The Shock of Generative AI ― ChatGPT and Foundation Models

In the previous episode, we saw the invisible struggle over the heart of AI—the computing power that is a crystal of sand. But the heartbeat that heart kept tapping out would, one day, break through the walls of the laboratory.

November 30, 2022. A single website was quietly made public. Its name was ChatGPT. No difficult operations were required. You simply typed in text, as if speaking to a person. And from beyond the screen, as though it were thinking, fluent words came back.

From that day on, the air of the world changed. Artificial intelligence, long the province of specialists, had at last leapt into the palms of ordinary people.

A Hundred Million in Just Two Months ― Intelligence Comes to the Palm

The speed of its spread outran everyone’s predictions. According to reporting, ChatGPT is said to have reached a million users just five days after its release and to have passed a hundred million within two months. For its time, it was one of the fastest-spreading consumer applications in history.

What was it that drew people in so powerfully? Until then, AI had mostly worked behind the scenes. It reordered search results, tagged photos, turned speech into text—useful, but an unseen helper in the wings. ChatGPT was different. It wrote prose, summarized, translated, and even composed program code. Ask it a question and it answered; add a request and it rewrote. For the first time, many people felt the tangible sensation of “conversing” with an AI.

Students asked it to draft their reports, office workers left the wording of their emails to it, and engineers made it a partner for puzzling out troublesome code. In schools, unease spread over the fact that assignments could simply be solved outright, and the media wrote daily of the promise and the peril of this new tool. A single chat window threw a question, all at once, at how people worked and how they learned.

Of course, its answers often contained errors, and at times it told plausible-sounding lies—what came to be called hallucinations. Even so, the shock of the fact that intelligence had become everyone’s possession was great enough to overshadow those flaws. Many companies vied to release their own generative AI, and within just a few months the phrase “generative AI” shifted from technical jargon into everyday speech.

The Machine That Learned Language ― LLMs and Foundation Models

Behind this shock lay a quiet but decisive accumulation of technology.

The key was a mechanism called the large language model (LLM). You feed it vast quantities of text and have it relentlessly predict “the word that comes next.” That alone, repeated at an almost unbelievable scale. The machine came to grasp, statistically, not only grammar and facts but also the very rhythm of context and turn of phrase.

The foundation for this was the structure called the Transformer, set out in the 2017 paper “Attention Is All You Need,” published by researchers at Google. This mechanism, which efficiently learns which words in a sentence should pay attention to which, became the shared basis for the language AI that followed.

Then, in 2021, researchers at Stanford University in the United States gave these enormous models a name: foundation models. You first train one giant model on broad data, then apply it to diverse uses such as translation, summarization, and dialogue—much as one might raise several buildings on a single base. The very way AI was made shifted from an era of building small, task by task, to an era of erecting one enormous foundation and reusing it.

Building this foundation requires the enormous computing power we saw in Episode 5. To train a foundation model, you must keep a great many high-performance chips running for long stretches. The dazzling arrival of generative AI was, behind the scenes, bound tightly to the competition over how much of the heart—computing power—one could hold.

The Horizon Called AGI ― America Out in Front, the Rest in Pursuit

The arrival of ChatGPT stirred yet another word in people’s minds. AGI—artificial general intelligence.

Not today’s AI, which handles only specific tasks, but an intelligence that can learn and carry out a wide range of challenges on its own, as a human does. Whether it can truly be realized, when it might come, whether it is even possible at all—experts’ views differ greatly. But having witnessed the dramatic progress of generative AI, people suddenly began to speak of AGI not as a distant fantasy but as a goal that might come within range.

In the development of this generative AI and these foundation models, it was a cluster of American companies that is said to have led the world in the early going. There, an environment in which funding, cutting-edge computing power, and talent could readily gather was at hand. As we saw in Episode 3, China had already placed AI at the core of its national strategy. China and the companies of other nations alike are reported to have hurried to develop their own large models, attempting to give chase. The competition over the cutting edge of intelligence thus entered a new phase, drawing in nations and companies.

What is intriguing is that this competition does not stop at a mere contest of technical strength. Which language’s data to train the model on, what values to reflect in its answers, what to have it say and what to keep it from saying—generative AI mirrors, in each of its answers, the culture and norms of the society that makes it. That is precisely why nations began not merely to compete on performance but to seek an intelligence that runs on their own language and values. The supremacy of intelligence came, before long, to take on the coloring of a contest of culture and thought as well.

That said, which companies or nations will ultimately stand at the front, and when and in what form AGI may arrive, are matters on which assessments differ greatly, and no firm conclusion can be drawn at present. What is certain is the fact that, from the close of 2022, AI worked its way into every corner of society and crossed a point of no return.

Just as oil moved the world and semiconductors sustained the age of electronics, now an “intelligence that wields language” has taken its seat at the center of people’s lives and of national strategy as a new wellspring of power. The “dream of making intelligence” we asked about in Episode 1 had, here at last, become a reality anyone could touch.

But the greater the power grew, the more humanity began to ask—who, in the end, makes the rules for this intelligence?