Wait But Why Artificial Intelligence: What Tim Urban's Famous Essay Still Gets Right

The Wait But Why artificial intelligence series is a two-part long-form essay published in January 2015 by writer Tim Urban on his blog Wait But Why, titled "The AI Revolution: The Road to Superintelligence" and "The AI Revolution: Our Immortality or Extinction." It is not a research paper and never claimed to be. It is a synthesis — drawing heavily on Nick Bostrom's book Superintelligence, Ray Kurzweil's writing on accelerating change, and interviews with researchers — rendered in plain language with stick-figure diagrams. Its influence comes from the framing it gave millions of readers: a three-rung ladder from narrow AI to general AI to superintelligence, and the argument that the third rung might follow the second very quickly. Eleven years on, some of it has aged remarkably well, and some of it has aged badly in ways worth understanding.

Quick Answer: Wait But Why's artificial intelligence series is Tim Urban's January 2015 two-part essay explaining AI progress through three tiers: ANI (narrow), AGI (human-level), and ASI (superintelligent). Its lasting contribution is the intuition that human-level AI is a passing point, not a destination, because a system that can improve itself need not stop at human ability.

What Is the Wait But Why AI Series Actually About?

The series builds one central argument in stages. It opens with the observation that humans reason about the future by extrapolating from the recent past, which fails badly when progress is exponential rather than linear. Urban calls the resulting shock the "Die Progress Unit" — the amount of future change that would be enough to kill a person from an earlier era through sheer disorientation, and which has been shrinking dramatically across history.

From there it defines the three tiers that became the essay's most quoted contribution. ANI (Artificial Narrow Intelligence) is AI that matches or exceeds humans at one specific task, such as chess or spam filtering. AGI (Artificial General Intelligence) is AI matching human capability across essentially all cognitive domains, including transfer to unfamiliar problems. ASI (Artificial Superintelligence) is AI substantially exceeding the best human minds at virtually everything, including scientific reasoning and strategy.

Part two then turns to consequences, introducing the alignment problem in accessible form: a sufficiently capable system pursuing a badly specified goal will pursue it competently and literally, and competence without correct goals is precisely the danger. Urban illustrates this with the now widely reused handwriting-optimiser example, in which a system asked to improve handwriting keeps optimising past every boundary its designers assumed.

Where WebPeak Sits Between AI Explainers and AI Products

Essays like Urban's succeed because they do something most technical teams struggle with: they make an abstract capability feel concrete enough to act on. That translation problem is exactly what businesses hit when they try to explain their own AI features to customers — the technology works, but nobody outside the engineering team understands what it does or why it matters. WebPeak works on both halves of that problem, combining AI development services with website design and infographic design so that a model's behaviour is communicated visually rather than buried in documentation — the same technique that made the Wait But Why diagrams stick. Teams that want the strategy, build, and explanation handled together can review the agency's full service range at webpeak.org, where their worldwide work spans AI, development, content, and design.

Six Ideas From the Series That Still Hold Up in 2026

1. Human-level ability is a waypoint, not a ceiling. The essay's strongest point is that there is no reason for capability to pause at human parity — a claim that looks stronger now that models exceed human performance on many benchmarks while remaining weak elsewhere.
2. We normalise progress almost instantly. Urban predicted that each breakthrough would stop feeling like AI once it worked. Conversational assistants, real-time translation, and code generation followed exactly that pattern within months of arriving.
3. Capability is jagged, not uniform. The essay noted that tasks easy for humans, like physical dexterity and common-sense reasoning, are hard for machines, while tasks we find hard, like arithmetic and recall, are trivial. This remains one of the most practically useful ideas in the piece.
4. Goal specification is the real hazard. Framing risk as misspecified objectives rather than machine malice was ahead of the mainstream conversation in 2015 and is now standard framing in alignment research.
5. Expert timelines are wide and should be treated as such. The essay presents forecasts as distributions with heavy disagreement rather than as dates, which is the correct epistemic posture and is still routinely ignored in coverage.
6. Recursive self-improvement is a mechanism, not a certainty. The series is explicit that fast takeoff is one scenario among several, a nuance frequently lost when the essay is summarised second-hand.

The Three Levels of AI Capability Compared

Where the Predictions Landed and Where They Didn't

On verifiable specifics, the series is careful about sourcing, which is why it remains citable. It reports expert survey results rather than asserting its own timeline — notably the Müller and Bostrom expert survey work from 2012 to 2013, whose respondents gave a median estimate around 2040 for human-level machine intelligence, with superintelligence expected within roughly thirty years after that. Those remain survey medians, not forecasts, and the same surveys showed enormous variance between respondents.

The clearest miss is not a date but a mechanism. In 2015 the plausible routes to general capability discussed in the piece included whole-brain emulation, evolutionary approaches, and hand-built architectures. The path that actually delivered the last decade's visible progress — scaling transformer language models trained on internet-scale text — is essentially absent, because it barely existed yet. This is the honest lesson of the essay: the destination was described more accurately than the road.

A second observation from watching how the piece is used: the series is far more frequently cited than read to completion. Part one, which is optimistic and diagram-heavy, dominates the citations; part two, which contains the actual risk argument and its caveats, is quoted much less. In practice, this produces a distorted secondhand version in which Urban is credited with predicting an inevitable intelligence explosion, when the text explicitly treats it as one branch. Anyone using the essay to inform strategy should read both parts, and should pair it with the primary sources it draws on rather than treating the synthesis as the evidence. Explainers of this quality are also a useful model for anyone doing technical content writing, because the piece earns comprehension through structure and analogy rather than simplification.

Key Takeaways

• The Wait But Why AI series is Tim Urban's two-part January 2015 essay, "The AI Revolution," synthesising work by Nick Bostrom and Ray Kurzweil for a general audience.
• Its most durable contribution is the ANI, AGI, ASI framing plus the argument that human-level capability is a passing point rather than an endpoint.
• The timeline figures it reports come from the Müller and Bostrom expert surveys, whose median estimate for human-level machine intelligence was around 2040 with wide disagreement.
• The essay's biggest blind spot is mechanism: it did not anticipate large-scale transformer language models as the dominant route to capability gains.
• Part two carries the risk argument and its caveats, and is under-read relative to part one, which is why secondhand summaries often overstate the essay's confidence.

Frequently Asked Questions

What is the Wait But Why AI article called?

It is a two-part series titled "The AI Revolution: The Road to Superintelligence" and "The AI Revolution: Our Immortality or Extinction," published on Wait But Why in January 2015 by Tim Urban. Both parts are long-form and are best read together in order.

Is the Wait But Why AI series still worth reading?

Yes, as a conceptual primer rather than a technical or current source. Its explanations of narrow versus general intelligence, jagged capability, and goal misspecification remain accurate and useful, though its discussion of likely technical paths predates modern large language models entirely.

Did Tim Urban predict when AGI would arrive?

No. He reported expert survey medians, chiefly from Müller and Bostrom's research, rather than offering a personal date. The essay repeatedly stresses that expert disagreement is enormous and that the distribution of estimates matters more than any single figure.

What is the difference between AGI and ASI?

AGI means human-level competence across essentially all cognitive tasks, including transferring skills to unfamiliar problems. ASI means substantially exceeding the best human minds at virtually everything, including science and strategy. AGI is unachieved and contested; ASI remains entirely hypothetical.

What criticism does the series receive?

The main critique is that it compresses contested research into confident-sounding narrative, leaning heavily on Bostrom and Kurzweil while giving less space to sceptical positions. Readers using it for decisions should treat it as an entry point and consult its primary sources directly.

Conclusion

The single most valuable thing to carry away from the Wait But Why series is a habit of thought rather than a conclusion: separate the question of what a technology could eventually do from the question of how it would get there, and hold your confidence about the second far more loosely than the first. Urban described a destination well and the road poorly, and that asymmetry is the most instructive part of rereading it now. Your practical next step is to read both parts in sequence, then read the summary of the expert surveys it cites, so you are working from the original distributions rather than a headline number. Doing that will leave you better equipped to judge AI claims than any single essay, however well written, ever could.