In 2005, Ray Kurzweil filled a chapter of The Singularity Is Near with a parade of confident neuroscience claims, citing them not as predictions but as the load-bearing facts on which his timetable for reverse-engineering the brain rested. Twenty years on, half of those claims have been quietly upgraded by better counting methods, half have been downgraded by better experiments, and one of them — the unsexy idea that the cortex is mostly a prediction machine — has become the dominant story in computational neuroscience.

This batch is unusual. There are no dates, no “by 2010” promises. Just ten anatomical and theoretical assertions Kurzweil treated as background scaffolding. The interesting question is which pieces of scaffolding are still standing.

What Kurzweil leaned on

Kurzweil grouped these claims into three pillars: a neuromorphic model of the cerebellum, a neuromorphic model of the auditory and visual systems, and a “higher-level functions” section about imitation, mirror neurons, recursion, and prediction. Each pillar was meant to show that the brain’s organizing principles were within reach of engineering. The Singularity Is Nearer (2024) returns to the same ground. Kurzweil writes that “the most notable change in the brains of pre-mammals was the region called the cerebellum. Human brains today actually have more neurons in the cerebellum than in the neocortex” (ch. “The Cerebellum: A Modular Structure”). He stands by the architectural picture but updates the framing: the cerebellum stores motor “scripts,” basis functions, and forward models; the neocortex does flexible recombination on top.

That distinction is what the 2025 evidence either supports or doesn’t.

Where the claims actually stand

The cerebellum is a bigger neuron warehouse than Kurzweil said. “The cerebellum contains more than half of the brain’s neurons and supports sensorimotor coordination” (ch. “A Neuromorphic Model: The Cerebellum”). The 2009 isotropic-fractionator work from Suzana Herculano-Houzel — now cited 2,253 times in our literature index — replaced 150 years of dilute estimates with hard counts. The current consensus: the human brain holds about 86 billion neurons, of which roughly 69 billion sit in the cerebellum and 16 billion in the cortex. That’s closer to 80 percent than 50. Kurzweil was directionally right but conservative. Verdict: ahead of schedule on his side; the number got bigger, not smaller, with better methods.

Purkinje synaptic load: directionally right, possibly understated. “Each Purkinje cell receives about 200,000 synaptic inputs” (ch. “A Neuromorphic Model: The Cerebellum”). Modern human work pushes this upward. A 2024 Communications Biology paper on human Purkinje cell dendritic complexity reports total synaptic inputs closer to one million per cell when climbing fibers, parallel fibers, and inhibitory interneurons are summed. The 200,000 figure refers cleanly to parallel-fiber inputs alone. Verdict: the textbook number is intact; the human-specific number is larger.

Alpha motor neurons and 600 muscles. “Some cerebellar outputs go to about 200,000 alpha motor neurons that determine final signals to the body’s approximately 600 muscles” (same chapter). This is straight gross anatomy. It hasn’t moved.

Purkinje cells sample handwriting sequences. “Studies of children’s handwriting learning indicate Purkinje cells sample movement sequences” (same chapter). The specific empirical claim Kurzweil cites — that a population of Purkinje cells with delayed response profiles encodes sequences and explains the inverse curvature-velocity relationship in handwriting — has held up and gotten more mechanistic. Hesslow and colleagues showed in PNAS that cerebellar Purkinje cells learn temporal response sequences. A 2023 Frontiers in Cellular Neuroscience paper modeled how a single Purkinje cell can discriminate temporal input sequences using its dendritic geometry. Verdict: on track and getting more concrete.

Cochlear nucleus is “one of the best understood regions.” “Gordon M. Shepherd described the cochlear nucleus as one of the best understood regions of the brain” (ch. “Another Example: Watts’s Model of the Auditory Regions”). Twenty years later this is still cited as the canonical example of a circuit whose cell types, projections, and computations are mapped end-to-end. Verdict: still true; nothing has displaced it as the gold-standard auditory subsystem.

The Reichardt motion detector. “Max Planck Institute of Biology motion-detection models based on comparing a receptor signal with a delayed adjacent signal closely match animal and human experimental results” (ch. “The Visual System”). The Hassenstein-Reichardt detector is now considered foundational across insects and mammals. A 2014 Nature Neuroscience paper directly identified the model in Drosophila. Human evidence is still psychophysical rather than cellular — the subtraction site has been localized to tangential cells in flies but not yet pinned down in primate cortex. Verdict: on track in spirit, with a thirty-year-old anatomical IOU still outstanding for humans.

Mirror neurons explain language origin and imitation. Here is where Kurzweil bet on a dominant 2002 hypothesis that has since been largely retired. “Giacomo Rizzolatti and Michael Arbib proposed the mirror system hypothesis that language emerged from manual gestures” (ch. “Understanding Higher-Level Functions: Imitation, Prediction, and Emotion”). Heyes and Catmur’s “What Happened to Mirror Neurons?” (Perspectives on Psychological Science, 2021, 240 citations and rising) is the field’s verdict. They report that F5 mirror neurons “do not encode observed actions using the same neural code as that underlying action execution,” undercutting the original simulation story. A 2025 bibliometric review in Brain and Behavior shows mirror-neuron publications peaked at 307 articles in 2013 and have been declining ever since. The “broken mirror” theory of autism, a high-profile child of the original hypothesis, was directly contradicted by Hamilton and colleagues’ “Unbroken mirror neurons in autism spectrum disorders” (197 citations) and by a 2015 Autism Research paper finding intact automatic imitation in autistic adults. Gregory Hickok’s broader critique — that action understanding requires far more than the mirror system — is now the consensus position. Verdict: behind schedule, leaning toward wrong mechanism.

Imitation is critical to language development. “The ability to imitate observed movements or vocal sounds is critical to developing language because it requires decomposition and recursive refinement” (same chapter). The general claim survives — children clearly imitate — but the neural story Kurzweil paired it with (mirror neurons doing the decomposition) has weakened. Verdict: the behavioral claim holds, the mechanistic claim doesn’t.

Recursion as the single key attribute of human language. “Marc Hauser, Noam Chomsky, and Tecumseh Fitch argued in 2002 that recursion is the single key attribute underlying the uniquely human language faculty” (same chapter). Twenty years of pushback — Daniel Everett’s claim that the Pirahã language lacks recursive sentential syntax, formal demonstrations that recursion appears in non-linguistic visual cognition, and a 2017 paper directly titled “Beyond Recursion: Critique of Hauser, Chomsky, and Fitch” — have eroded the strong form of the hypothesis. A 2010 paper “On two incompatible theories of language evolution” (23 citations) frames the standoff. The Hauser-Chomsky-Fitch group has since softened to a distinction between the biological potential for recursion and how individual languages express it, which is closer to what their critics argued in the first place. Verdict: contested; the strong claim is gone, the weak claim is uncontroversial.

Prediction as the primary function of cortex. “Some scientists believe prediction, including prediction of one’s own actions and decisions, is the primary function of the cerebral cortex” (same chapter). This was the throwaway line. It became the field. Predictive-coding papers in our literature index went from 18 in 2005 to 274 in 2025, a fifteen-fold increase. Karl Friston’s free-energy and active-inference framework — work that didn’t really start until 2005, the year Kurzweil’s book was published — now anchors a substantial fraction of computational neuroscience. A 2024 Current Biology paper reports clear prediction-error signals in layer 2/3 of primary visual cortex, exactly where the theory said they should be. Verdict: ahead of schedule. Kurzweil hedged with “some scientists believe.” Twenty years later, “some” is a serious understatement.

The scorecard

What Kurzweil missed and what he nailed

The pattern is striking. Every claim Kurzweil borrowed from anatomy and circuit-level neurophysiology — neuron counts, synapse densities, well-mapped subsystems, motion-detector models — has held up or gotten stronger. Every claim he borrowed from the early-2000s “cognitive neuroscience grand theories” press cycle — mirror neurons explaining language, recursion as the single human distinction — has eroded.

That tracks a deeper truth about technology forecasting. Hardware-side facts (counts, wiring diagrams, response curves) tend to survive because they’re produced by methods that get cheaper and more accurate over time. Theory-side facts about what a brain region “does” are produced by interpretations of small experiments, and those are exactly the claims most likely to be quietly retracted by the next decade. Anyone planning to reverse-engineer a brain on the strength of 2005 cognitive neuroscience would now be rebuilding the upper half of their stack.

The exception is prediction. Kurzweil tagged it almost in passing — “some scientists believe” — and that quiet hedge happens to mark the only line in the chapter that has become more central, not less, to how the field thinks the cortex works.

Method note

We checked each claim against a corpus of 357 million scientific papers, a 9.3 million patent collection, and targeted full-text reads of recent review articles, neuroanatomy papers, and bibliometric analyses. Specific findings — neuron counts, citation numbers, paper titles, year-by-year publication trends — were drawn from those primary sources, not from secondary summaries. Where Kurzweil restated a 2005 claim in The Singularity Is Nearer (2024), we used his updated phrasing as the canonical version.