Kurzweil Scorecard: The Brain Got Reverse-Engineered Backwards

In 2005, defending himself against critics, Ray Kurzweil laid out a careful program for understanding intelligence: map the brain’s circuitry, extract its computational tricks, and port them into machines. Reverse-engineer the visual system. Build spiking-neuron hardware big enough to run those circuits in real time. He was right that it would happen. He had the direction almost exactly backwards.

The machines learned to see first. Then they taught us how the brain does it. And by late 2025, a computer model of the visual cortex was accurate enough to reach into a living monkey’s brain, pick which neurons to stimulate, and bias what the animal chose to look at. The reverse-engineering arrived — but the artificial system led, and the biology followed.

The predictions

These four claims all come from the same place: The Singularity Is Near, Chapter Nine, “Response to Critics.” Kurzweil was answering skeptics who argued the brain was either too simple to matter (just read the genome) or too complex to ever copy. His rebuttal was a set of testable assertions about scale and method.

First, an arithmetic claim. Kurzweil argued that the human brain has roughly 100 billion neurons and about 10^14 connections, requiring on the order of 10^16 bits to specify, at least a million times more information than the uncompressed 6-billion-bit genome (ch. “Response to Critics”). From that he drew a conclusion: understanding intelligence requires reverse-engineering brain structure and dynamics, not reading the genome alone.

Then two forward-looking bets. That spiking-neuron approaches would soon make it possible to simulate very large neuronal networks in real time, and that reverse-engineering the primate visual system would be a long-term project progressively incorporating the computational tricks that biological vision uses — a prediction he credited in part to the French vision scientist Simon Thorpe and his work on fast feed-forward recognition.

Where we actually are

The bit-budget arithmetic holds — and connectomics is now measuring it. Kurzweil’s numbers were sound and remain so. The genome is about 3.2 billion base pairs; at two bits each, that’s the 6.4 billion bits he cited. Modern neuron counts put the human brain near 86 billion cells, with synapse estimates in the 10^14 to 10^15 range — comfortably the million-to-one ratio he claimed. What’s changed since 2005 is that we can now weigh the connectome directly instead of estimating it. In May 2024, the Lichtman lab at Harvard and Google’s connectomics team published a reconstruction of a single cubic millimeter of human cortex — “A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution” (Science, 2024; 288 citations and counting). That one speck contained roughly 57,000 cells and 150 million synapses, and it took 1.4 petabytes to store. Extrapolate a cubic millimeter to a whole brain and Kurzweil’s 10^16-bit figure stops looking like rhetoric and starts looking conservative. Verdict: verified.

Large spiking networks now run faster than the brain itself. Kurzweil expected real-time simulation of very large spiking networks “soon” — he was writing for the 2010s. The capability exists, and then some. In April 2024, Intel switched on Hala Point at Sandia National Laboratories: 1,152 Loihi 2 chips, 1.15 billion neurons and 128 billion synapses across 140,544 neuromorphic cores, in a chassis the size of a microwave drawing 2,600 watts. Intel reports it can run its full neuron capacity 20 times faster than a human brain, and up to 200 times faster at lower loads. The patent record tracks the same arc. Filings mentioning neuromorphic hardware climbed from a handful a year before 2015 to 70–80 annually since 2019, led by IBM, Samsung, HRL, and Intel. The inventions are concrete: US 11,580,366 describes an event-driven architecture of interconnected “core circuits,” each an array of digital neurons wired through digital synapses along axon and dendrite paths — silicon that only computes when a spike arrives. US 12,468,927, granted recently, claims an expandable neuromorphic circuit whose neuron and synapse arrays can be tiled to grow the network. This is exactly the substrate Kurzweil described. Verdict: ahead of schedule.

And yet — here’s the first crack — neuromorphic hardware did not become the road to AI. The breakthroughs of the last decade ran on GPUs executing dense, non-spiking deep networks. Hala Point is a research instrument for studying brain-like efficiency, not the engine behind the systems people actually use. Kurzweil got the capability right and the relevance wrong.

The visual system got reverse-engineered — in reverse. This is where the batch turns strange. Kurzweil pictured a patient, biology-first project: study the primate ventral stream, learn its tricks, and gradually build them into machines. What happened instead inverted the order of discovery. In 2014, Daniel Yamins and James DiCarlo at MIT showed that a deep convolutional network trained purely to recognize objects — never shown a single brain measurement — spontaneously developed internal representations that matched the firing patterns in monkey inferotemporal cortex better than any model neuroscientists had hand-built. “Deep Neural Networks Rival the Representation of Primate IT Cortex for Core Visual Object Recognition” ran in PLOS Computational Biology that year. The trick wasn’t extracted from biology and ported to silicon. It emerged in silicon, and then turned out to match biology.

Kurzweil himself, in The Singularity Is Nearer (2024), tells a version of this without quite naming the inversion. He recounts how in October 2014 the MIT vision expert Tomaso Poggio called describing the content of an image “one of the most intellectually challenging things of all for a machine to do,” needing “another cycle of basic research” — at least two decades out. “The very next month,” Kurzweil writes, “Google debuted object recognition AI that could do just that.” The field had been quietly reverse-engineered by gradient descent while the experts were still estimating timelines.

Since then the relationship has flipped entirely. Artificial networks became the reference models of the visual brain — DiCarlo’s lab runs Brain-Score, a public leaderboard ranking which ANN best predicts real neural and behavioral data, with architectures like CORnet-S and ResNet near the top. And in October 2025, a team including Martin Schrimpf pushed it to the logical endpoint. In “Model-Guided Microstimulation Steers Primate Visual Behavior,” they used a computational model of the ventral stream to choose where to electrically stimulate two macaques’ cortex — and the stimulation shifted the animals’ perceptual choices, with per-site model predictions strongly correlated to actual behavior. The model was a good enough copy of the visual system to drive it. They note the approach points toward visual prosthetics that could induce complex visual experiences. Verdict: ahead of schedule, wrong mechanism.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern in this batch is consistent and instructive. Kurzweil was excellent on scale and direction, and unreliable on mechanism. The bit budget was right. Real-time spiking networks arrived. The visual system did get reverse-engineered, faster than his “long-term” hedge implied. Every destination on his map was reached.

But he assumed the route would run biology-first: understand the brain, then build the machine. The actual route ran capability-first. We built systems that worked, discovered after the fact that they resembled the brain, and then used those systems as our best theory of how the brain works — to the point of steering a living one. The genome-versus-connectome debate he was refereeing turned out to be beside the point, because the path to machine vision went through neither. It went through a lot of labeled images and a lot of GPUs.

That is the recurring failure mode in this kind of forecasting, and it’s a generous one. The honest skeptic in 2005 would have said reverse-engineering the visual cortex was impossibly hard. Kurzweil said it would happen and was right. He just expected us to copy the brain to build the machine. Instead we built the machine, and it handed us the brain.

Method note

Verdicts here draw on a full-text index of roughly 9.3 million US patents and 357 million scientific papers, queried for neuromorphic hardware, connectomics, and computational models of vision and filtered by citation impact and year; on the actual claims of named patents; and on the published findings of the papers cited, read this session. Web searches confirmed current figures for Intel’s Hala Point, the FlyWire and H01 connectome projects, and the 2025 microstimulation work. Every number is from a source accessed while writing. Kurzweil’s predictions are paraphrased from The Singularity Is Near (2005), with restatements quoted from The Singularity Is Nearer (2024).