Kurzweil Scorecard: The Eye Got Its Chip. The Synapse Kept Its Memory.

In 2005, Ray Kurzweil used a batch of neuroscience facts as load-bearing walls for a philosophical argument. The brain’s molecules, he argued, turn over so fast that you cannot be your atoms — microtubules every ten minutes, actin filaments every forty seconds, synaptic proteins every hour. You are a pattern. A pattern is uploadable. He also pointed to then-hot findings: Benjamin Libet’s readiness potential, John Allman’s spindle cells, Antonio Damasio’s body maps, and retinal implant work at MIT and Harvard.

Twenty-one years later, most of the facts held up in some form. One — the synaptic protein turnover rate — was off by more than a hundred-fold, and the error matters for the argument built on top of it. And the retinal implant prediction was fulfilled spectacularly in October 2025, but not by MIT or Harvard, and only after the first commercial attempt left roughly 350 blind patients with obsolete chips in their eyes.

The predictions

These ten claims come from Kurzweil’s chapters on higher brain function and body engineering (The Singularity Is Near, chs. “Understanding Higher-Level Functions: Imitation, Prediction, and Emotion,” “on the Human Body,” and “Who Am I? What Am I?”). They were not timelined predictions so much as the factual grounding of his argument that identity is information, not hardware. The question is: did the biology hold up, and does the argument built on top of it still stand?

Where we actually are

Retinal implants: right direction, wrong institutions, survivor pile. Kurzweil wrote that “researchers at MIT and Harvard were developing neural implants to replace damaged retinas” (ch. “on the Human Body”). It happened — twice. Second Sight’s Argus II received FDA approval under a humanitarian device exemption in 2013 and was implanted in about 350 patients worldwide. In 2019 the company quietly discontinued the product; by 2020 it had laid off most of its staff and stopped providing technical support. Patients were not reliably informed. Many now carry non-functioning electrodes in their eyes — painful and expensive to remove, a complication for future MRI scans. The retinal-prosthesis patent pipeline ran 15–21 issuances per year through the mid-2010s in our local index, then collapsed to low single digits from 2019 onward.

The second attempt came from a different direction. The PRIMA system — a 2mm × 2mm subretinal photovoltaic chip developed at Stanford, spun out as Pixium Vision, and acquired by Science Corporation — published its PRIMAvera trial in NEJM on October 20, 2025. Thirty-eight patients with geographic atrophy from age-related macular degeneration showed a mean visual-acuity improvement of 25.5 letters on the ETDRS chart, more than five lines. Eighty percent gained at least ten letters. Patients read letters, numbers, and words. European regulatory approval is pending; U.S. review is underway. Recent U.S. grants track the second wave: US 12,544,572 (February 2026) and US 12,415,065 (color variant, September 2025). Verdict: Arrived, through a different door, over a pile of abandoned patients.

Libet’s readiness potential: fact real, interpretation gone. Kurzweil cited Libet’s finding that “neural activity initiating an action occurs about one-third of a second before conscious awareness of the decision” (ch. “Understanding Higher-Level Functions”) and presented it as settled. The 300-millisecond gap is still measurable. The causal story is not. Aaron Schurger’s stochastic accumulator model (PNAS 2012; Trends in Cognitive Sciences 2021 review; Gavenas et al. 2024) reframes the readiness potential as a statistical artifact of time-locking the average to threshold crossings of noisy accumulation — what averaging random walks looks like when you align them to the moment they crossed a decision threshold. A 2025 Imaging Neuroscience paper from Schurger’s group (“Probing for intentions”) explicitly concludes the early readiness potential does not reflect awareness of motor preparation. The 2021 review has 30 citations in our index; the 2018 “Rise and Fall” analysis has 67. Verdict: Fact intact, interpretation overturned.

Confabulation from stimulation: still holds. Kurzweil wrote that “when neurophysiologists electrically stimulate emotional centers, subjects immediately invent rationales for the induced feelings, similar to confabulation in split-brain patients” (same chapter). Replicated many times since 2005 in depth-electrode studies during epilepsy surgery and in continued analysis of Gazzaniga’s split-brain work. Kurzweil himself restates this in The Singularity Is Nearer (2024). Verdict: Verified.

Spindle cell counts: the numbers held. Kurzweil wrote that “humans have about 80,000 spindle cells, gorillas about 16,000, bonobos about 2,100, chimpanzees about 1,800” (same chapter). CARTA’s anthropogeny database, citing later Allman-group counts, gives 82,855 / 16,710 / 2,159 / 1,808. Verdict: Verified.

Spindle cells are not uniquely hominid. Kurzweil wrote they “first appeared 10 to 15 million years ago in a common ape-hominid ancestor and increased rapidly around 100,000 years ago,” and that “other mammals lack them entirely.” Both claims have softened. Von Economo neurons have since been identified in dolphins, beluga whales, humpback whales, elephants, macaques, sheep, cows, white-tailed deer, and the pygmy hippopotamus. A 2021 Frontiers in Neural Circuits review asked whether VENs are “primate-specific or commonplace in the mammalian brain” and landed closer to the latter. Human densities remain unusually high, but the clean phylogenetic story is gone. Verdict: Wrong framing.

Spindle cell development: on track. Kurzweil wrote they are “absent in newborn humans, begin appearing around four months, and increase significantly from ages one to three” (same chapter). The anatomical trajectory remains the textbook Allman picture; a 2020 Nature Communications paper (99 citations in our index) added molecular specificity without overturning it. The link to moral or love-related capacities is harder to pin down empirically. Verdict: Verified.

Damasio’s body maps: verified and extended. Kurzweil wrote that “recent findings support Antonio Damasio’s view that emotions are closely linked to brain areas containing maps of the body” (same chapter). Interoception — the brain’s perception of internal bodily states via the insula and anterior cingulate — is now one of the most active areas in cognitive neuroscience, with direct implications for anxiety, depression, and eating-disorder research. Verdict: Verified and extended.

Microtubule half-life: the fast-pool number only. Kurzweil wrote that “the half-life of a neuronal microtubule is about ten minutes” (ch. “Who Am I? What Am I?”). That figure comes from dividing cells and the dynamic microtubule population. Neurons are different. Reviews in 2020–2024 make the case explicitly: neuronal microtubules are relatively stable, held in place by post-translational modifications (acetylation, detyrosination, polyglutamylation). Polymer turnover in dendritic segments has been measured in hours, not minutes. There is a fast pool and a slow pool; Kurzweil cited only the fast one. The mistake matters because the philosophical argument — you cannot be your molecules, they turn over too fast — depended on there being no molecular continuity at all. Verdict: Half right.

Actin in dendrites: verified. Kurzweil wrote that “actin filaments in dendrites are replaced about every forty seconds” (same chapter). The rapid dendritic-spine actin turnover figure has been confirmed many times; a 2006 JCB paper (“Actin turnover–dependent fast dissociation of capping protein,” 128 citations) became the canonical reference, and 2021–2022 eLife computational models tightened the numbers. Verdict: Verified.

Synaptic protein turnover: the one Kurzweil got badly wrong. He wrote that “proteins that power synapses are replaced about every hour” (same chapter). Cohen and Ziv’s 2013 study measured synaptic protein half-lives in cultured neurons in the range of 2–5 days. Their 2018 PNAS follow-up (“Identification of long-lived synaptic proteins,” 113 citations) identified a distinct class with half-lives of several months or more. A 2018 eLife paper (“Local and global influences on protein turnover in neurons and glia”) corroborated the spectrum view. This inversion matters because Kurzweil used the rapid-turnover figure as the clinching argument that you are a pattern, not a substance. Some synaptic proteins last about as long as the memories they encode. The pattern view survives; the “all flesh is rushing water” metaphor does not. Verdict: Behind (off by roughly two orders of magnitude on the slow pool).

The scorecard

What Kurzweil missed (and what he nailed)

The philosophical edifice survived even when a numerical receipt cracked. The synaptic-protein-turnover figure was off by roughly two orders of magnitude on the slow tail, but the larger point — you are more pattern than specific molecules — does not require every molecule to turn over every hour. It only requires the pattern to be substrate-agnostic. The 2018 long-lived-protein work says something more interesting than either Kurzweil’s 2005 version or its naive negation: specific synapses carry molecular scaffolds that persist for months or years, and those scaffolds may be part of how long-term memories stay put. If anything, that is a better foundation for upload-talk, because it gives you a specific substrate to read.

Kurzweil was also overconfident about clean phylogenetic and institutional stories. Spindle cells were supposed to be a hominid breakthrough; they turned out to be widespread. Retinal implants were supposed to come from MIT and Harvard; they came from Stanford and a pair of spin-outs, one of which cratered and stranded its patients before the next one got the NEJM paper. The readiness potential was supposed to be a clean window into consciousness; it turned into a cautionary tale about averaging artifacts. Seven of ten claims verified or partially verified is a strong showing for a 2005 book. The failures are specific and informative. The patients who believed the retinal-implant prediction first paid the price for being early.

Method note

This scorecard draws on our local index of about 9.3 million patent records (tracking retinal-prosthesis issuances by year and reading recent claims), our local index of about 357 million scientific papers with citation counts (ranking foundational and rebuttal literature in each subfield), and web searches for clinical trial results and recent reviews. Specific patents, papers, and trial outcomes are named above by number or DOI.