Kurzweil Scorecard: The Spines Became the Interface

In 2005, Ray Kurzweil opened a short section on synapses and spines with a
small wonder: that brain scanning could now see new connections forming —
that learning was visible as physical change. He cited a Cold Spring Harbor
mouse lab, a Harvard piano study, and a 2002 actin-turnover paper. He
treated it as a basic-science footnote on the road to mind uploading.

He was right on every count. The footnote also turned out to be load-bearing.
The same dendritic-spine biology Kurzweil flagged as a curiosity has become
the operating principle of two distinct industries that did not exist in 2005:
implantable speech prosthetics for paralyzed patients, and rapid-acting
antidepressants built specifically to regrow lost spines. Mental practice
no longer just reshapes motor cortex. It drives cursors and writes sentences.

The predictions

Kurzweil’s Subneural Models: Synapses and Spines and Brain Plasticity
sections (in The Singularity Is Near) made four testable claims about the
state of neuroscience circa 2005:

1. “Recent scanning results show rapid growth of dendrite spikes and new
   synapses, making real-time connection changes an important learning
   mechanism.” (ch. “Subneural Models: Synapses and Spines”)
2. “Karel Svoboda’s mouse studies at Cold Spring Harbor showed dendrites
   continually grow new spines, most lasting only a day or two, with
   stabilized spines reflecting favorable rewiring.” (same chapter)
3. “High-resolution brain scanning can now directly observe dendritic-spine
   growth and new synapse formation, showing the brain physically adapts to
   use and thought.” (ch. “Brain Plasticity”)
4. “Alvaro Pascual-Leone showed that merely thinking about practicing a
   piano exercise, without moving muscles, produces motor-cortex changes as
   pronounced as physical practice.” (same chapter)

In The Singularity Is Nearer, Kurzweil revisits this terrain only
glancingly, citing Star, Kwiatkowski, and Murthy’s 2002 Nature Neuroscience
paper on actin turnover in dendritic spines as evidence that “the
computational substrate of the brain is itself in motion.” That phrasing has
held up better than most of the book.

Where we actually are

The spines are real, and we watch them turn over in real time.

The technical claim that two-photon imaging could see spines come and go was
already true when Kurzweil wrote it. What has happened since is volume and
resolution. A 2018 eLife paper used two-photon STED microscopy through a
chronic hippocampal window in living mice and measured 40% spine turnover
within four days, with turnover rate dependent on spine size — exactly the
“most lasting only a day or two” picture Kurzweil sketched.

The Yale group around Alex Kwan made the same imaging methodology useful for
drug discovery. Their 2021 Neuron paper — now cited 547 times in our
literature index — showed that a single dose of psilocybin produced ~10%
increases in dendritic spine size and density in mouse frontal cortex within
24 hours, with the new spines still present a month later. Ketamine
produces a similar pattern at subanesthetic doses. The same actin-turnover
machinery Kurzweil cited from a 2002 basic-science paper is now the explicit
mechanism that pharmaceutical companies cite to justify dosing humans.

A 2025 Cell paper extended the finding to network-scale rewiring: psilocybin
triggers an activity-dependent reorganization of large-scale cortical
networks, not just local spine growth. Spines remain the unit of currency.
The trading floor has gotten larger.

Mental practice changes motor cortex. And it now writes English.

Pascual-Leone’s 1995 piano study — the one Kurzweil leaned on — used TMS to
measure motor-cortex expansion in subjects who only imagined practicing a
five-finger exercise. The expansion was similar in magnitude to the
expansion in subjects who physically practiced. The finding has been
replicated repeatedly in the music-performance literature; the cortical
representation expanded similarly in both groups, and accuracy and temporal
consistency improved in both.

What Kurzweil could not have anticipated is the engineering payoff. The same
plasticity that lets a pianist rehearse without a piano is what lets a
paralyzed patient operate a computer without muscles. In 2023, the Stanford
group around Jaimie Henderson and Francis Willett published a Nature paper
demonstrating a speech neuroprosthesis with four microelectrode arrays
implanted in area 6v (ventral premotor cortex) and area 44 (Broca’s). A
patient with ALS produced text at 62 words per minute — 3.4 times the
previous record — with a 9.1% word error rate on a 50-word vocabulary and
23.8% on a 125,000-word vocabulary. Natural conversation is roughly 160
words per minute. The gap is closing fast.

In January 2024, Neuralink implanted its N1 device — 1,024 electrodes across
64 flexible threads — in the motor cortex of Noland Arbaugh, a quadriplegic
patient. Within a week post-surgery he was operating a computer cursor by
imagined movement; within months he was playing online chess. When most of
his threads retracted from cortex, the system was salvaged by retraining the
decoder — i.e., by letting plasticity work on both ends of the interface.

The patent record confirms this is no longer an academic curiosity. Patent
filings mentioning “brain computer interface” climbed from a handful per year
in 2010 to 30 in 2022, 27 in 2023, 30 in 2024, and 42 in 2025, with
another 10 already in the first weeks of 2026. US 12,548,570, granted
February 2026 to a major BCI company, describes a neural foundation model
that ingests features extracted from microelectrode-array recordings,
transforms them through encoders, and predicts phonemes via decoders — a
direct application of transformer architecture to the silicon side of the
motor-imagery loop. The claims describe subject-specific embeddings
concatenated with user profile embeddings, language about plasticity on the
machine side that mirrors plasticity on the cortical side.

US 12,508,079 (Dec 2025) goes further on the hardware end: it claims three
distinct minimally invasive delivery routes for electrode arrays — cortical
surface, ventricular, and endovascular — that would let BCIs piggyback on
existing interventional radiology workflows rather than requiring craniotomy.
The patent reads like an attempt to industrialize what is currently a
research procedure.

Watching the human cortex itself remains hard.

The one prediction Kurzweil reaches a little too far on is the third —
that scanning can directly observe dendritic-spine growth in the human
brain. In 2005 he was describing two-photon imaging in mice. Twenty years
later, the technical situation is roughly the same. Two-photon and STED
methods work routinely in rodents and have been extended to marmosets in
eNeuro (2015), but as that paper explicitly notes, “there has been no
report in which dendritic spines were imaged and analyzed in vivo in
primate brains” with the same fidelity, and human in vivo work at the
spine level remains absent. Functional MRI sees macroscopic plasticity at
millimeter resolution. Spine-level plasticity in the living human cortex
still requires inference from indirect signals. Kurzweil was right that
the brain physically adapts to thought. He elided that seeing the
adaptation in your brain is still beyond instrument reach.

The scorecard

What Kurzweil missed (and what he nailed)

What he nailed is the directional bet. Plasticity, in 2005, was a topic
mainly of interest to systems neuroscientists and a small cohort of
rehabilitation researchers. Kurzweil placed it in the path of his merger
argument — the brain is a thing that rewires itself, therefore the substrate
is malleable, therefore the merger is mechanically possible. That argument
has held up. Two industries — implantable BCIs and rapid-acting
antidepressants — are now built on top of that one footnote.

What he missed is the asymmetry. He framed plasticity primarily as a
property to be replicated in silicon: if a neural network can rewire
itself the way a biological one does, we can simulate intelligence. What
actually happened is closer to the opposite. The plasticity got
exploited without being replicated. Speech neuroprostheses do not need
to simulate a cortex. They need to ride along on the cortex they already
have. The decoder learns; the cortex learns; the loop closes. Kurzweil’s
2005 framing — substrate-independent intelligence — pointed at a future in
which biology is left behind. The 2020s engineering reality leaned the
other way: biology is the substrate that the new technology is built on
top of, not around.

There is also a methodological observation worth noting. The four
predictions in this batch were all retrospective — claims about what 2005
neuroscience had already established. They were, in effect, footnotes
Kurzweil used to scaffold larger forecasts. They have aged better than most
of his timeline predictions. The lesson, perhaps, is that Kurzweil’s
science reportage was reliable and his timelines were ambitious. When he
told you what was true, he was usually correct. When he told you when it
would matter, the answer was almost always later than he said — except
here, where the answer was sooner.

Method note

Spine-imaging and plasticity papers were surveyed in an index of 357 million
scientific works, filtered to the high-citation tail, and read at the
abstract level where DOIs allowed. The BCI claims were grounded in the
2023 Stanford speech-prosthesis Nature paper and the Neuralink human-trial
disclosures. Patent claims were read directly from a 9.3-million-document
US patent corpus, restricted to grants since 2010 that mention brain-computer
interfaces. The Svoboda and Pascual-Leone references were checked against
Kurzweil’s footnote citations in The Singularity Is Nearer.