The four clocks of every new technology: why AI predictions keep breaking them

A sixty-year overnight success is the cleanest illustration of Rodney Brooks' framework for reading any technology timeline. In 1943, Warren McCulloch and Walter Pitts published the first computational model of a neuron, the math abstraction that imagines brain cells as threshold logic gates. By 1960, Bernard Widrow had refined the idea into the linear-threshold neuron, the basic unit still recognizable in today's neural networks. It then took another half-century of dead ends, a 2012 image-labeling breakthrough by Geoffrey Hinton, and roughly a decade more for that work to surface as the large language models now reshaping every business press headline.

Brooks argues that every technology runs on four clocks, and that researchers, journalists, CEOs, and readers routinely jump between them. The result is predictions that are confidently wrong and sometimes damaging. The clocks are not the same speed. Treating them as one is what makes AI forecasting unreliable.

Clock 1: research ideas. New ideas take 10 to 20 years to form before they reach a solid lab demonstration, and some take much longer when there are many false starts or a single hard step that resists decades of work. After the demonstration comes a gold rush: major tweaks on the same underlying concept arriving every six months, the ground apparently shaking when in fact the underlying concept has already been settled. The McCulloch-Pitts to LLMs arc is one long, repeatedly declared-dead example. Deep learning was pronounced finished more than once before it became the substrate of the present moment.

Clock 2: hype generation. An idea goes from near-invisibility to daily business press while every adjacent researcher re-markets themselves as having always worked on it. The advertising surfaces are an honest thermometer. Almost no San Francisco buses carried AI agent ads in mid-2025; by the time of Brooks' writing, the city's buses were plastered with them. A 30-year-old reader will recognize the shape around blockchain and the metaverse. Anyone older will remember IBM Watson, the nanotube-infused chinos of the early 2000s, and the expert systems of the 1980s. Each wave was going to change everything. None did on the schedule promised.

Clock 3: at-scale deployment. Even when a product is solid, getting it to mass adoption takes time. Software, which has zero marginal cost to copy, still typically needs 20 years or more. Unix shipped at Bell Labs in 1969; commercial versions arrived about 15 years later, but Windows dominated the consumer desktop for decades before Linux reshaped the server landscape. Hardware goes slower. Ernst Dickmanns demonstrated a self-driving car on a Munich freeway in 1987; the DARPA Urban Challenge put self-driving in public consciousness in 2007; Waymo now holds licenses for about 4,000 vehicles in San Francisco, a fraction of the city's cars and a sliver of the US fleet. The number is the point: even after four decades, "scaled" in autonomous vehicles is still a fraction of what the rhetoric implies.

Clock 4: reshaping the economy. This is the slowest clock by far. Domesticated animals, sailing ships, electrification, commercial aviation, and shipping containerization each took more than 50 years of continuous at-scale deployment to reshape the world economy. The current twin hypes, LLMs replacing white-collar labor and humanoid robots replacing blue-collar labor, are pitched as if a one- or two-year horizon is realistic. Hyperloop was once that kind of pitch; most Hyperloop companies have ceased operations or pivoted to slower-speed rail projects. A 50-year technology does not become a 5-year economy.

One serious objection sits on the Hacker News thread about the essay. A commenter concedes the framework but argues AI deployment and adoption are moving much faster than Brooks' historical examples. The reasons offered: massive capital expenditure, distribution through established software platforms, and unusually high user willingness to experiment, even among people only a few years from retirement who are using Copilot at work and ChatGPT for fun. It is the canonical "this time is different" argument, and it has a real point. Distribution through established software platforms is genuinely new. So is the depth of consumer experimentation outside the developer audience.

The framework's reply is structural, not dismissive. The HN comment is still about clocks 1 and 2, research and hype, where speed is genuinely possible. Clocks 3 and 4, deployment at scale and economic reshaping, are governed by the time it takes to rewire business processes, train workforces, retrain regulators, and build the physical infrastructure the AI actually runs on. None of those move at the rate of a Windows update, and that is what most predictions forget.

Next time a CEO promises agents next quarter, or a futurist promises AGI by 2030, run the four clocks. Which one is the claim actually about? If it is research, the timeline can be fast. If it is hype, the timeline is whatever the marketing department wants. If it is deployment, give it 20 years. If it is economic reshaping, give it a lifetime.