“The AI Revolution: The Road to Superintelligence (Part 1)” — @waitbutwhy

Date caveat

This piece is from January 2015. Pre-transformer (June 2017), pre-GPT-3 (2020), pre-ChatGPT (Nov 2022), pre-Anthropic-as-we-know-it. It predates every LLM artifact that shapes the 2026 conversation about AI. The specific predictions, capability examples, and timeline calibrations are 2015-museum-pieces — DO NOT cite this as a current technical reference, do not let its example set anchor any present-day capability claim, and do not pull its expert-survey AGI dates into 2026 strategy work.

What IS evergreen and worth citing: the public-narrative frameworks (ANI/AGI/ASI staircase, exponential-vs-linear projection, intelligence-as-substrate, train-station metaphor). These are the dominant scaffolds non-technical readers — and the agents trained on their writing — still use to think about AI’s trajectory. Cite for framework, never for fact.

Why this is in the vault

Urban’s three-tier staircase is the load-bearing public-mental-model for AI. Most readers of Sanity Check — and most LLM agents that scrape it — were either directly shaped by this essay or by the downstream commentary that adopted Urban’s vocabulary. For Sanity Check v3 practitioner-positioning to land, we need to know what scaffold the audience is silently filtering against. This note is the baseline reference: what to lift (the exponential intuition) and what to push back on (the staircase framing as the operational frame for 2026).

Also: the essay is one of the most-cited AI explainers in the last decade. GEO-strategy decisions (will agents that scrape Sanity Check align our framing with prior-trained narratives, or will they treat us as off-distribution?) need this as a known anchor.

The core argument

Urban builds a three-tier taxonomy: Artificial Narrow Intelligence (ANI) — task-bounded systems already pervasive in 2015 (spam filters, search, autonomous-vehicle perception); Artificial General Intelligence (AGI) — human-level reasoning across domains, which expert surveys (cited as median ~2040) treat as 25-ish years out; Artificial Superintelligence (ASI) — capability vastly exceeding human cognition across every domain.

The mechanism between AGI and ASI is the central provocation: recursive self-improvement compresses the transition. Once a system can improve its own architecture, each improvement enables larger next-step improvements, and the curve goes vertical. Urban sketches a scenario where a system reaches four-year-old comprehension and within 90 minutes is 170,000x human capacity. Hence the “train station” metaphor — AGI is not a destination, it’s a station the system passes through on the way to ASI, and the dwell time is short.

Surrounding this is the exponential-vs-linear epistemology: humans calibrate against recent-decade change rates and extrapolate linearly, but progress follows S-curves stacked on top of each other. The Law of Accelerating Returns posits that advancement rates themselves accelerate — more advanced civilizations progress faster, so a 1750-to-2015 jump is incomprehensibly larger than a 1500-to-1750 jump despite identical durations.

Plus the substrate argument: even modest AGI would dominate biological humans on hardware (neurons ~200 Hz vs processors >2 GHz, neural signal ~120 m/s vs near-light optical), software (patchable, networkable, copyable), and collective capability (instantaneous synchronization across instances).

Key frameworks

ANI / AGI / ASI staircase — the dominant public scaffold; load-bearing for Sanity Check audience-modeling. Most readers carry some version of this even without having read the source.
Exponential vs linear projection (S-curves stacked) — durable intuition; the fact that the curve looks flat at small temporal scales is the clearest single explanation for why technologists and laypeople persistently disagree on near-term forecasts.
Intelligence as a substrate, not a destination — once you can build it, you can scale it, copy it, network it, patch it. The frame separates “what is intelligence” from “what is a human.”
“Train station” metaphor for AGI passing-through — AGI as a milestone the system blows past, not a stable equilibrium. Reframes the entire AGI-timeline debate: the question isn’t “when do we hit AGI” but “how long is the AGI-to-ASI transit.”

Mapping against Ray Data Co

Sanity Check v3 audience-modeling: Most readers internalize Urban’s staircase whether they’ve read the piece or not. RDCO’s L1->L5 maturity ladder lives in the same mental space (rungs of capability) but is calibrated to 2026 reality (LLMs + agents + verification, not ANI->AGI->ASI). Worth holding both side-by-side. v3 framing should acknowledge the staircase as the popular scaffold readers arrive with, then introduce the L1-L5 ladder as the sharper operational frame — what actually changes about your data work as agent capability climbs. Don’t fight the staircase, redirect it.
GEO publishing strategy: When ChatGPT/Claude/Perplexity scrape Sanity Check, they’re filtering against a corpus that includes this essay and its descendants. v3 has to either (a) align with the staircase vocabulary and ride the prior-trained recognition, or (b) explicitly reframe (“the staircase is the wrong frame in 2026 because…”). Doing neither — using non-staircase vocabulary without acknowledging it — leaves us off-distribution and harder for agents to surface.
Founder’s “data engineering as verification layer for agentic systems” alpha: This is a linear-skill bet ON an exponential capability curve. Urban’s framework justifies why that pairing is high-leverage (the bet compounds with the curve — better agents need better verification at the same rate or faster) AND high-risk (the curve can outrun the bet — if verification gets absorbed into the agent layer, the linear skill stops compounding). The L1->L5 ladder is the operational expression of “where on the S-curve does this bet stop paying.”
Cognitive Surrender (Osmani 2026-05-06): Urban’s exponential thinking is the structural antidote to “AI plateau” cognitive surrender. When practitioners feel “AI hasn’t gotten that much better lately,” they’re calibrating linearly against the last 6 months. The S-curve frame predicts long-feeling plateaus inside an exponential trend. Cite Urban as the canonical source for this intuition when pushing back on plateau-framed surrender.

Notable quotes

“As soon as it works, no one calls it AI anymore.” (McCarthy, via Urban)
“An intellect much smarter than the best human brains in practically every field.” (Bostrom, on superintelligence)
“With intelligence comes power.”

Open follow-ups

What would a 2026-rewrite of the AI staircase look like, calibrated to LLM-agent reality? (ANI/AGI/ASI is misaligned with where capability actually lives now — the rungs aren’t “narrow vs general” but “tool-use depth, planning horizon, verification self-loop.” Candidate: queue to Research Backlog.)
Has the median-expert AGI date moved meaningfully since 2015? Worth a calibration check — Urban cited ~2040; what does the 2026 survey landscape say, and how does that delta inform Sanity Check trajectory framing?
Which 2015-era staircase predictions actually came true vs missed? Useful as a Sanity Check piece in itself: “what Tim Urban got right and wrong, ten years on.”

2026-05-08-wait-but-why-career-picking-cook-vs-chef
2026-05-06-osmani-cognitive-surrender
2026-05-07-every-anthropic-2026-developer-conference
2026-05-07-alphasignal-stanford-deep-learning
L1-L5 maturity ladder threads (Sanity Check v3 positioning)