Kurzweil Scorecard: The Brittleness Argument Lost

In 2005, Jaron Lanier was the most quoted critic of The Singularity Is Near. His argument was straightforward: software is brittle. Real systems crash, hang, segfault, leak memory, and bring planes down. Scaling them up only multiplies the failure modes. Therefore the smooth exponential curves Kurzweil drew for AI capability could not survive contact with engineering reality. Twenty-one years later, the curves are still going up and the brittleness argument is the one that broke.

This batch covers five claims Kurzweil made in defense — that self-organizing methods escape the complexity ceiling, that machines can harness the same emergent principles that let biology recover from internal failure, that software is accelerating in capability even if more slowly than hardware, that failure rates would not stop AI from scaling, and that nonbiological intelligence will eventually exceed biological by a factor of trillions of trillions. Four of the five now look correct or ahead of schedule. The fifth is unfalsifiable on a 2026 timescale.

What Kurzweil actually wrote

In The Singularity Is Near, Kurzweil argued that “machines can and already do harness the same self-organizing and emergent principles found in biological systems” (ch. “A Panoply of Criticisms”), that “self-organizing methods increase system complexity without the brittleness associated with explicitly programmed logical systems” (ch. “The Criticism from Software”), and that “software is accelerating in effectiveness, efficiency, and complexity, though with a slower doubling time than hardware.” He extended this to a long-horizon claim — that “nonbiological intelligence will become trillions of trillions of times more powerful than biological intelligence as information-technology trends continue.”

In The Singularity Is Nearer (2024), Kurzweil restates the argument in plain language. He calls the failure mode of rule-based AI the “complexity ceiling” — the property that “when MYCIN and other such systems made a mistake, correcting it might fix that particular issue but would in turn give rise to three other mistakes that would rear their heads in other situations.” His thesis in 2024 is that connectionism solved this: he writes that transformers above 100 billion parameters “could suddenly answer questions on their own with intelligence and subtlety” — a property absent below that scale. The frame had not changed since 2005. The evidence had.

The connectionist landscape, by the numbers

The shift from explicitly programmed logic to self-organizing learned representations is the dominant fact of the past decade of computing. Patents tell the story: filings whose claims explicitly invoke “deep learning neural network” rose from 17 grants in 2017 to 373 in 2024 and stayed near that level through 2025. Filings invoking “generative adversarial network” — a self-organizing technique that did not exist when Kurzweil wrote the original book — went from a single 2014 grant to 178 in 2025. Patents naming “large language model” in their claims went from four in 2018 to 443 in 2025. None of these existed as a category when Lanier was arguing that software couldn’t scale past brittleness.

The scientific record is steeper. High-citation papers (≥50 cites) explicitly studying large language models grew from a single paper in 2018 to 579 in 2023. Wei et al.’s “Emergent Abilities of Large Language Models” (2022, doi:10.48550/arxiv.2206.07682) has 1,015 citations in the literature index alone — a paper whose central claim is that capabilities not present at small scale appear suddenly above a threshold. That is a 2024 academic restatement of Kurzweil’s 2005 prediction that emergent properties would arise from self-organizing systems.

The scaling itself is faster than Kurzweil claimed. Epoch AI estimates that frontier-model training compute has grown 5.3× per year since 2010. Going from GPT-3’s 3 × 10²³ FLOPs in May 2020 to Gemini Ultra’s 5 × 10²⁵ FLOPs in December 2023 is a hundredfold increase across three and a half years, well above Kurzweil’s then-stated doubling cadence for hardware. The first model estimated to use over 10²⁶ FLOPs was Grok-3, released in February 2025. The cost growth — 2.4× per year — is on track to push the largest training runs past a billion dollars by 2027.

Reading the actual inventions

Counting patents tells you the field is real. Reading the patents tells you what was built.

US 12,596,914, granted to a generative-adversarial neural-architecture-search system in April 2026, describes the method without metaphor: “a generative adversarial network comprising a generator and a discriminator receives a query for neural network architecture, the query including a search space. The generator generates a plurality of neural network architectures responsive to the received search space. The discriminator selects an optimal neural architecture.” That is, in plain terms, a population of candidate architectures that mutates and competes against a learned fitness function. It is the design pattern Kurzweil pointed to in 2005 — an evolutionary search wrapped around a self-organizing classifier — and it is now the way frontier AI labs design their own networks.

US 12,475,360 (“Distributed cellular computing system and method for neural-based self-organizing maps,” granted November 2025) describes a neuromorphic grid in which each cell is connected only to its direct neighbors and learns through unsupervised competition — Teuvo Kohonen’s 1982 self-organizing map cast as a hardware substrate. US 12,519,700 (January 2026) builds an entire telecommunications-network controller around a generator-discriminator pair that simulates the network and tunes itself through adversarial training. The “self-organizing” qualifier in both titles is not aspirational marketing. It is the working principle of the claim.

Outside the patent record, the most aggressive validation is DeepMind’s AlphaEvolve, announced in May 2025. AlphaEvolve wraps an evolutionary loop around two Gemini models — Gemini Flash for breadth, Gemini Pro for depth — and lets them rewrite candidate algorithms while a held-out evaluator scores the results. Across more than fifty open mathematical problems it rediscovered state-of-the-art solutions in 75% of cases and improved the best known solution in 20%. It found a way to multiply two 4 × 4 complex matrices using 48 scalar multiplications, beating Strassen’s 1969 algorithm — the first improvement in that setting in 56 years. In production at Google, its scheduling heuristics recover 0.7% of worldwide compute, and it sped up a kernel inside Gemini training by 23%. The system improved the AI training that produces it.

That is the literal closure of Kurzweil’s 2005 sentence about machines harnessing biological self-organization. A population of programs mutates, recombines, and is selected against an environment. The resulting code is then used to train the descendants of the system that wrote it.

What about the brittleness?

Lanier was not wrong that software fails. The Boeing 737 MAX MCAS system, which routed all of its inputs through a single angle-of-attack sensor and applied repeated nose-down inputs that pilots could not counteract, killed 346 people across two crashes. That failure was structural and rule-based — exactly the kind of explicitly-programmed logic Kurzweil’s 2005 book named as the brittle counterexample. It is also exactly the kind of system that has not scaled. The failure-prone code did not stop being failure-prone. It stopped being the dominant kind of code.

Frontier AI systems have plenty of failure modes — hallucination, reward hacking, distributional brittleness — but the systems are scaling through them rather than being stopped by them. Compute, parameters, citation counts, and patent grants all climb together while the underlying components break, leak, and surprise their users. Kurzweil’s claim was not that software would stop failing. It was that failures would not be the binding constraint. On that, he has been right so far.

The scorecard

What the pattern suggests

Kurzweil’s 2005 critics were arguing about the wrong thing. They were measuring brittleness at the line-of-code level — counting bugs in MYCIN, in CYC, in Windows NT. Kurzweil’s bet was that the substrate would change, and that the new substrate would be one in which “lines of code” was no longer the unit of complexity. He did not predict transformers. He predicted that something with their properties would arrive on roughly the right timeline.

The interesting forecasting lesson is what he got wrong about what the self-organizing technology would look like. He spent more pages on genetic algorithms than on attention mechanisms because attention mechanisms did not yet exist. The mechanism was different. The destination was where he said it would be.

The remaining open question is the trillions-of-trillions claim, which Kurzweil restates in 2024 by saying the doubling will continue “and will multiply our intelligence millions-fold.” On a 2026 timescale that is unfalsifiable. The compute trend — 5.3× per year since 2010 — could sustain it for another decade or run into thermodynamic walls before it gets close. Either outcome will be visible by 2035. Until then, the scorecard for this batch is a defensible win for Kurzweil and a clear loss for the brittleness school.

Method note

Counts came from a full-text search across 9.3 million U.S. patent documents and 357 million OpenAlex scholarly works, indexed locally. Patent abstracts and claims were read directly. Compute-trend numbers come from Epoch AI’s ongoing AI training-compute analysis. AlphaEvolve numbers come from DeepMind’s May 2025 announcement and the corresponding paper. Quotations from The Singularity Is Near and The Singularity Is Nearer were verified against the source text.