Kurzweil Scorecard: The Shortcut Around the Brain

Ray Kurzweil’s 2005 roadmap to machine intelligence was specific and, in
retrospect, patient. First, reverse-engineer the brain. Then implement its
operating principles on faster substrates. Then merge with it. Twenty-one
years later, the intelligence arrived. The brain reverse-engineering did
not. A different path got there first — and it runs on a learning rule
Kurzweil himself flagged in the same book as “not biologically realistic
for mammalian brains.”

What Kurzweil claimed

Across twelve predictions in chapters including “The Software of the
Brain,” “Reverse Engineering the Brain: An Overview of the Task,”
“Modeling the Brain,” “A Neuromorphic Model: The Cerebellum,” and
“Uploading the Human Brain,” Kurzweil sketched a sequence. A
“functional simulation of human intelligence will pass the Turing test
by 2029” (ch. “Uploading the Human Brain”). Getting there required
building “a verifiable, real-time, high-resolution model of significant
parts of human intelligence” (ch. “Achieving the Software of Human
Intelligence”). Once the algorithms were in hand, they would run on
“synthetic neural equivalents” on substrates “already far faster than
neural circuitry” (ch. “Peeling the Onion”). Humans would then
“effectively upload themselves gradually through increasing nonbiological
augmentation” (ch. “Uploading the Human Brain”), and by the 2040s the
nonbiological portion of human intelligence would be “billions of times
more capable than the biological portion.”

In The Singularity Is Nearer (2024), Kurzweil restated the claim
almost verbatim: “Passing the Turing test, which I have been
anticipating for 2029, will bring us to the Fifth Epoch.” He added
that Metaculus had moved from a 2040s–2050s consensus to agreeing with
him on 2029 by May 2022, and since drifted as early as 2026 —
“putting me technically in the slow-timelines camp!”

Where we actually are

Turing-level performance is close. The Kapor-Kurzweil bet is still
open. The 2023 “Sparks of AGI” study on an early GPT-4 — 1,517
citations — concluded the model showed “remarkable capabilities
across a variety of domains and tasks, challenging our understanding
of learning and cognition.” The 2022 “Emergent Abilities of Large
Language Models” paper (1,015 citations) made the sharper point:
multi-step arithmetic, multilingual reasoning, and instruction-following
appear discontinuously as models scale. Those are the abilities the
Turing test probes. But Kurzweil and Kapor’s specific 2002 Long Bets
protocol is stricter than the public “does it feel human” version, and
both parties have said recent LLMs don’t meet it. The bet remains
unresolved — and the Metaculus crowd now finds 2029 plausible.

The brain isn’t what we reverse-engineered — the shortcut was.
Nearer describes transformers as using “attention” in a way
reminiscent of the neocortex, but attention-as-implemented is a matrix
operation over token embeddings, not a neuromorphic model. Patent
filings mentioning “large language model” training or fine-tuning
jumped from 13 in 2023 to 112 in 2024 and 269 in 2025 — a 20x
expansion. Recent grants read like plumbing for commodity technology:
US 12,432,128 (“Efficient generation of specialized large language
models for network traffic analysis,” September 2025) describes
“performing transfer learning on a base large language model…
trained using network traffic capture files” so specialists “may be
developed in an expedient and efficient manner.” That is industrial
mass-production. It is not cortical simulation. Google’s Conformer
patent US 12,373,666 (July 2025) claims “a first half-step
feed-forward block… a self-attention block… a convolutional block…” —
a stack of linear algebra optimized to minimize speech-recognition
error, not to resemble any known circuit.

The cerebellum story is the one place Kurzweil got the mechanism
right. His 2005 summary of the Texas cerebellum simulation — it
matched animal eyeblink conditioning data and reproduced rabbit lesion
phenotypes — was accurate, and the line of work continued. A 2024 PLOS
Computational Biology paper (“Mesoscale simulations predict the role of
synergistic cerebellar plasticity during classical eyeblink conditioning”)
used a spiking cerebellar model with biologically realistic plasticity
rules and “closely reproduced the behavioral phenotypes of mutant mice
with altered cerebellar synapses.” A 2024 Frontiers paper implemented a
real-time cerebellar spiking network on FPGA for adaptive motor control.
These are working systems, built from circuit biology, doing what
Kurzweil said they would do. They just have very little to do with the
systems passing graduate exams.

The Watts auditory model actually shipped. Lloyd Watts founded
Audience in 2000 and commercialized chips that, per the company’s
engineering material, reverse-engineered the cochlea and the first
stages of the auditory brainstem for noise suppression in mobile phones.
First-generation two-microphone noise suppression chips launched in 2007
and reached market-leading smartphones by 2010 — roughly contemporaneous
with Singularity Is Near. Kurzweil’s claim that a real-time software
model of “a significant portion of the human auditory-processing
system” existed and could localize and identify sounds was, at time of
writing, already demonstrably true.

The whole-brain model is not close. The FlyWire consortium’s
October 2024 nine-paper package in Nature reported the first
whole-brain connectome of an adult animal: 139,000 neurons, 50+ million
synapses, 8,453 cell types, reconstructed from a Drosophila
melanogaster. It is a landmark. It is also a fruit fly. The human
brain is ~86 billion neurons and ~100 trillion synapses — roughly six
orders of magnitude larger on each axis. Nectome, the leading
commercial brain-preservation effort, published a March 2026 preprint
on ultrastructural preservation of a large mammal brain. Anders
Sandberg, who co-authored the 2008 whole-brain emulation taxonomy,
recently estimated a ~5 percent chance of brain emulation within ten
years. The gradual uploading Kurzweil sketched for the 2040s is not on
that timetable.

Hybrid minds arrived, but not through neurons. Neuralink’s PRIME
trial had reached roughly 21 implanted participants by early 2026 —
cursor control and assistive communication, not the cortex-to-cloud
merge of Nearer. The hybrid the public actually uses is a chat
window. People extend their memory, reasoning, and writing through
GPT-style systems they talk to. That is real augmentation, arguably
more effective at the population level than neural implants will be
this decade. It is not the mechanism Kurzweil described.

Knowledge-sharing among AIs arrived early. Kurzweil wrote that
“when one AI learns something it will quickly share that knowledge
with many others.” Knowledge distillation — training a “student”
network to imitate a “teacher” — is now a default engineering move. US
patents on neural-network distillation grew from 1 in 2021 to 7 in
2025; the real footprint is larger because most practitioners treat it
as table stakes. US 12,346,813 (“Online knowledge distillation for
multi-task learning system,” July 2025) and US 12,327,085 (“Sentence
similarity scoring using neural network distillation,” June 2025)
codify specific variants. The field also normalized training new
models on the outputs of older ones. That is “one AI learns, others
absorb it” — implemented as file copy, not neural transfer.

The scorecard

Tally: 2 ahead, 3 on track or verified, 1 behind, 5 wrong mechanism,
1 verified historical fact, 1 too early to call.

What the pattern says about Kurzweil — and about forecasting

Strip out the two historical claims (backprop’s 1980s resurgence; the
Texas cerebellum work) and the pattern is sharp. Where Kurzweil forecast
outcomes — Turing-level performance, AI-to-AI knowledge transfer,
hybrid minds — he was closer to right than most of his peers thought in
2005. Where he forecast mechanisms — neuromorphic substrates, gradual
neural uploading, working from biological operating principles — the
field took a hard left. The intelligence showed up on a path made of
gradient descent on autoregressive objectives over internet text.
Attention-plus-MLP blocks, not cortical columns. Biology was the muse,
not the blueprint.

There is a tempting version of this scorecard that reads as a Kurzweil
loss. That reading ignores what he got right: the accelerating-returns
argument implied that whichever substrate paid off would scale fast
and surprise forecasters. Forecasters who were more correct on the
mechanism — expecting neuromorphic hardware, spiking networks,
biologically-faithful models — were dramatically more wrong on the
timeline. Kurzweil’s bet was that the timeline was the load-bearing
variable. With 2029 approaching, he looks less wrong every year.

The FlyWire connectome, the 2024 cerebellar simulations, Nectome’s
preservation work, and Neuralink’s trials all continue. They may yet
matter — for medicine, for neuroscience, for a later, deeper kind of
artificial mind. But the intelligence that will either settle or lose
the 2029 bet is already here, and it came from the shortcut.

Method note

This scorecard draws on the twelve-prediction AI / brain-software
cluster from The Singularity Is Near, full-text reads of The
Singularity Is Nearer (2024), trend and deep-read searches across
~9.3M US patent filings, citation-ranked OpenAlex abstracts on
connectomics, cerebellar simulation, and emergent LLM capabilities, and
targeted web reads of the Kapor-Kurzweil Long Bet, Metaculus AGI
forecasts, FlyWire’s 2024 Nature package, Nectome’s 2026 preservation
preprint, Neuralink PRIME enrollment, and the commercial history of
Audience. Specific numbers come from those sources.