The AI moment was seventy years in the making
April 2026 • 481 words • 3 min read
What we call the AI moment isn't a single breakthrough. It's three separate lines of work that finally crossed: a way to describe information mathematically, hardware fast enough to run it, and an architecture that used both. What changed in 2022 was access.
In 1947, Bell Labs invented the transistor. The next year, in the same building, Claude Shannon published a paper on the mathematics of communication. Neither one knew what it was starting.
What we call “the AI moment” isn’t a single breakthrough. It’s three separate lines of work that finally crossed: a way to describe information mathematically, hardware fast enough to run that math, and an architecture that used both well. Each one matured on its own schedule. None was waiting for the others.
A language for information
In 1948, Claude Shannon published A Mathematical Theory of Communication. He was trying to solve a narrow problem: how do you get a signal through a noisy channel without losing it?
What he wrote down was far bigger than the question. Entropy, compression, channel capacity, and the idea that meaning can be separated from whatever medium carries it and measured on its own.
Every token a language model predicts today is downstream of that paper. Shannon thought he was working on a telephone problem.
Hardware that could run the math
The transistor arrived in 1947, and then came seven decades of making it smaller, faster, and more parallel.
Hardware visibly caught up with modern AI in 2012. A group at the University of Toronto trained a neural network called AlexNet on gaming GPUs and won the ImageNet contest by a margin wide enough that the rest of the field stopped to recheck its assumptions. GPUs were built to draw graphics. They turned out to be the right shape for deep learning too: lots of parallel cores, fast at matrix math, cheap enough that a couple of graduate students could afford to experiment.
Google went further and built its own chips. TPUs shipped in 2016, made only for neural network work. The hardware was now being designed on purpose for AI.
An architecture that used both
In June 2017, a team at Google published Attention Is All You Need. It proposed the Transformer, which dropped recurrence and used attention to read a whole sequence at once instead of one step at a time.
Google understood it well enough to ship BERT in 2018 and T5 the year after. What it didn’t do was push the architecture as hard as it could have. The likely reason was search revenue. If your business depends on people clicking into web pages for answers, a model that just gives the answer is a threat you manage carefully.
OpenAI had nothing to protect. GPT-2 in 2019, GPT-3 in 2020. Anyone watching could see the trajectory. The three threads had already met.
What happened in 2022
By the time ChatGPT launched in November 2022, none of the fundamentals had moved. The models were bigger and better tuned, but the architecture, the hardware, and the training recipe had all been in place for years.
What changed was the front door. A chat box, no API key, nothing to learn first. In 2022, seventy years of work finally became visible to everyone with a connection.
The three threads took very different amounts of time to pay off. Shannon’s 1948 paper needed sixty years of hardware before it could run at any real scale. The Transformer took less than five years to go from a research PDF to the engine inside a product with a hundred million users.
That gap, between something being discovered and the same thing being in everyone’s hands, has been shrinking the whole way through. That, more than any single model or paper, is what makes this moment feel different from the ones before it. Whatever the next thread turns out to be, it will reach you faster than this one did.
The follow-on question, where those curves are headed and where they break, is what I wrote about in AI doesn’t have a Moore’s Law.