Kurzweil Scorecard: An Owl’s Brain, a Cubic Millimeter, and the Deadline That Slipped

In April 2024, Intel wheeled a six-rack box the size of a microwave oven into Sandia National Laboratories. Inside: 1,152 Loihi 2 chips wired into 1.15 billion silicon neurons and 128 billion synapses, drawing no more than 2,600 watts. Intel pegged the scale at roughly the brain of an owl.

This matters because Kurzweil wrote in 2005 that “by the 2020s, with at least a millionfold increase in computational power and vastly improved scanning resolution and bandwidth, we will have the tools needed to model and simulate the entire brain” (The Singularity Is Near, ch. “The Accelerating Pace of Reverse Engineering the Brain”). Hala Point is not the entire brain. It is hardware for roughly one percent of one, and a research system rather than a working simulation of any specific animal. The 2020s deadline is burning down while the brain-simulation community has delivered something narrower than Kurzweil promised — and while the AI field overtook it entirely via a path he barely anticipated.

This batch concerns reverse-engineering the brain: cortical chips, retina chips, cerebellum models, cocktail-party front ends, the hypothesis-and-test cortex, and the grand claim that understanding the brain would “vastly extend” human intelligence (ch. “Reverse Engineering the Brain: An Overview of the Task”). On the historical facts, Kurzweil is mostly right. On the forecasts, he is mostly behind or on the wrong mechanism.

Where we actually are

The whole-brain target slipped from 2020s to unknown. Europe’s Human Brain Project ran 2013 to September 2023 and concluded with an independent review noting “major contributions” to digital brain atlases, neuromorphic computing, and brain-inspired AI — but no working whole-brain simulation. Its EBRAINS infrastructure continues as a platform. Switzerland’s Blue Brain Project wound down in December 2024; its deliverable was a “reference digital model of the entire mouse brain” integrating multiscale data across 70 million neurons — a structural atlas, not a living simulation. Henry Markram’s group spun out the Open Brain Institute in January 2025 to continue the work as an independent nonprofit.

The deepest structural map so far — the MICrONS consortium’s paper in Nature in April 2025 — covers a single cubic millimeter of mouse visual cortex: about 200,000 cells, 523 million synapses, 4 kilometers of axons, 1.6 petabytes of data. Beautiful science. It is also roughly 1/350th of one mouse brain, from one animal, at one moment. Kurzweil’s 2005 roadmap assumed that by now we would have the scanning, the compute, and the algorithmic understanding to do the full organ. Two out of three is a generous read.

Verdict on entire brain modeling possible by 2020s: Behind schedule. The compute arrived. The scanning arrived for cubic-millimeter-scale pieces. The integration into a working emulator did not.

The chip predictions are the most interesting column. Kurzweil’s two 2005 chip claims — Carver Mead’s analog retina–optic-nerve chip, and the MIT/Bell Labs integrated circuit with 16 excitatory and 1 inhibitory silicon neurons — were true when he wrote them, and they were seeds. Mead’s native-analog-mode approach is the direct ancestor of today’s event-based “silicon retina” vision sensors that Prophesee, Sony, and Samsung now ship at volume, and of research filings like US 11,501,432 (granted 2022) for a spiking retina microscope.

The MIT/Bell Labs “16+1” chip has been swamped. IBM’s TrueNorth (2014) hit 1 million neurons. Intel’s Loihi 2 (2021) powers Hala Point (April 2024) at owl scale. Dresden’s SpiNNaker 2, funded through the Human Brain Project, began 2024 deployment with 5.2 million ARM cores at roughly 10× better neural-simulation efficiency per watt than its predecessor. Heidelberg’s BrainScaleS-2 runs analog neural dynamics 1,000× faster than biological time. Patents fill in the plumbing: US 11,593,623 (2023) claims a spiking-neural-network accelerator with per-neuron address generation and external memory; US 12,142,263 (November 2024) describes a self-learning neuromorphic acoustic front end for speech recognition. Neuromorphic-chip literature grew from 56 papers in 2015 to 413 in 2025.

Verdict on MIT/Bell Labs cortical circuit chip and Mead retina chip: Verified historical claims, both extended far beyond Kurzweil’s footnote.

The cerebellum bet was more specific — and more vindicated. Kurzweil highlighted Javier Medina, Michael Mauk, and colleagues at the University of Texas Medical School for building a bottom-up cerebellum simulation with more than 10,000 neurons and 300,000 synapses across principal cell types. That work was real. A 2020 eLife paper, “Principles of operation of a cerebellar learning circuit” (10.7554/eLife.55217), extended the lineage: circuit-level cerebellum models now make predictions that match motor-learning experiments. In The Singularity Is Nearer (2024), Kurzweil describes the cerebellum as “largely of small and simple modules… a feed-forward structure” whose understanding feeds the AI field.

Verdict on the Texas cerebellum claim: Verified and extended.

The cocktail-party prediction is a “right direction, wrong mechanism” case. Lloyd Watts’s biologically inspired auditory model did demonstrate speaker isolation as a speech-recognition front end in the early 2000s. But the cocktail-party problem as industry now solves it runs through transformer- and Mamba-based separation networks: Wavesplit and SepFormer post SI-SNRi of 22.3 dB on WSJ0-2mix; diffusion vocoders layered on deterministic separators pushed the state of the art further in 2023–2024. None of that came from a cochlear model. The insight carried; the architecture was replaced.

Verdict on Watts cocktail party: On track on outcome, wrong mechanism on how.

The hypothesis-and-test cortex is aging well. Kurzweil’s 2005 claim that the cortex “guesses and verifies” features against sensory input mapped directly onto what is now called predictive coding. The 2023 Nature Human Behaviour paper “Evidence of a predictive coding hierarchy in the human brain listening to speech” (10.1038/s41562-022-01516-2, 268 citations) is a capstone: large language models reproduce cortical activation patterns during speech comprehension precisely when their internal representations align with hierarchical predictive-coding dynamics. The cortex does appear to operate something like that.

Verdict on hypothesis-and-test visual strategy: On track / verified.

The big forecasts are where the scorecard gets uncomfortable. Kurzweil predicted “detailed and implementable replicas… for all brain regions” by the 2020s (ch. “Peeling the Onion”). We have implementable replicas for the retina, parts of the cerebellum, V1, and early auditory cortex. We do not have them for the hippocampus, basal ganglia, thalamus, or most of the neocortex — at least not in the “run it and get behavior” sense. Behind schedule. He predicted the “nonbiological portion of human intelligence will predominate” by the 2030s — a decade we have not entered. Kurzweil still holds to 2029 for human-level AI; in The Singularity Is Nearer he writes, “My expectation was that in order to pass a valid Turing test by 2029, we would need to be able to attain a great variety of intellectual achievements with AI by 2020. And indeed, since that prediction, AI has mastered many of humanity’s…” Too early to call on the 2030s statement, but trending ahead — for reasons mostly unrelated to brain reverse-engineering.

The upload Turing test — a “Jane Smith” test in which judges cannot distinguish a digital copy from the original person — has no candidates at all. Too early to call is charitable; nothing in the current pipeline leads there.

The scorecard

What Kurzweil missed and what he nailed

The consistent pattern across this batch is direction right, timeline wrong, mechanism often replaced. Kurzweil’s 2005 thesis was that understanding the brain would power the AI rise. Two decades later, the AI rise has plainly happened — the Metaculus consensus on strong AI briefly hit 2026 in 2022 — but the hardware that did the work was the GPU, not the cortical simulator, and the algorithm was the transformer, not the cortical column. Neuromorphic systems like Hala Point and SpiNNaker 2 are scaling, but as tools for neuroscience and edge inference, not the substrate of current frontier models.

Where Kurzweil nailed it: the compute curve, the predictive-coding architecture of the cortex, the extendability of Mead’s analog approach, and the vindication of the cerebellum-as-simple-modules thesis. Where he missed: he assumed the path to machine intelligence ran through biological fidelity. It ran around it. The brain atlas got built. The brain is not yet running.

Method note

This post compared Kurzweil’s 2005 predictions against United States patent grants and pre-grant publications, the OpenAlex scholarly-literature corpus, and open-web coverage of the Human Brain Project, Blue Brain Project, MICrONS, Hala Point / Loihi 2, SpiNNaker 2, BrainScaleS-2, and recent cocktail-party and predictive-coding benchmarks. Citations to The Singularity Is Near (2005) and The Singularity Is Nearer (2024) are by chapter; verbatim passages are from the 2024 edition. Patent numbers are granted US utility patents. Where an outcome is happening via a different mechanism than Kurzweil described, the scorecard notes both the outcome and the mechanism mismatch rather than collapsing them.