In 1915, Albert Einstein published general relativity and predicted that starlight would visibly bend as it passed the sun — a claim no instrument on Earth could test that year. It took four years and a solar eclipse before Arthur Eddington's expedition to Príncipe and Sobral could point a camera at the right moment and confirm Einstein had been right the whole time. The theory didn't change between 1915 and 1919. What changed was that an eclipse, and the equipment to photograph it, finally lined up.^[1] Artificial intelligence ran the same experiment twice, on a much longer clock, with almost nobody framing it that way at the time.

In 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams published backpropagation — the algorithm that still trains nearly every neural network running today, the method by which a network corrects itself by tracing its own errors backward through its own layers. It took twenty-six more years, a repurposed graphics chip, and a public internet stacked high enough to scrape by the billions of pages before anyone could prove the idea actually worked at the scale it needed.^[2] Same shape as Einstein and the eclipse: the theory wasn't wrong for decades. Nobody had built the instrument yet.

AI's own instrument had already failed to arrive once before backpropagation, and failed again after it, for the same underlying reason each time, described in the language of its own era. In 1973, mathematician James Lighthill told the British Science Research Council that artificial intelligence had failed to deliver on its promises, singling out the "combinatorial explosion" — the way the computation a problem required grew exponentially as its scope grew even slightly. The UK gutted its AI funding within the year; DARPA cut off Carnegie Mellon's speech-recognition program the year after.^[3] A second collapse followed in the late 1980s, when hand-coded expert systems — human knowledge translated one rule at a time into thousands of if-then statements — hit the same wall from the opposite direction: they worked in narrow domains and broke the moment reality got wider than the rules anyone had thought to write down.^[8] Neither generation had a way to let a system learn its own rules from more data than a person could hand-encode, because neither generation had the data or the raw calculation available to do it.

The unlock, when it came, touched none of the underlying math. Ian Buck, later the creator of Nvidia's CUDA platform, was already experimenting with small neural networks in 2005 — a year before Nvidia opened its GPUs, chips built to render video-game triangles, to general-purpose calculation. Between 2006 and 2012, a small community of researchers found that a GPU's core design — thousands of simple processors running the same matrix multiplication in parallel — happened to be exactly the operation neural-network training needed most.^[4] The proof arrived in 2012: a University of Toronto team trained a deep network, AlexNet, on two consumer Nvidia GTX 580 cards against ImageNet's fourteen million labeled photographs, and beat the next-best entry in the field's top competition by more than ten percentage points — a gap too large to be incremental.^[5] Same 1986 algorithm underneath. What was different in 2012 was that the computation and a dataset large enough to be worth computing over finally existed at the same time.

The data half of that equation had one specific source: the public internet itself, crawled at scale. When OpenAI trained GPT-3 in 2020, 60 percent of its training data came from a single filtered slice of Common Crawl, a nonprofit archive that has scraped the open web since 2008^[6] — ahead of every book and encyclopedia in the mix combined.^[7] A chatbot answering a question today exists because backpropagation existed in 1986, because a graphics chip turned out to be good at the same math neural networks needed, and because a public, openly crawlable internet had, by then, spent over a decade accumulating more written language than any research lab could have paid a person to type.