Kurzweil Scorecard: The 2005 Neuroscience Textbook Held. The Instruments Didn’t.

In 2021, a team at Hebrew University of Jerusalem trained a deep neural network to reproduce, at millisecond resolution, the input‑output behavior of a single layer‑5 pyramidal neuron from rat cortex. They needed a temporally convolutional network with five to eight hidden layers to do it — and when NMDA receptors were knocked out of the biophysical model, one hidden layer was suddenly enough. The implication, made explicit in their Neuron paper: a single cortical neuron is doing work roughly equivalent to a small deep network, not a threshold gate with 200 switching events per second.

Ray Kurzweil, in 2005, wrote that “individual neurons typically execute about 200 transactions per second, but the brain’s massive parallelism yields up to 100 trillion computations simultaneously” (ch. “Is the Human Brain Different from a Computer?”). On the firing‑rate side he was right, and still is. On the ops‑per‑neuron side, the 2021 paper puts him off by roughly three orders of magnitude. That single swap — a biophysical neuron becoming an 8‑layer network — is the cleanest way to describe what changed in neuroscience since The Singularity Is Near shipped.

Batch 46 is ten neuroscience claims Kurzweil made circa 2005: historical facts about Adrian, Barlow, and Hebb, and instrument specs for fMRI, MEG, and PET. Twenty years on, the textbook facts are holding up. The instrument descriptions, not so much — because almost none of the progress since 2005 came from tuning the instruments Kurzweil named. It came from replacing them.

The predictions

Kurzweil’s 2005 map had three layers: foundational discoveries about how neurons signal (Adrian in 1928; Barlow’s trigger features in the frog and rabbit retina; Hebb’s reverberatory memory), instrument envelopes (fMRI at 1–3 mm and a few seconds; MEG at 1 ms and about 1 cm; PET and fMRI reading out via the subtraction paradigm), a specific bet that Fritz Sommer at Redwood Neuroscience Institute would fuse fMRI and MEG, and one big claim about total compute (≈200 Hz per neuron × ~10¹¹ neurons → ~10¹⁴ ops/sec, with experience boosting brain “complexity” about a billion‑fold beyond what the genome encodes).

In The Singularity Is Nearer (2024), Kurzweil sharpened the imaging envelope: fMRI down to “cubic voxels about 0.7 to 0.8 millimeters to a side” with “a couple of seconds” temporal lag, and scalp electrical methods at “about one millisecond” temporal but “six to eight cubic centimeters” spatial. He added: “even though we may see marginal improvements from AI and improved sensor technology, they probably won’t be sufficient to allow a sophisticated brain‑computer interface.” That’s the moving target. The 2005 claims are the ones the batch asks us to score.

Where we actually are

The textbook facts held. Adrian’s 1928 demonstration that nerve cells produce electrical pulses, and that pulse frequency scales with stimulus intensity, is still the first thing a neurophysiology student learns. Horace Barlow’s retinal “trigger features” — neurons that respond to specific directions, shapes, or velocities — became the conceptual seed for Hubel and Wiesel’s orientation columns, and from there for the convolutional filter banks that underlie every modern vision model. Hebb’s reverberatory memory idea, which Kurzweil flagged as “less established than Hebbian synaptic learning,” has kept accruing evidence — recurrent dynamics, gamma‑band oscillations that bind features, hippocampal replay. Not settled, but not losing ground.

fMRI got sharper by going stronger, not by solving the BOLD problem. The 1–3 mm / few‑second envelope is now the clinical floor, not the ceiling. A 2024 Nature Methods paper described a next‑generation 7‑tesla scanner imaging human cortex at 0.35–0.45 mm isotropic — small enough to resolve individual cortical layers. A 2025 submillimeter 7T study mapped laminar response changes in primary motor cortex in Parkinson’s patients with 3D passband bSSFP at that resolution. The temporal side has barely moved, though, because the BOLD signal itself lags neural activity by hundreds of milliseconds. Spatial: beaten by roughly an order of magnitude. Temporal: stuck where Kurzweil left it.

MEG didn’t improve. It got replaced. Magnetoencephalography in 2005 meant a subject sitting inside a cryogenic helmet cooled to 4 K, inside a magnetically shielded room — 1 ms, 1 cm, and immovable. Optically pumped magnetometry made Kurzweil’s spec obsolete. Our literature index shows OPM‑MEG papers jumping from 3 in 2010 to 61 in 2025. The landmark work — a 2020 Science Advances paper recording brain activity in the unshielded Earth’s field; a 2022 Radiology evaluation of on‑scalp OPM‑MEG versus cryogenic MEG for pediatric epilepsy; a 2023 Frontiers in Neuroscience description of a 128‑sensor full‑head OPM‑MEG system from FieldLine Medical — all converge on the same shift: the helium dewar is gone, the helmet fits children, the sensors sit close enough to the scalp to partly collapse the old 1 cm floor. Cerca Magnetics (Nottingham, UK) built 128‑sensor wearable systems that entered pediatric epilepsy and Parkinson’s tremor trials in 2024–2025. Kurzweil’s MEG envelope was correct for the instrument he named; it does not describe the instrument clinicians now reach for.

Fritz Sommer’s combined‑modality dream showed up — without Fritz Sommer. Kurzweil specifically credited “Fritz Sommer at Redwood Neuroscience Institute” with developing methods to combine fMRI and MEG (ch. “Peering into the Brain”). Sommer is still affiliated with Berkeley’s Redwood Center — he co‑founded it in 2005 — but his day‑to‑day work migrated to neuromorphic computing at Intel Labs; he isn’t running the fMRI–MEG fusion program Kurzweil attributed to him. The fusion itself arrived on a different track. A framework for simultaneous EEG‑fMRI at 7 T, published in Imaging Neuroscience in 2025, achieves millisecond temporal resolution alongside sub‑millimeter spatial resolution — the spec Kurzweil wanted from fMRI+MEG, built from concurrent EEG and ultra‑high‑field fMRI instead. And a trio of recent patents from a Japanese industrial team — US 11,914,012 (February 2024), US 11,957,472 (April 2024), and US 12,262,974 (April 2025) — describe a single helmet‑mounted instrument that pairs an MRI detection coil with an optically pumped magnetometer, registering the magnetic field distribution and the MR image in one pass. That is the Sommer concept, commercialized, with OPMs standing in for the cryogenic MEG that made the original dual‑instrument plan a nightmare.

PET + fMRI by subtraction is still standard — but narrower. The subtraction paradigm (compare brain state under task A to state under task B; map the difference) is still default for activation mapping, and PET regional cerebral blood flow is still the reference against which BOLD is calibrated. It now shares the stage with resting‑state connectivity, arterial spin labeling, and model‑based decoding. Verified, and narrower than it was.

The neuron itself was denser than Kurzweil assumed. His “200 transactions per second” per neuron still matches measured maximum cortical firing rates, so the top‑line 10¹⁴ ops/sec number isn’t wrong by the rule Kurzweil was using. What shifted is the rule. The 2021 Beniaguev, London, and Segev paper in Neuron showed that faithfully reproducing the spike‑timing output of a single cortical pyramidal cell — dendrites, NMDA receptors, and all — requires a five‑ to eight‑layer temporal convolutional network per neuron per millisecond. Multiply that by 10¹¹ neurons and the effective compute ceiling is several orders of magnitude above Kurzweil’s 10¹⁴ estimate. The firing rate is right. The neuron isn’t the unit of computation he thought it was.

The scorecard

What Kurzweil missed (and what he nailed)

Where Kurzweil reported settled neuroscience — Adrian, Barlow, Hebb, rate coding, the subtraction paradigm — the record has held. Where he reported hardware specs, the numbers got beaten, but not by tuning the instruments he named. 7T cortical‑layer fMRI is not a polished 1.5T scanner; it is a different machine at a different field strength aimed at a different scale of question. OPM‑MEG is not a faster cryogenic MEG; it is a room‑temperature atomic sensor array that killed the helium dewar.

Kurzweil’s one genuinely load‑bearing quantitative claim — the 10¹⁴ ops/sec brain compute ceiling — rested on treating the neuron as a ~200 Hz switching device. The 2021 single‑neuron‑as‑deep‑network result suggests the ceiling may be several orders of magnitude higher. For a book arguing that substrate‑level brain simulation is near, the neuron becoming a small network is the expensive kind of wrong: not about timelines, about what the units of the budget are.

The 2005 neuroscience textbook held. The 2005 neuroscience instruments, and the 2005 accounting unit, didn’t.

Method note

Evidence for this scorecard came from our local index of U.S. patent grants and pre‑grant publications (9.3 million documents), our index of roughly 357 million scholarly works from OpenAlex, and targeted web research into recent 7T fMRI, OPM‑MEG, and brain‑computer‑interface literature. Counts were established by full‑text search on patent and paper abstracts; specific claims were verified by reading the cited papers and patents directly. Patent numbers named above refer to issued U.S. patents; single‑neuron‑as‑deep‑network results come from Beniaguev, London, and Segev, Neuron, 2021.