Kurzweil Scorecard: The 2005 Receipts Cashed Out — and the Wave That Actually Hit Wasn’t in the Book

In October 2014, MIT’s Tomaso Poggio said teaching a machine to describe the contents of an image would take “another cycle of basic research” — at least two decades out. One month later, Google demoed exactly that. Kurzweil recounts the story in The Singularity Is Nearer (2024): “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.”

Poggio is a named source behind this batch of Kurzweil’s 2005 predictions, alongside Newell, Simon, Evans, and Ross King. Across ten interwoven claims about AI circa 2005 plus one forward bet on 2030s VR, the pattern is sharper than any single verdict. Kurzweil’s 2005 receipts all cashed out. But the wave that actually hit after 2017 didn’t look like any of the examples he named.

The predictions

These are Kurzweil’s “AI is already everywhere” claims from Chapter Nine (“Response to Critics”), marshaled against skeptics who thought AI had stalled in the 1990s. He argued the field had never stopped: fuel injection in cars, airport gate scheduling, spam filtering, game characters, Microsoft’s software, NASDAQ fraud detection, and face recognition were all running on AI by 2005. He cited specific landmarks: the 1957 General Problem Solver, Thomas Evans’s ANALOGY program, SpamBayes, a 2003 USC Information Sciences Institute translation system “romancing the Rosetta Stone,” and Ross King’s 2004 Nature paper on a robot scientist doing functional-genomic hypothesis generation on yeast. The one forward-looking piece: “Humans will spend increasing portions of their time in virtual environments and will be able to have any desired experience with anyone, real or simulated, in virtual reality” (ch. “Promise and Peril of GNR”), slated for the 2030s.

Where we actually are

AI everywhere — verified, then exploded past Kurzweil’s description. The 2005 claim cited fuel injection and airport scheduling. Fair. The post-2022 build-out is a different animal. Patents mentioning “large language model” went from 11 grants in 2021 to 443 in 2025 — a roughly 40x surge in four years. Reading the claims makes the pattern concrete: US 12,602,206 describes an LLM pipeline where input code is paired with a retrieved policy and schema, then rewritten by the model to conform. US 12,602,357 chains two prompts — one to generate a natural-language description of a data schema, another to use that description plus a target schema to emit preprocessing code. These aren’t research demos; they’re claims written for issue and enforcement. Kurzweil was right about AI being “everywhere around you every second of the day.” He could not have known how literal that phrase would become twenty years later.

AI techniques moving from research to product — verified, lag collapsing. Kurzweil wrote that “many software applications, from search engines to games, routinely use AI techniques that were only research projects a decade earlier” (ch. “A Panoply of Criticisms”). In The Singularity Is Nearer, he describes the next mechanism: “Invented by Google researchers in 2017, this mechanism has powered most of the enormous AI advances of the past few years.” Attention paper 2017; GPT-2 2019; GPT-3 2020; ChatGPT to 100 million users by early 2023 — two months, not a decade. The research-to-product lag has flipped: products now ship faster than the peer-reviewed literature can document them.

Bayesian spam filtering — verified, superseded. SpamBayes was real; Bayes was the serious answer in 2005. Gmail today processes more than 15 billion unwanted messages daily and blocks over 99.9% of spam, phishing, and malware. The mechanism is no longer Bayes alone. Google’s RETVec detects deliberately obfuscated text using homoglyphs and LEET substitution that word-probability filters miss entirely. Bayes is a legacy layer inside a deep-learning stack. The 2005 claim was correct when made; the problem was solved, just not by the tool Kurzweil named.

Machine translation — verified, then paradigm replaced. The 2003 USC ISI system was statistical phrase-based. Google ran on statistical methods from 2007 until November 2016, when GNMT switched to neural translation and cut errors on major language pairs by more than 60% in a single release. A year later the attention paper reworked the foundation again. Our database shows patents on “neural machine translation” going from zero in 2014 to 21 grants in 2022, continuing through 2025. The actual trajectory — statistical → neural → transformer, each transition compressing the prior generation into obsolescence — is the story the 2005 example doesn’t tell.

Robot scientist — verified and scaled. King’s 2004 Adam did functional-genomic hypothesis generation on yeast. Adam was followed by Eve (drug repurposing) and now Genesis, which King estimates at about £1 million to build. The broader wave is the self-driving laboratory. A 2025 Nature Chemical Engineering paper reports dynamic flow experiments with at least an order-of-magnitude gain in data collection over prior SDLs. Berkeley’s A-Lab integrates target selection, ML-driven recipe generation, robotic solid-state synthesis, and active learning end-to-end. ChemAgents runs an LLM-based multi-agent hierarchy coordinating Literature Reader, Experiment Designer, Computation Performer, and Robot Operator agents. Adam’s grandchildren, doing science at scale.

Face recognition — verified, Poggio’s caveat aged badly. Kurzweil paraphrases Poggio’s distinction between identification (commercially deployed) and categorization (harder). The database bears out the first half: face-recognition-and-neural-network patents climbed from single digits a year in the mid-2000s to 17 grants in 2024. US 12,573,238 (2025) describes biometric facial recognition paired with a liveness detector scoring live faces vs. 2D/3D spoofs via deep convolutional networks — a 2005 technology plus a 2020s anti-spoofing layer. The second half is where Poggio’s 2014 estimate embarrassed itself and Kurzweil gets to take a victory lap: image captioning, flagged as two decades out, fell within a month.

VR, “any desired experience” by the 2030s — behind schedule. Meta’s own developer docs recommend a “Goldilocks” session length of 20–40 minutes. Roughly 13% of US households own a VR device, using it 30–45 minutes, 2–4 times a week. Global headset sales hit 65 million cumulative units in 2025. These are not the numbers of a population “spending increasing portions of their time” in virtual environments. They are the numbers of a niche medium that has repeatedly failed to break out. The social moment Kurzweil described — arbitrary experiences with arbitrary people, real or simulated — is not visibly on a ramp to 2030s delivery. The cultural energy he expected VR to absorb is being absorbed instead by conversational AI and video-first platforms. Wrong mechanism, behind schedule, or both.

AI Winter, GPS, ANALOGY, industry applications — historical citations Kurzweil used to establish the field’s pedigree. Accurate as history, no longer predictive. The 2005 industry survey (medicine, customer service, education, manufacturing, defense, finance, fraud) has been made quaint by the LLM wave. Saying “AI is in industry” in 2026 is like saying electricity is in industry.

The scorecard

What Kurzweil got right, and what the pattern says

The 2005 examples were all valid receipts. Spam filters worked. The robot scientist ran. Machine translation was getting better. Face recognition was in production. Kurzweil was not the hype merchant he gets caricatured as — he was, if anything, careful in Chapter Nine to anchor his claims to specific papers, specific products, specific years.

The pattern is what’s interesting. Every 2005 example was eventually replaced by a technology Kurzweil did not name. Bayes gave way to transformer-based filters. Phrase-based translation gave way to sequence-to-sequence and then attention. Adam gave way to LLM-orchestrated self-driving labs where the model writes the experiment plan, not just the regression. Direction of the arrow — AI penetrates more domains, more deeply, every year — was exactly right. The mechanism that carried the weight after 2017 was the transformer, and the transformer is not in The Singularity Is Near. It’s in the 2024 update, written retrospectively.

The VR miss is the mirror: the one forward-looking claim, and the one where the population data refuses to cooperate. Kurzweil expected social presence to migrate into simulation. It migrated into chat windows. People who were supposed to be in VR with their avatars are on video calls and talking to models. Same underlying appetite — synthetic presence, conversational partners that aren’t physically there — routed through a completely different substrate.

The forecasting lesson: direction is easier than mechanism, and mechanism is easier than timeline. Kurzweil nailed direction across this batch. Mechanism he got partially — the transformer surprised him. Timeline on the one forward call is slipping. That is roughly the order of difficulty any honest forecaster should expect.

Method note

Patent counts come from our search over 9.3 million US patent documents; specific patent numbers read from granted text. Literature counts come from an index of ~357 million scientific works. The Singularity Is Nearer passages were pulled from a local copy. Web searches supplied current Gmail filter architecture, Google translation history, self-driving laboratory benchmarks, and Meta Quest usage. Every number came from a query or source accessed this session.