Kurzweil Scorecard: The Threshold Unit Ate the Action Potential

In 1939 the biologists Alan Hodgkin and Andrew Huxley sketched a theory of how a nerve’s action potential propagates. In 1952 they measured it on a squid axon and wrote down four coupled differential equations that are still on the wall of every computational neuroscience lab. Nine years before that measurement, in 1943, Warren McCulloch and Walter Pitts had already cheated. They proposed a stripped-down neuron — sum the weighted inputs, fire if you cross a threshold, otherwise stay quiet — and proved you could build logic gates out of them.

Two neurons, two intellectual lineages, both still alive in 2026. One won so completely that nobody patents it anymore. The other survives in a research niche that does about half a percent of the revenue of the people who ignored it.

This is the scorecard for the foundations Ray Kurzweil pointed at in The Singularity Is Near (2005) when he argued that “early neuron models led to the field of connectionism, perhaps the first self-organizing paradigm introduced to computation” (ch. “Trying to Understand Our Own Thinking”).

What Kurzweil claimed

Three statements from the chapter, paraphrased close to the source:

1. “A. L. Hodgkin and A. F. Huxley described the axon’s action potential theory in 1939 and measured an action potential on squid axons in 1952.”
2. “In 1943 W. S. McCulloch and W. Pitts developed a simplified neural-net model with synaptic weights and a nonlinear firing threshold that motivated decades of artificial neural-net research.”
3. “These early neuron models led to the field of connectionism, perhaps the first self-organizing paradigm introduced to computation.”

The first two are history; we don’t grade Kurzweil on the historical record. What’s interesting is the implicit forecast wrapped around them: that both lineages would feed forward into computational substrates capable of thinking, and that connectionism — neurons learning patterns from data rather than executing symbolic rules — would furnish the road to general intelligence.

By the time Kurzweil wrote The Singularity Is Nearer (2024), he had already conceded the score on his own terms. “Connectionist approaches to AI were largely ignored until the mid-2010s, when hardware advances finally unlocked their latent potential,” he wrote. He compared the field to “the flying-machine inventions of Leonardo da Vinci — prescient ideas, but not workable until lighter and stronger materials could be developed.” He pegged the computational price-performance gain from 1969 to 2016 at “a factor of about 2.8 billion.” It wasn’t biology that paid off. It was hardware.

Where we actually are

The McCulloch-Pitts neuron won the production economy. A modern transformer is, mechanically, a stack of weighted sums followed by a nonlinearity. The most common nonlinearity in deep learning today — the rectified linear unit, or ReLU — is the McCulloch-Pitts threshold neuron with the cliff edge filed off into a ramp. Every GPT, every Claude, every Gemini call runs through trillions of these per inference. The architecture won so decisively that filing a patent on the activation function is no longer something serious teams do. Across the last eight years of US patent text we see only fourteen grants whose titles or abstracts pivot on “rectified linear unit” — fewer than a single quarter’s worth of new neuromorphic patents at the field’s current pace. Foundational primitives don’t get patented; only the wrappers around them do.

The Hodgkin-Huxley neuron survived in a niche. Our patent corpus shows steady growth in “neuromorphic” filings — 2 grants in 2010, 79 in 2021, holding around 56–71 a year through 2025. Spiking neural network filings track a similar curve but lower. Compare that to “self-supervised learning” — the data-hungry pre-training trick that quietly carries every modern foundation model — which went from 2 patents in 2020 to 66 in 2025. The pure-deep-learning direction is outrunning the bio-faithful direction roughly two to one in patent volume, and by an order of magnitude in commercial revenue.

The dollar gap is wider than the patent gap. Outside analyst estimates put the 2025 neuromorphic chip market somewhere between $43 million and $480 million. NVIDIA’s data-center segment alone runs an order of magnitude larger every quarter. Brain-inspired-but-not-biologically-faithful won — and won big.

The hybrid is where the field actually lives. IBM’s US 11,636,317, granted April 2023, deploys long-short-term-memory cells on a spiking neuromorphic core. Claim 1 describes neurons with a “membrane potential” — pure Hodgkin-Huxley vocabulary, action-potential lineage — gating an LSTM memory cell. Claim 2 then specifies: “the memory cell comprises a rectified linear unit.” That sentence is the truce. The biologically faithful neuron does the routing; the McCulloch-Pitts descendant does the math.

The recent hardware launches tell the same story. Intel’s Hala Point, deployed at Sandia National Laboratories and built from 1,152 Loihi 2 chips, delivers up to 20 quadrillion operations per second; Intel positions it as a research substrate, not a commercial product, and accesses route through the Intel Neuromorphic Research Community. IBM’s NorthPole, unveiled in 2023, gets cited as 25× more energy-efficient than conventional systems for specific inference workloads. BrainChip launched the Akida Cloud in August 2025 to give developers remote access to its second-generation spiking processor. None of these have made a dent in the training side of foundation models — all are pitched at edge inference, where energy budgets are tight and millisecond-scale latency matters.

The literature concedes the same point. The most-cited recent spiking-network paper in our literature corpus — Eshraghian and colleagues’ “Training Spiking Neural Networks Using Lessons From Deep Learning” (Proceedings of the IEEE, 2023, 613 citations) — is, by title, the field admitting it has to import the toolkit. Backprop-through-time, gradient surrogate methods, and gradient-descent training were developed for the un-biological McCulloch-Pitts descendants. Spiking networks now use them too.

“Self-organizing” is the part Kurzweil got most wrong, and most importantly wrong. Modern deep learning is not self-organizing in any meaningful sense. GPT-style models are trained on curated text dumps; vision models are trained on human-labeled or human-clicked datasets; the post-training step is reinforcement learning from human feedback. The closest cousin to self-organization is self-supervised pre-training, where the model predicts masked tokens or missing patches — and even that requires human-built data pipelines, training objectives, and reward functions. Self-supervised learning paper volume in our corpus went from 5 in 2016 to nearly 3,000 in 2025. That curve is real. But Kurzweil’s “self-organizing” implied autonomy of goal, not just absence of labels.

The scorecard

What Kurzweil missed, and what he nailed

He nailed the direction. Connectionism did inherit the future of AI. The 2.8-billion-fold hardware improvement he flagged in retrospect did, in fact, do the unlocking. And the McCulloch-Pitts neuron survives, mostly unmodified at its core, inside every state-of-the-art model.

What he missed was that biological fidelity is not, on this evidence, a precondition for intelligence at the level we care about. The neuron model that won is the one McCulloch and Pitts admitted was a deliberate simplification. The one that more accurately captures what a real cortical neuron does — Hodgkin-Huxley’s membrane voltage dynamics and its descendants — exists today as a parallel ecosystem of edge-inference chips, doctoral dissertations, and review papers titled like apologetics. The patent record shows IBM, Samsung, and Intel all hedging their bets, filing in both lineages, building hybrids. But the production money flows to the simpler, less brain-like model.

The lesson for forecasting is that “biology will win” is a tempting trap. Aircraft do not flap. Computer vision does not saccade. And general intelligence, so far, does not spike — it multiplies and accumulates, rectifies, and softmaxes. The brain remains the existence proof. It is not turning out to be the blueprint.

Method note

Patent counts are drawn from a US patent corpus covering published applications and granted patents through April 2026. Literature counts are from a 357-million-paper bibliographic index covering 2005–2026, filtered for full-text search matches on the relevant terms. Web sources are listed below. The post quotes Kurzweil from The Singularity Is Near (2005, chapter “Trying to Understand Our Own Thinking”) and The Singularity Is Nearer (2024). Patent numbers and granted dates are cited as filed.

Sources:
– Intel Builds World’s Largest Neuromorphic System (Hala Point)
– Advancing Neuromorphic Computing With Loihi: A Survey of Results and Outlook
– NorthPole, IBM’s latest neuromorphic AI hardware
– BrainChip Akida — first commercial SNN AI chip
– Training Spiking Neural Networks Using Lessons From Deep Learning
– Fifty years of gating currents and channel gating
– ReLU activation function — modern guide
– Neuromorphic Chip Market — Fortune Business Insights
– AI Chip Statistics 2025 — NVIDIA share