Kurzweil Scorecard: Brain Barriers and Silicon Nerves

In 2005, Ray Kurzweil placed three small bets buried in the neurotech chapters of The Singularity Is Near. A pair of gut peptides, zonulin and zot, would be the key that pops open the blood-brain barrier on demand. A silicon chip growing neurons on its surface, built by Infineon with the Max Planck Institute, would mature into the standard interface between nervous systems and electronics. And a line of visual-recognition models from Tomaso Poggio, James DiCarlo, and Christof Koch would eventually let engineers pipe images directly into a brain.

Twenty years later, all three outcomes are, in some form, real. Two arrived by mechanisms Kurzweil never named. One arrived exactly on the research path he identified — and is in a human patient’s head right now.

What Kurzweil actually said

Kurzweil’s BBB claim appeared in the nanobots section: “Proteins zonulin and zot have been discovered that react with brain receptors to temporarily open the blood-brain barrier at select sites” (ch. “Scanning Using Nanobots”). He used it to motivate a later argument that medical nanobots could enter the brain through a chemically opened door.

On the hardware side: “An Infineon chip developed with the Max Planck Institute allows neurons to grow on a substrate providing direct contact between nerves and electronic sensors and stimulators” (ch. “Interfacing the Brain and Machines”). That was a 2005 statement of fact, citing the Fromherz-group Neuro-Chip announced at the 2003 International Solid-State Circuits Conference — a 1 mm² array with 16,384 field-effect transistors reading from cultured neurons at 6 kHz.

On the visual side: “Visual object-recognition models being developed by Tomaso Poggio, James DiCarlo, and Christof Koch could eventually be used to transmit images directly into the brain” (ch. “Interfacing the Brain and Machines”). That was the boldest of the three — a prediction with a target horizon somewhere in the 2020s.

In The Singularity Is Nearer (2024), Kurzweil returned to the visual-recognition story by reporting what happened to that lab’s program. He noted that “in October 2014 Tomaso Poggio… said, ‘The ability to describe the content of an image would be one of the most intellectually challenging things of all for a machine to do. We will need another cycle of basic research to solve this kind of question.’ Poggio estimated that this breakthrough was at least two decades away. The very next month, Google debuted object recognition AI that could do just that.” The timeline compressed by two decades in thirty days.

Where we actually are

The blood-brain barrier didn’t yield to zonulin

Zonulin exists. Larazotide, a zonulin-receptor-modulating peptide, has moved through clinical trials — including a 2024 phase 2a study in children with post-COVID multisystem inflammatory syndrome where it accelerated gastrointestinal recovery and viral antigen clearance (Science Translational Medicine, 2024). But the clinical story played out at the gut lining, not at the brain’s vasculature. The only zonulin-related patent granted in 2025 in our index — US 12,194,065 — is about metabolic detoxification.

The outcome Kurzweil wanted — targeted, temporary BBB opening at select sites — arrived through two different paths.

The first is focused ultrasound with injected microbubbles. In January 2024, a West Virginia University team reported in the New England Journal of Medicine that they had used MR-guided focused ultrasound to open the BBB over six monthly aducanumab infusions in three Alzheimer’s patients. Amyloid-β levels in the targeted regions fell 48–63% relative to untreated contralateral regions. The barrier closed within three days each time. Our clinical-trials index shows 23 ongoing or completed ultrasound-plus-BBB-opening studies, including a 2024 Science Translational Medicine report of neuronavigation-guided focused ultrasound opening the barrier for chemotherapy delivery in pediatric diffuse midline glioma.

The second is engineered AAV capsids. US 12,398,181, granted August 2025, claims an AAV capsid protein with one of four short targeting sequences — TVSALFK, TVSALK, KLASVT, or KFLASVT — inserted between amino acids 588 and 589 of AAV9 VP1. Those seven residues turn a virus that normally bounces off the BBB into one that crosses it. US 12,281,335 (April 2025) and US 11,981,705 (May 2024) extend the approach to gene-therapy and immunotherapy cargo for brain cancers. Across the patent corpus, 118 filings since 2020 describe some form of BBB crossing or opening technology. None of that looks like zonulin.

Verdict: Wrong mechanism. The predicted outcome — routine, temporary, spatially selective BBB opening — has arrived. The specific molecular lever Kurzweil named for it has not.

The Infineon/Max Planck chip had descendants

The 2005 claim was correct at the time — the Fromherz-group Neuro-Chip was a real device. What Kurzweil didn’t spell out was what would happen once that playbook got picked up by the rest of the field.

The direct lineage is now a commercial industry. Switzerland’s MaxWell Biosystems ships CMOS high-density microelectrode arrays with 26,400 electrodes per chip and 1,024 simultaneous recording channels. Italy’s 3Brain sells systems with 4,096 channels. Axion BioSystems does the same for pharmaceutical screening.

The more consequential descendants are the ones that left the petri dish. US 11,944,822 (April 2024) is Neuralink’s patent on high-density cylindrical packaging — fitting the electronics for 1,024 thread-based electrodes into an implantable enclosure the size of a coin. US 12,248,629 (March 2025) adds static and dynamic multiplexing. US 12,369,863 (July 2025) describes on-chip neural-signal compression paired with short-range wireless telemetry, so the implant filters and streams cortical activity without a wired tether.

Two human-implant stories now depend on this class of technology. Neuralink’s first patient, Noland Arbaugh, received a 1,024-channel implant in January 2024 and, as of mid-2025, uses it to control a cursor, play Mario Kart, and run his household electronics through thought alone. The company reported 21 enrolled participants worldwide by late 2025. The second is Precision Neuroscience, whose US 12,324,910 (June 2025) takes a different path — a conformal flexible electrode array threaded along the cortical surface into ventricular or endovascular spaces, stimulating visual pathways without penetrating brain tissue.

Across the patent corpus, filings on brain-computer or brain-machine interfaces rose from 2 per year in 2010 to 36 in 2025, with 193 filings since 2020. 113 clinical trials currently involve BCI implants.

Verdict: Ahead of schedule. A 2005 lab curiosity — neurons growing on a silicon substrate — has produced both the commercial CMOS-MEA industry and the intracortical implants now in human trials.

Visual models really did transmit images into brains

This is the prediction Kurzweil got most right, and it arrived through the research line he explicitly flagged.

In May 2019, Pouya Bashivan, Kohitij Kar, and DiCarlo himself published “Neural population control via deep image synthesis” in Science. They trained a deep neural network to model neurons in monkey V4 cortex, then ran the model in reverse to synthesize images that, when shown on a screen, pushed chosen V4 neurons to chosen firing rates — including beyond naturally occurring levels, with independent control of overlapping receptive fields. That is image-to-brain transmission via the natural visual pathway, driven by exactly the recognition-model architecture Kurzweil had pointed at.

The direct cortical version followed in December 2021, when Eduardo Fernández and collaborators published “Visual percepts evoked with an intracortical 96-channel microelectrode array inserted in human occipital cortex” in the Journal of Clinical Investigation. A 57-year-old woman who had been blind for 16 years received a Utah array in her V1. Over six months, she learned to identify letters and object boundaries from phosphene patterns evoked by multielectrode stimulation. Penetrating electrodes required one-tenth the current of surface electrodes for a percept. The paper has 207 citations. It is the first demonstration of patterned cortical visual input in a human.

Subretinal stimulation arrived in parallel. The Pixium Vision PRIMA system — a 2×2 mm, 30 μm thick photovoltaic chip with 378 electrodes, now owned by Science Corporation — received FDA breakthrough designation in March 2023. A 2024 Ophthalmology Science report on four-year follow-up showed a 32-letter improvement in visual acuity in AMD patients; 27 of 32 implanted patients regained reading ability within a year. Not quite “images directly into the brain” — the retina is still the transducer — but the first durable restoration of letter-reading vision from a silicon chip.

Verdict: On track. The timeline is correct to within a few years. The DiCarlo-Poggio-Koch lineage is the specific intellectual path that delivered the result. The resolution is coarse — letters, object boundaries, phosphenes — not yet “images” in the everyday sense. But the mechanism and the people are the ones Kurzweil nominated.

The scorecard

What Kurzweil missed, and what he nailed

The pattern across this small batch is unusually clean. When Kurzweil identified a research program — a lab, a method, a class of technique — he tended to be right about where it would lead. The DiCarlo-Poggio-Koch visual-recognition agenda really did produce image-driven neural population control and human cortical visual prostheses. The Fromherz-Infineon neuron-on-substrate program really did scale into the chips now inside Noland Arbaugh’s motor cortex.

When Kurzweil named a specific molecule to solve a general problem, he was more likely to be wrong. The BBB really did get pried open on demand — but by acoustic microbubbles and engineered virus coats, not peptides targeting brain tight junctions. The outcome landed. The mechanism swapped.

There is a forecasting lesson in that split. Betting on the research community, the lab bench, and the broad architectural vector tends to survive contact with reality. Betting on a named candidate molecule tends not to — drug-discovery pipelines are unforgiving, and biology almost always has more paths into the same room than any single 2005 paper imagined.

Method note

Predictions were drawn from extracted statements in The Singularity Is Near (2005), cross-referenced with The Singularity Is Nearer (2024). Verdicts are based on full-text search of roughly 9.3 million US patents and 357 million scholarly works for each prediction’s key terms and named entities, targeted reading of the most-cited papers and most-recent patents surfaced by those searches, and web research for clinical-trial status and product launches. Every named patent number, paper DOI, citation count, and product claim was pulled from primary sources accessed this session.