The 30-Year Lag

The gap between seeing the thing, naming the thing, and deploying the thing is measured in human generations, not news cycles.

In 1831, Michael Faraday discovered that moving a magnet near a wire produced an electric current. He didn’t have the math to explain why. He didn’t have the theoretical framework. He just observed it, wrote it down, and moved on. Most physicists at the time thought it was an interesting curiosity, possibly useful for novelty toys, definitely not the basis of a civilization.

It took thirty years for James Clerk Maxwell to write the equations that explained what Faraday had seen. It took another thirty years after that for engineers to build the first generators that turned Maxwell’s equations into electric power. From Faraday’s discovery to the first electrified city was about seventy years. From the first electrified city to electricity being unremarkable was another fifty.

This is the normal pace at which a fundamental discovery becomes a civilization. The gap between seeing the thing and naming the thing and deploying the thing is measured in human generations, not news cycles.

In 1900, Max Planck noticed that black-body radiation didn’t fit the equations everyone was using, and proposed a strange fix involving quantized energy. He didn’t believe the fix was real. He thought it was a mathematical trick. It took twenty-five years before quantum mechanics, as we know it, was a coherent theory. It took another twenty for the transistor, in 1947. It took another fifteen for the integrated circuit. From Planck’s reluctant fix to your phone is roughly a hundred years.

In 1948, Claude Shannon wrote a paper called “A Mathematical Theory of Communication.” Almost nobody outside Bell Labs read it for a decade. The paper invented information theory. It defined the bit. It is, arguably, the most important scientific paper of the twentieth century. It took thirty years for engineers to start building communications systems on Shannon’s principles. It took another twenty for those principles to become the foundation of the internet.

The lag, in each case, is roughly thirty years from discovery to deployment, plus another thirty for deployment to feel normal.

This brings us to 2017.

In 2017, eight researchers at Google published a paper called “Attention Is All You Need.” It introduced the Transformer architecture. It was nine pages long. It received polite attention in the natural language processing community and was largely ignored by the broader computing industry for about three years.

Five years later, in 2022, ChatGPT was released, and the world acted as if a meteor had hit. Articles began appearing about the speed of AI development, the unprecedented pace of change, the impossibility of regulation because everything was moving so fast.

But the math doesn’t change. Vaswani et al. is to the Transformer what Faraday was to electromagnetism, what Planck was to quantum mechanics, what Shannon was to information. The discovery happened. The deployment is just beginning. The thirty-year clock started in 2017, and we are eight years in.

By the historical pattern, this means we are at the equivalent of 1839 in electricity. 1908 in quantum mechanics. 1956 in information theory. We have a working theory, the first impressive demonstrations, and a public that has noticed something is happening but doesn’t yet have the vocabulary to describe it. We are still mostly naming the thing. We have not yet deployed the thing.

This matters because it changes how to read the present moment.

Almost every confident prediction about AI — both the apocalyptic and the utopian — assumes that what we have now is what AI will be. This is the error of looking at Faraday’s induction coil in 1831 and concluding either (a) electricity is a novelty that will fade or (b) we are about to enter the age of electric mind control. Both predictions were made. Both were absurd. The actual future of electricity was something neither side could have imagined, because the deployment takes thirty years and looks nothing like the early demonstrations.

If the historical pattern holds — and there’s no obvious reason it shouldn’t — then the AI of 2050 will look as different from ChatGPT as the modern electrical grid looks from Faraday’s first coil. The underlying physics will be the same. The applications will be unrecognizable. Most of them will involve domains where we don’t yet have the words to describe what’s happening.

This is not a comfortable timeline for an industry that has trained itself to think in quarters. It is also not a comfortable timeline for the people predicting imminent collapse. The honest answer is: we are early, the pace will continue to feel fast and slow at the same time, and the most important applications haven’t been imagined yet — by anybody, including the people building the models.

Naming a thing is the last step before it becomes invisible. We are still in the loud, public, naming phase. The quiet, invisible, deployed phase is still ahead.

It will not arrive on Twitter. It will arrive in the way electricity arrived — a wire at a time, until you forget to notice.

Part of the Aperture series at Nous. The longer arc of why the present moment looks fast and slow at once pairs with Naming Is the Last Step in this series, and with Why Long Tasks Break AI in the Logocachexia series.