Four time scales explain why AI hype outpaces real deployment

An essay published on rodneybrooks.com lays out four distinct time scales for technology and argues that jumping between them is what produces outrageously wrong, sometimes damaging predictions about when a technology will actually do something. The author does not name themselves in the text, so no identity is asserted here beyond the publication.

The first scale is new research ideas, which the essay says take ten to twenty years to form before reaching a solid lab demonstration, with some ideas needing far longer because of false starts or a single hard step that takes decades to crack. Once a technology is established in the lab, a gold rush phase can follow where variants appear every six months. The essay traces this for neural networks: the first computational neuron models were published in 1943 by McCulloch and Pitts, the dominant linear threshold neuron model was established in 1960 by Widrow, convolutional networks with backpropagation followed, and in 2012 Hinton's deep learning let trained neural networks overtake conventional computer vision. Today's large language models arrived roughly a decade after that, making the total research arc about sixty years, with the idea declared dead multiple times along the way.

The second scale is hype generation, which moves far faster than research. The essay points to how quickly the term AI agents went from obscure to plastered on San Francisco buses, absent in mid 2025 and now hard to avoid. It lists earlier hype cycles that have since faded: blockchain and the metaverse, IBM Watson, nanotechnology (including a chinos maker's ads touting nanotechnology in its pants), and expert systems that were supposed to capture expert knowledge and let companies cut staff. The essay warns that people unfamiliar with a technology often cannot separate genuine ongoing research from hype about how it will change everything, and that the ratio of hype events to genuinely transformative technologies is far too high.

The third scale is at scale deployment, the time from a solidly engineered product to mass adoption. The essay says software typically takes 20 years or more to scale up even though it has zero marginal cost to copy, because businesses do not automatically re-engineer around new tools. Unix, started at Bell Labs in 1969, had commercial versions shipping 15 years later, though Microsoft Windows became the dominant operating system instead. Linux, begun in 1991, was known to every computer science graduate student within five years, but Microsoft did not adopt it until 2012; today it runs most backend systems and billions of mobile devices. Hardware takes even longer: the author saw a self-driving car demonstrated on a freeway outside Munich in 1987 by Dickmanns, but the idea only entered public consciousness after the 2007 DARPA Urban Challenge. The author first rode a Waymo predecessor, then part of Google X, on Highway 101 in 2012; Waymo is now licensed for about 4,000 vehicles in San Francisco and leads the US market, though that is still tiny next to the city's total car count. The essay adds a personal anecdote that despite the app promising to take him home, a Waymo dropped him off somewhere else, and he spent over 20 minutes on the phone with customer support.

The fourth scale is reshaping the economy. The essay names the current two biggest hype concentrations, large language models replacing white collar labor and humanoid robots replacing blue collar labor, both promising a magically richer world. It counters that many technologies have reshaped the economy over the millennia, from domesticated animals and sailing ships to electrification, commercial air transport, and shipping containerization, but each took over 50 years of continuous at scale deployment. Some heavily hyped technologies fail outright, the essay notes that every company formed to commercialize Hyperloop has since shut down. The essay concludes that reshaping the world economy at scale genuinely takes decades, essentially a human lifetime, and that people mistakenly expect a new research result to change everything within one, two, or ten years.

Key facts

• New research ideas take ten to twenty years to reach a solid lab demonstration; large language models took about sixty years total, from the first computational neuron models in 1943 (McCulloch and Pitts) through 1960 (Widrow) and 2012 (Hinton) to today.
• Hype cycles move far faster than research: the term AI agents went from absent to covering San Francisco buses within about a year, following earlier faded hype cycles like blockchain, the metaverse, IBM Watson, nanotechnology-branded chinos, and expert systems.
• Software typically takes 20 years or more to reach mass adoption at scale; Linux, started in 1991, was not adopted by Microsoft until 2012, and Unix's 1969 origins took 15 years to reach commercial shipping versions.
• Hardware deployment is slower still: a self-driving car was demonstrated near Munich in 1987, the idea reached public consciousness only after the 2007 DARPA Urban Challenge, and Waymo, whose predecessor the author first rode in 2012, now runs about 4,000 licensed vehicles in San Francisco, still a small share of the city's cars.
• Reshaping the whole economy has historically taken over 50 years of continuous at scale deployment, per the essay, which argues this undercuts claims that large language models and humanoid robots will replace human labor within a few years.

Why it matters

The essay's core claim is that people routinely conflate four separate clocks, research maturity, media hype, at scale deployment, and economy wide impact, and that conflating them is precisely what produces outrageously wrong predictions about AI and robotics. It offers a simple diagnostic: before believing a claim about a technology's near term impact, ask which of the four time scales it actually describes.

Who it affects

Anyone forming or repeating predictions about AI and robotics, including executives, investors, journalists, and policymakers exposed to claims that large language models will soon replace white collar work or that humanoid robots will soon replace blue collar work, as well as ordinary readers trying to judge how seriously to take such claims.

How to use it

The essay suggests checking, for any hyped claim, whether the underlying technology has actually cleared each earlier stage: has it left the research lab, has the hype outrun genuine deployment, has it been engineered and adopted at real scale, and has it been deployed continuously for the decades that economy wide change has historically required. If a claim skips straight from a lab result or a hype cycle to economy wide transformation, the essay's own historical examples suggest skepticism.

How solid is it

This is a single author's argument, not a peer reviewed study, but it is backed by specific, checkable historical dates and examples rather than vague assertions: named researchers and years for neural networks (McCulloch and Pitts 1943, Widrow 1960, Hinton 2012), Unix and Linux's development and adoption dates, the 1987 Munich self-driving demonstration and 2007 DARPA Urban Challenge, and Waymo's current licensed vehicle count in San Francisco. The text itself does not name the essay's author or give a publication date.

Risks and caveats

As an opinion essay, it generalizes a pattern from a relatively small, self-selected set of historical examples rather than a systematic survey, and it does not quantify how many technologies fit the pattern versus how many do not. The Waymo vehicle count and other figures are a snapshot in time rather than a fixed rule, and the essay's framework is presented as the author's own observation rather than an established or tested model.

“Getting things to work at scale is orders of magnitude harder than getting them to work at first and having your first few dozen satisfied customers.”

— the essay