in the cell body rather than in its processes. When in-vivo recordings showed plasticity faster than synaptic models could explain, the 2005 inference was that the soma itself must be learning. The actual answer was more interesting: nonsynaptic plasticity of ion channels, dendritic calcium spikes, somato-dendritic coupling. The direction was right. The geography was wrong. Both the soma-learning and holographic-memory predictions suggest Kurzweil read too quickly from “the 2005 model is insufficient” to “this specific alternative is the answer.” The right posture would have been “the mechanism is somewhere in this neighborhood.”

What he nailed: the multi-decade persistence of dendritic spines, the eventual arrival of whole-mouse-brain structural imaging, the continuing relevance of the Dayan-Abbott framework, and — most quietly — that Kanerva’s mathematics would turn out to matter, even if not for the reason he first gave.

Method note

For each prediction we ran full-text searches against local corpora (9.3 million US patent records, 357 million OpenAlex works), pulled year-by-year counts to establish trends, and read the specific papers anchoring each claim — Gan’s spine-imaging series, the MICrONS cubic-millimeter paper, the Neuropixels 2.0 paper, and the 2024 work on dendritic calcium-spike plasticity. Where Kurzweil’s 2005 text named specific labs, we searched for their subsequent publication records directly. Web searches filled in conference presentations, funding history, and the 2024–2025 news events that define the current state of the art.

Sources consulted:

• Nature (2025): Functional connectomics spanning multiple areas of mouse visual cortex
• Nature Methods (2025): Method of the Year: electron microscopy-based connectomics
• Science (2021): Neuropixels 2.0: a miniaturized high-density probe
• Nature (2009): Stably maintained dendritic spines are associated with lifelong memories
• Science (2014): Sleep promotes branch-specific formation of dendritic spines after learning
• Scientific Reports (2024): Fast learning without synaptic plasticity in spiking neural networks
• MIT Press: Theoretical Neuroscience (Dayan & Abbott)
• CNN (April 2025): World’s most detailed brain map constructed from speck of mouse tissue
• Allen Institute: Revealing the largest wiring diagram through MICrONS