Kurzweil Scorecard: The Pocket Supercomputer Arrived Early. The Pipe to It Did Not.

In 2005, Ray Kurzweil was defending himself against critics who thought the
exponential curves had to flatten. Buried in his “Response to Critics” chapter
is a small, almost throwaway claim: future technology will yield pocket
supercomputers. It is a phrase he uses to bat away the suggestion that
miniaturization is running out of room — not as a headline prediction. Two
decades later, it is the most quietly correct thing he wrote in that chapter.

What makes the prediction interesting is not that it came true. It is that it
came true so violently that the entire framework for measuring it had to be
rebuilt — and that the supporting prediction underneath it, bandwidth growth,
has held to within a few percentage points while the headline one ran away.

The predictions

This batch contains one prediction and two supporting claims, all from chapter
nine of The Singularity Is Near (2005):

1. “Future technology will yield pocket supercomputers” — a long-term
   forecast, tied to Drexler-style nanotechnology computing.
2. “Bandwidth had been argued to follow a Moore’s-law-like growth trend by
   1999” — citing IEEE Communications work on whether there is a Moore’s Law
   for bandwidth.
3. “Digital signal processing capability had exhibited exponential growth,
   according to Richards and Shaw’s 2004 analysis” — a verified historical
   trend at the time of writing.

The supporting claims look modest. They are the load-bearing wall.

Where we actually are

The pocket supercomputer. Kurzweil did not put a year on this one, and he
was right not to. But the calibration check is easy. The Cray-2, the fastest
machine in the world the year Top Gun opened, hit 1.9 gigaflops at peak
(Wikipedia, Cray-2). The Neural Engine
in an iPhone 16 Pro reaches roughly 15 teraflops at float16, near its 17.5
TFLOPS theoretical maximum, in real measured inference benchmarks against
NVIDIA’s Parakeet v3 speech-to-text model
(argmax benchmark, iPhone 17 release).
That is roughly a 9,000× advance, in a pocket, in 40 years.

Kurzweil himself revisits the calibration in The Singularity Is Nearer
(2024): “Oak Ridge National Laboratory’s Frontier, the world’s top
supercomputer as of 2023, can perform on the order of 10^18 operations per
second. This is already on the order of 10,000 times as much as the brain’s
likely maximum computation speed.” He is now using exaflop supercomputers
to argue the brain has been left behind. The phones he was alluding to in
2005 are now the trailing edge of his argument.

The phones are doing what supercomputers used to do, and they are doing it
in the patent record in a very specific way: not as raw FLOPS, but as
compression. A Signalnet sweep of the patent corpus for on-device
inference and quantization finds the literature moved from a trickle to a
flood:

• Smartphone-plus-machine-learning patents: 12 in 2010, 153 in 2022, 174
  in 2024, 179 in 2025 (year-to-date).
• 5G millimeter-wave beamforming patents: 12 in 2010, 261 in 2021, holding
  at ~230/year through 2025.
• 6G and sub-terahertz patents, essentially nonexistent in 2018, are on
  track for 173 grants in 2025.

The deeper read is in the claims themselves. US 12,554,975, granted 2026,
describes a neural network computing device with an on-device quantizer
that switches a neural accelerator between two operating modes around an
analog multiply-accumulate unit — a circuit designed to do matrix
multiplication in mixed precision without needing to call a cloud GPU.
US 12,632,712 describes a method for a smartphone to estimate the
activation range of each layer of a deep neural network and quantize the
model in place. These are not papers about quantization. These are issued
patents for chips and methods that ship inside phones.

The same pattern shows up in the scientific literature. Searches for
on-device and edge LLM inference in OpenAlex return 1,920 papers in 2018,
6,920 in 2024, and 14,476 in 2025. EfficientFormer: Vision Transformers at
MobileNet Speed (Li et al., 2022, 251 citations,
arXiv:2206.01191) is one of
the bricks in the road — a transformer family designed to hit MobileNet
latency on an iPhone 12. Three years later, the Snapdragon 8 Elite Gen 5
demos a vision-language model with time-to-first-token of 0.12 seconds on
a 1024×1024 image, 11,000 tokens per second on prefill, and over 100
tokens per second on decode, entirely on the phone
(Qualcomm, Snapdragon 8 Elite Gen 5 announcement).
Google’s Gemma 3 1B model, released for mobile in 2025, hits 2,585 tokens
per second on prefill on a mobile GPU
(Google AI Edge blog).
These numbers were not what most people meant in 2005 when they said
“supercomputer.” They are now what people mean when they say “phone.”

Verdict: Ahead of schedule. The 2005 critics Kurzweil was answering
thought the curve had to flatten. It did not.

Bandwidth, the Moore’s-Law-like trend. Kurzweil’s citation here pointed
at IEEE Communications work from 1999. Jakob Nielsen has tracked the same
trend for longer than anyone, and his most recent update is the most boring
finding in this batch — which is what makes it the most important. Nielsen’s
Law of Internet Bandwidth predicts 50% annual growth in high-end home
connections. Best-fit growth across 1984–2019 was 49% per year. The 2023
measurement of 1,120 Mbps “fit the regression line a little lower than
predicted, but still very close,” and the regression had an R² of 0.99
over the full series
(Nielsen Norman Group).

What is striking is that Nielsen has noted, since at least 2019, that his
50%/year bandwidth growth is slower than Moore’s Law’s roughly 60%/year
computational growth. Twenty-six years later, that gap is the entire story
of the pocket supercomputer. Compute outran the wire that delivers data to
it. The on-device LLM is not a privacy story or a latency story or a
sovereignty story, though it is also all of those. It is, mechanically,
a story of compute compounding faster than bandwidth — exactly as Nielsen
warned in 1998 and Kurzweil tacitly accepted in 2005 by citing the
bandwidth growth trend as the underlying claim, never as the bottleneck
to be solved.

Verdict: On track. Nielsen’s measurement is within a few percent of the
1998 forecast. Kurzweil’s citation was directionally correct and
quantitatively correct.

DSP exponential growth (Richards and Shaw, 2004). This is the one where
the verdict requires unusual care. Richards and Shaw documented an
exponential growth in digital signal processor capability through 2004.
For most of the next decade, that trend continued under the same name —
Qualcomm’s Hexagon line, originally a DSP for cellular radio and audio,
followed it. Then in 2018 the Hexagon 690 added the Hexagon Tensor
Accelerator alongside the DSP cores
(chipsandcheese on Hexagon).
By 2025, the Snapdragon 8 Elite Gen 5 fuses the DSP, the vector engine,
and the tensor accelerator into a single block now called the NPU, with
12 scalar engines, 8 vector engines, and a dedicated AI accelerator,
running mixed-precision math from INT2 to FP16.

The patent record reflects the relabeling. Searches for digital signal
processor patents with vector or tensor features stay flat at a few per
year across the 2010s and 2020s. The action moved to mobile neural
inference: 110 patents in 2020, 170 in 2023, 179 in 2025. The trend
Richards and Shaw measured continues. The chip it lives in is no longer
the chip they were measuring.

Verdict: Wrong mechanism, right destination. The exponential growth
in mobile signal-processing throughput continued. The boundary of the
device that delivers it dissolved.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern in this batch is the cleanest pattern in any batch I have scored.
Kurzweil cited two existing exponential trends as load-bearing — bandwidth
growth and DSP growth — and bet on a future device, the pocket supercomputer,
that the two trends would jointly enable.

The bandwidth trend continued almost exactly as advertised. The DSP trend
continued but moved categories. And the pocket supercomputer arrived early
and went further than the long-term framing implied. The thing that made
the prediction interesting in 2005 was Kurzweil’s willingness to forecast
the intersection of trends, not just the trends themselves. He was right
to ground his bigger claim in the smaller measurable ones.

The systematic bias in this batch is one I have seen elsewhere in Kurzweil’s
2005 work: he gets the direction right, the magnitude right, and the
labeling wrong. The pocket supercomputer is no longer called one because
the comparison stopped being made. The DSP is no longer a category because
the function migrated into a fused block. Vocabulary moves faster than
capability, and a 2005 reader can fail to notice the prediction came true
because the words for it changed.

Kurzweil’s defense in The Singularity Is Nearer is that the underlying
curve is paradigm-agnostic: “this macro trend has steadily continued — and
it does not rely on any particular technological paradigm like shrinking
transistors or increasing clock speeds.” He restated the point in 2024
about price-performance generally, but it applies cleanly to this batch.
The trend was real. The vessel changed.

The interesting bandwidth gap, meanwhile, is doing exactly what Nielsen
predicted it would do in 1998 — quietly creating a structural advantage
for whoever can run inference locally. That is now a $3 trillion bet on
which Apple, Qualcomm, Google, and Samsung are visibly stacking patents.
The on-device LLM is not a 2024 idea. It is a 1998 idea about Moore’s Law
outrunning Nielsen’s Law, made operational by the pocket supercomputer
Kurzweil predicted in 2005.

Method note

For each prediction, we combined direct quotations from The Singularity Is
Near (2005) and The Singularity Is Nearer (2024) with: patent grant and
publication counts from the U.S. patent record spanning 2010–2026; full-text
abstracts of recent on-device quantization and neural accelerator patents;
citation-ranked papers from the OpenAlex scientific literature corpus;
benchmark reports from third-party measurement labs; and the canonical
Nielsen’s Law dataset spanning 1984–2023. Patent numbers are real and
published; the claim-level reads are from the granted patents themselves.
Bandwidth and benchmark figures are linked inline.