The Early History of the AI Revolution: Google, Musk, and the Race That Changed Everything

When ChatGPT burst onto the scene in late 2022, it felt as if the whole world woke up to AI overnight. But before this "big bang," a small group of people had been working quietly for over a decade. Kevin Roose's new book The AGI Chronicles and this episode of The Vergecast take us back to that overlooked era to examine what the earliest "believers" got right—and where they went wrong.

Why Did Google Become the "Public Enemy"?

The story begins with Google's acquisition of DeepMind. At the time, this small London company was virtually unknown, yet it publicly declared "artificial general intelligence (AGI)" as its goal—which most people considered sheer madness. Google's acquisition would later prove to be one of the most successful in history, but it also thoroughly ignited Elon Musk's anxiety.

Musk had been an early investor in DeepMind, knew its CEO Demis Hassabis, and considered him extremely competitive. According to the book, Musk worried that Hassabis would use Google's resources to fulfill his "evil genius" ambitions. This concern prompted him to write to Sam Altman and others, ultimately leading to the creation of OpenAI—an organization conceived as a "counterbalance to DeepMind."

During the interview, host David Pierce offered a sharp observation: every company seems to need an "enemy" to drive itself forward. He half-jokingly asked whether Hassabis was truly an "evil empire" or simply a "useful symbol" placed on the opposite side to motivate employees. Roose's answer was nuanced: Hassabis is indeed fiercely competitive (a chess prodigy and five-time "Pentamind" champion at the Mind Sports Olympiad), and his "Singleton hypothesis"—that the first AGI will control the universe and form a natural monopoly—was interpreted by peers as "I want to become the god of humanity." This interpretation itself may explain better than the facts why this race has become so frenzied.

DeepMind was founded in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman in London, with the goal of "solving intelligence" and ultimately building artificial general intelligence (AGI). In 2014, Google acquired DeepMind for approximately £400 million, when AGI was widely regarded as science fiction. DeepMind stunned the world in 2016 by defeating Lee Sedol with AlphaGo, proving the potential of deep learning in complex games. It went on to produce breakthroughs such as AlphaFold (protein structure prediction) and merged with Google Brain in 2023 to form Google DeepMind. DeepMind's existence kept Google at the forefront of foundational AI research for years, but commercializing its research was repeatedly hampered by internal constraints—a key backdrop for the later discussion in this article.

Did They Really Believe Humans Don't Matter?

An uncomfortable fact is that the most radical proponents of AGI often lack, at a fundamental level, any reverence for human specialness. In his book, Roose calls these people "seekers"—they believe intelligence lies on a continuous spectrum, evolving from ancient prokaryotes all the way to humans and then inevitably beyond. They want to see what the next stop on the spectrum is.

This worldview is complemented by "computational functionalism": consciousness, emotion, and memory are merely computational processes in the brain—input-output transformations among neurons and synapses. Since the human brain is essentially a computer, it can be replicated digitally. The host's exclamation during the interview captures the reaction of most people: "I've never spent time seriously thinking about this worldview."

This philosophical foundation explains why these companies can keep moving at full speed while believing their own products might destroy humanity. Roose emphasizes that he didn't write this book to convey a particular message, but to record a period of history that might otherwise be forgotten. Yet his account reveals a key fact: these world-changing decisions came from a very small group—about 150 people were interviewed for the book, but fewer than 50 may have been involved at every critical juncture.

Computational functionalism is a position in the philosophy of mind holding that mental states are essentially defined by their functional roles, regardless of the physical substrate that implements them. In other words, as long as a system (whether a carbon-based brain or a silicon-based chip) implements the same input-output causal structure, it possesses the same mental states. The idea traces back to Hobbes and Leibniz and was systematized by Putnam and others in the 20th century. It directly supports the view that "the mind can be computationally simulated" and is one of the philosophical cornerstones of strong AI. Opponents argue that consciousness involves irreducible subjective experience (qualia) that cannot be replicated by functional equivalence alone. The beliefs of the AI company leaders mentioned in the article are directly in line with this philosophical tradition.

Google Could Have Won It All

The most "parallel universe" detail in the book is that Google had already built a prototype of ChatGPT back in 2020—a conversational Transformer model called Meena. In internal tests, it comprehensively outperformed Google Assistant, and someone even wrote an internal memo titled "Meena Eats Google," foreseeing how a chatbot could disrupt search.

But Google did nothing. The search team worried about cannibalizing ad revenue, the translation team feared their product would be replaced, and the legal and ethics teams worried about bias and hallucinations. Every department had a reason "not to ship." As a result, a product that could have reshaped the search landscape was shelved.

Roose's assessment of Google: it is still the company we thought it was—full of smart people, inventing countless technologies, but also mired in bureaucracy. "Google is still seen as a frontier AI company, which is almost a miracle," he said. "It could have become Yahoo."

The Original Sin of Data: They Knew It Was Wrong, But Did It Anyway

The host's notion of the AI industry's "original sin"—the way training data was collected—is confronted head-on in the book. Roose reveals that OpenAI's early attitude was "it's just research, no need to ask anyone"; Google's legal department, by contrast, objected explicitly, which was one reason its products were slow to launch.

Anthropic's Dario Amodei once proposed an economic model to compensate data producers: track each training input's contribution to model output quality and distribute revenue proportionally. But it was ultimately abandoned because it was "too hard and too complex," and they were in a race. So they chose to simply dump the data in—which ultimately led to one of the largest copyright settlements in U.S. history.

Roose himself is a member of the class action in that case, because his book was used for training, and he expects to receive a check for about $1.75. He admits that if AI companies had managed data ethics better from the start, "today's debate about AI might be completely different." But he also acknowledges that, at a scale of billions of tokens, precisely tracking contributions is nearly impossible—it might just amount to "a few cents' worth of atoms."

One of the largest copyright settlements in U.S. history refers to multiple class-action lawsuits filed against OpenAI, Anthropic, and others between 2023 and 2024. The Authors Guild, representing authors including John Grisham and George R.R. Martin, sued OpenAI for using their works without authorization to train ChatGPT; Getty Images sued Stability AI for using its images to train image-generation models; music publishers also sued Anthropic and OpenAI. In 2024, Anthropic agreed to pay approximately $1.5 billion to settle an authors' class action, making it one of the highest-value AI copyright settlements at the time. The core dispute in these cases is whether using copyrighted works to train AI constitutes "fair use." Courts have not yet established a uniform precedent, but the trend toward settlements suggests that AI companies are being forced to pay for their data sources.

Chatbots: First, They Have to Be Fun to Talk To

On the eve of ChatGPT's launch, OpenAI employees bet on how many people would sign up on the first day, with most guesses between 10,000 and 100,000. No one anticipated it would become a billion-user product. After all, they had been using GPT-4 internally and had become desensitized to "chatting with a computer."

One thing the industry collectively overlooked: these models took off first and foremost because they were fun. As David Pierce quoted Amazon's Panos Panay: "You can't imagine how much tolerance people will give it as long as it can string together coherent sentences." Sundar Pichai later admitted to Roose that Google underestimated how willing people would be to talk to an imperfect chatbot.

Panos Panay is Microsoft's former Chief Product Officer, responsible for the Surface and Windows experiences, known for his obsessive pursuit of product detail and user emotion; he joined Amazon in 2023 to lead its Devices and Services division. His remark that "as long as it can string together coherent sentences, people will give it immense tolerance" reveals an underappreciated dimension of success for conversational AI products: fluency of interaction and personification. ChatGPT's interface is minimal, its tone natural, and it even allows users to correct and follow up—this "chatting like a human" experience dramatically lowers the barrier to use. Technical "imperfections" (such as factual errors) are masked by conversational coherence and usefulness; users care more about "it understands what I'm saying" than "it's always right." This insight explains why ChatGPT reached 100 million users in just two months, while more precise models before it attracted little attention.

Have We Reached AGI?

At the end of the interview, the host posed the "clichéd" question. Roose gave his famous "time machine definition": if you took today's models back to 2017 and let researchers then try them, they would say without hesitation, "This is AGI."

His personal definition is more down-to-earth: we already have systems that can reliably answer factual questions and complete work tasks like a random stranger pulled from a bar. As for the AGI that "ends humanity"? He doesn't believe in it. He only believes we are on a path that will profoundly reshape life.

David Pierce's remark perhaps best captures the core of this interview: AGI is a loathsome term because you can declare you've reached it at any time. What matters is not the label, but the change we are all "feeling."

One-line Summary

Google's acquisition of DeepMind fueled Musk's anxiety and led to OpenAI; Google missed the Meena opportunity; the AI race is reshaping the landscape amid data original sin and product intuition.

Paragraph Summary

Based on Kevin Roose's new book The AGI Chronicles and an interview on The Vergecast, this article reviews the decade-plus history of the AI race before ChatGPT's breakout. Key threads include: how Google's acquisition of DeepMind triggered Musk's anxiety and directly led to OpenAI; the computational functionalist worldview held by the "seekers" pushing AGI; how Google developed a ChatGPT prototype, Meena, as early as 2020 but failed to release it due to internal departmental conflicts; how OpenAI and others, under competitive pressure, abandoned data compensation schemes and used scraped data directly, ultimately leading to major copyright settlements; and how ChatGPT's breakout stemmed from a simple, overlooked fact—people are willing to talk to an imperfect but "fun" chatbot. The article concludes by discussing the definitional divide over AGI, noting that what matters is not the label but the profound changes already underway.