Kurzweil Scorecard: The Road to Strong AI

In March 2026, a startup called Eon Systems closed the loop on something that had never been done before: they ran a complete fruit fly brain — all 127,400 neurons, 50 million synaptic connections — inside a physics simulation, connected to a virtual body that could perceive and act. The first whole-brain emulation of any animal with a real nervous system. It took two decades of connectome mapping to get there. And it arrived just in time to illustrate the central irony of Kurzweil’s predictions about strong AI: the path he bet on is finally producing results, but the race was won years ago by a completely different horse.

What Kurzweil predicted

This batch of twelve predictions from The Singularity Is Near lays out Kurzweil’s roadmap for achieving and then scaling artificial general intelligence. The core thesis: “strong AI will be achieved by short-circuiting long evolutionary search through reverse engineering the human brain” (ch. “Can We Evolve Artificial Intelligence from Simple Rules?”). He predicted “human-level artificial intelligence will emerge in the 2020s as hardware and software for full emulation of human intelligence become available” (ch. “The Singularity Is Near”), with machines passing both the Turing test and “Edward Feigenbaum’s expert-scientist test in at least some disciplines around the same time” (ch. “Strong AI”). Then the feedback loop: “once machines can design technology as humans do, they will improve their own abilities in an accelerating feedback cycle that unaided human intelligence cannot follow” (ch. “The Singularity Is Near”). The whole arc culminates in “Epoch Five” — the merger of human and machine intelligence — around 2045.

Where we actually are

The reverse-engineering bet

Kurzweil’s mechanism for achieving strong AI was explicit: “short-circuit long evolutionary search through reverse engineering the human brain, while also using evolutionary methods” (ch. “Can We Evolve Artificial Intelligence from Simple Rules?”). In The Singularity Is Nearer, he maintained that “computers will be able to simulate human brains in all the ways we might care about within the next two decades or so” — while also acknowledging that “the trade-off between spatial and temporal resolution in brain scans is one of the central challenges in neuroscience as of 2023.”

The connectome field has made genuine progress. The scientific literature shows a steady climb from 4 papers per year in 2005 to a peak of 297 in 2020 on brain emulation, connectomics, and neural circuit reconstruction. The fruit fly connectome, completed in 2024 after years of electron microscopy work, was a landmark — three orders of magnitude more neurons than the only prior complete connectome (the C. elegans worm, 302 neurons, mapped in 1986). But the gap between a fruit fly and a human brain is another five orders of magnitude. Current estimates put mouse whole-brain simulation around 2034, human later than 2044.

Meanwhile, in the patent record, brain emulation and reverse-engineering grants grew modestly from 4 per year in 2000 to 74 in 2023. That’s real work — but it’s a rounding error next to the 401 US patents granted in 2024 for automated machine learning and neural architecture search alone. The money and the talent followed the path that was working, and that path was not neuroscience.

The mechanism that actually won

Large language models achieved Kurzweil’s predicted capabilities through brute-force statistical learning on text, not brain emulation. Training runs for frontier models now take 2-4 months on thousands of GPUs. Hardware efficiency has improved 30-50% since 2023. Anthropic reports that 70-90% of code for its next models is already written by Claude. Andrej Karpathy’s open-source AutoResearch project, released March 2026, ran 700 automated experiments in two days and produced an 11% training speedup on a small language model.

Kurzweil predicted that “once machines can design technology as humans do, they will improve their own abilities in an accelerating feedback cycle that unaided human intelligence cannot follow” (ch. “The Singularity Is Near”). That prediction is arriving in 2026 — not through brain-inspired architectures, but through language models writing and evaluating their own code. An ICLR 2026 workshop is dedicated entirely to recursive self-improvement in AI systems. The field has moved from thought experiment to deployed capability.

IBM leads in automated ML patents with 107 grants since 2020 (combining entity variants), followed by Capital One (62), Amazon (53), Google (36), and Microsoft (30). The presence of financial institutions — Capital One and Bank of America (51) — in the top five is telling: recursive optimization of machine learning pipelines is already a production concern, not a research curiosity.

The “AI education” prediction

Kurzweil predicted that “the education of future AIs will be compressed from the roughly twenty years required for humans to a matter of weeks or less” (ch. “Strong AI”). Current frontier model training runs take 2-8 months depending on scale — not weeks, but far closer to Kurzweil’s target than to twenty years. More importantly, techniques like knowledge distillation and few-shot learning mean that once a large model is trained, smaller specialized models can be derived in days or hours. The literature on training efficiency peaked at 1,399 high-citation papers in 2021 and remains a major research front.

The Feigenbaum test

Kurzweil predicted that “machines will pass Edward Feigenbaum’s expert-scientist test in at least some disciplines around the same time they pass the Turing test” (ch. “Strong AI”) — sustained expert-level dialogue in a specialized domain. Feigenbaum himself, honored at AAAI’s 40th conference in January 2026 on his 90th birthday, would recognize the irony: the expert system paradigm he helped create was sidelined by deep learning, but the goal he articulated — machines that can hold their own with domain experts — is being realized by a technology he didn’t anticipate. Current frontier models score above 90% on graduate-level science questions (GPQA) and can engage in sustained technical dialogue across multiple disciplines.

The computational capacity question

Kurzweil predicted that “in the early 2030s, the amount of nonbiological computation produced per year will roughly equal the total capacity of all living biological human intelligence” (ch. “Setting a Date for the Singularity”) — somewhere between 10^26 and 10^29 operations per second. Estimates of human brain equivalence range wildly, from 10^13 to 10^25 FLOPS, with a median around 10^15-10^18. Global computing capacity was estimated at roughly 10^20-10^21 FLOPS as of 2015. With AI-specific compute growing rapidly — frontier labs are building data centers an order of magnitude larger than anything before — the 10^26 target by the early 2030s is plausible for aggregate AI compute, if not for general-purpose computing.

The scorecard

What Kurzweil missed (and what he nailed)

This batch crystallizes the deepest pattern in Kurzweil’s forecasting: he was a better prophet of what than of how. He predicted human-level AI in the 2020s, and got it. He predicted machines that design their own successors, and it’s happening in 2026. He predicted the compression of AI training time, and frontier distillation is nearly there. He was right about nearly every destination.

But every route was wrong. Strong AI arrived through statistical learning on text, not brain reverse-engineering. Recursive self-improvement is happening through code generation, not neuromorphic computing. The software didn’t lag the hardware by a decade — if anything, the algorithmic insight (transformers, attention, scaling laws) arrived first, and the hardware is scrambling to keep up.

The fruit fly brain running in simulation at Eon Systems is a beautiful scientific achievement. It’s also a monument to the road not taken. The intelligence Kurzweil predicted didn’t need to understand the brain. It just needed enough text and enough GPUs.

Method note

This scorecard draws on 9.3 million US patent grants from USPTO bulk XML (full-text keyword indexed), 357 million scientific papers from OpenAlex (citation-filtered where noted), and current reporting from technology publications accessed via web search in April 2026. Patent assignee counts combine variant spellings. Brain emulation progress data is from the State of Brain Emulation Report 2025 and the Carboncopies Foundation. AI training statistics are from public disclosures by Anthropic, OpenAI, and Epoch AI. Compute estimates reference AI Impacts and Coefficient Giving analyses.