Artículos relacionados a The Bitter Struggle: Superintelligence, Superpowers,...

The Bitter Struggle: Superintelligence, Superpowers, and the Fate of the World - Tapa dura

Buchanan, Ben; Collins, Tantum

 
9798217089482: The Bitter Struggle: Superintelligence, Superpowers, and the Fate of the World

Sinopsis

From two former White House officials who shaped America’s AI strategy and now work with leading AI companies, an urgent case for why democracies must achieve preeminence in AI—and a bracing warning about what victory will require.

“Essential reading for understanding how AI will impact geopolitics and how the great powers are shaping AI.”—Chris Miller, New York Times bestselling author of Chip War

Humanity may soon create AI systems that outperform humans at virtually every cognitive task. Whoever wields that technology will gain a power unlike any in history. In The Bitter Struggle, Ben Buchanan, the first White House Special Advisor for Artificial Intelligence and currently an advisor to Anthropic, and Tantum Collins, who served on the National Security Council after five years at Google DeepMind, argue that the United States and its democratic allies must not lose this contest. But how democracies win matters most of all.

Buchanan and Collins take readers into the briefing rooms where AI CEOs privately warned about their own creations, into the Situation Room where Cabinet Secretaries debated how to choke off China’s access to AI technology, and into the Oval Office, where they helped shape President Biden's decisions. From semiconductor factories in Taiwan to city-sized data centers in America, they show why a million chips may matter more than a million troops, why AI progress has been so stunning and yet so predictable, and why the next few years may decide the next few decades. Yet inventing more powerful AI is only the beginning. Democracies must also deploy AI’s capabilities without surrendering their own values—and so far, they are failing.

The deepest challenge is not the contest with autocracy but the conflict AI will force within democracies themselves. Every choice pits security against liberty, speed against caution, power against principle. This is the bitter struggle. Democracies have every advantage they need to prevail—except, perhaps, the conviction to act.

"Sinopsis" puede pertenecer a otra edición de este libro.

Acerca del autor

Ben Buchanan was the first White House Special Advisor for AI, leading the development and execution of the Biden administration’s AI policy. He is the author of four books and is currently a professor at Johns Hopkins University and an advisor to Anthropic.

Tantum Collins was Director of Technology and National Security during the Biden administration. Before joining government, he spent five years at Google DeepMind and coauthored the New York Times bestseller Team of Teams. He is currently the CEO of Inherent, an AI startup.

Fragmento. © Reproducción autorizada. Todos los derechos reservados.

One

Scale

On December 18, 1903, Lorin Wright approached Frank Tunison, the Associated Press representative in Dayton, Ohio, with a pitch for a news story: Lorin’s brothers Orville and Wilbur had just achieved humanity’s first heavier-­than-­air flight. The brothers had flown four times. The longest voyage, Lorin told Tunison, had lasted fifty-­seven seconds.

Tunison could have sent the story out on the AP wire, where it would have been picked up by newspapers all over the country. But he doubted its value. “Fifty-­seven seconds, hey?” he reportedly said. “If it had been fifty-­seven minutes, then it might have been a news item.”

Had Tunison known the rest of the facts, he would likely have harbored even greater doubts. The first flight had lasted a mere twelve seconds, consisting of a 10-­foot ascent and a 120-­foot voyage. The plane’s ground speed had never exceeded seven miles per hour, making it slower than a bicycle (fittingly, since the brothers owned a bicycle shop). Its flimsy wings were covered in a light muslin best known for its use in underwear. It had no seat or steering wheel; the brothers took turns lying on the bottom portion of the wing and “piloting” the craft by tilting their hips from side to side. After the final flight on that first day, a gust of wind blew the world-changing invention over, damaging it so badly that it would never fly again.

Scientific authorities were also skeptical of aviation. Seven years earlier, the eminent physicist Lord Kelvin had written, “I have not the smallest molecule of faith in aerial navigation other than ballooning.” Simon Newcomb, one of the world’s best-­known astronomers, voiced a similar opinion less than two months before the flight. The New York Times was a touch more optimistic: Around the same time, it estimated that humans might achieve such a feat in “one million to ten million years.”

Like Tunison, Scientific American turned down the story. The most favorable early press the brothers received was from a beekeeper who had witnessed some later flights and wrote about it in his hobbyist magazine. Even several years later, the Paris edition of The New York Herald questioned the brothers’ success, writing, “They are in fact either fliers or liars.” And though the brothers were insistent about the military utility of the plane, the U.S. government rejected their first offer of a contract.

The story of modern AI in many ways looks nothing like the brothers’ saga. It is corporate, not personal; primarily digital, not physical; and proceeding at a breakneck pace. But the pursuit of super­intelligence shares one crucial attribute with the development of the aviation industry: a dependence on scale.

A plane that could stay aloft for less than a minute would never, on its own, have revolutionized shipping, travel, or war, but it demonstrated that the brothers had solved the fundamental problem of heavier-­than-­air flight. Their invention offered a clear path to new heights. Larger engines, stronger materials, broader wings, and more rigorous engineering enabled extraordinary aviation accomplishments in commerce, science, and military conflict. Sixty-­six years after those first twelve seconds in the air, Neil Armstrong walked on the moon, carrying in his spacesuit a strip of cotton—­that underwear muslin—­from the wing of the brothers’ first plane.

Humanity has already witnessed a major AI paradigm shift with the transition to machine learning systems. Continued progress now comes not just from new ideas but from scaling established architectures across ever-­greater numbers of ever-­more-­powerful chips in an expanding industrial process. Fundamental innovation continues, but increasing scale drives progress, providing both predictability and power to those who seek it.

The Bitter Lesson

Richard Sutton is an AI legend. He won the 2024 Turing Award, considered to be the Nobel Prize of computer science, for his foundational contributions to reinforcement learning, a subfield critical to the progress of recent decades. Nearly every serious AI student and scientist has read his textbook and many of his research papers.

For our purposes, though, Sutton’s most important contribution is his short 2019 essay, “The Bitter Lesson,” which reflected on the preceding seven decades of AI research—­a period defined by commercial booms and busts, breathless media hype, and shifting academic fads. Amid all that churn, Sutton identified one constant: “general methods that leverage computation are ultimately the most effective, and by a large margin.” He argued that the greatest AI performance comes not from giving a machine precise instructions but from constructing flexible algorithms that can learn by using copious computational power. As a writer would put it, “Show, don’t tell.”

That lesson was bitter because it diminished the role humans had hoped to play. Researchers wanted their own expertise and skill, rather than the brute scaling of computing power (or “compute,” as it has come to be known), to be the essential ingredients in creating AI. As Sutton wrote, “researchers always tried to make systems that worked the way the researchers thought their own minds worked.” Building in human knowledge was “personally satisfying to the researcher,” but “in the long run it plateaus and even inhibits further progress.” Even as scientists sought to make machines as smart as humans, the larger vision for humanity’s role remained clear: We wanted our creations to depend on us.

Humanity does not get to choose which research paradigms succeed any more than it chooses the laws of physics. The results may bruise the collective human ego, but scaling compute has propelled AI forward. Scaling turns modest advances into breakthroughs and will likely prove decisive in the geopolitical AI competition.

Sutton provided many historical examples of the bitter lesson. Image recognition is a classic case. For decades, researchers had handcrafted formulas to detect specific visual features, such as the curves of a cat’s ear or the shape of a stop sign, stitching systems together bit by bit. Progress was steady but very slow.

In 2012, a team from the University of Toronto took a different approach. They deployed a neural network—­layered mathematical functions with many machine-­adjustable parameters that collectively map an input, such as a set of pixels, to an output, such as a label like “cat.” Instead of giving the machine precise instructions about what to look for, they trained the system, AlexNet, on millions of labeled images, letting it discover for itself what mattered most. Neural networks had been around for decades, but there had not been enough compute to make them work at scale.

AlexNet slashed the error rate for image recognition nearly in half. Google quickly purchased the team’s hastily formed company, and within two years virtually every competitive image recognition system used neural networks. Similar transitions followed in nearly every subfield and application of AI. In each case, the systems that dominated emerged from a neural network training process rather than top-­down human direction. As Chris Olah, a leading expert on the inner workings of AI systems, put it, “We don’t make them. We grow them.”

Then came mastery of Go, one of the world’s most ancient and popular board games. There are more possible Go board configurations than the number of atoms in the universe squared. Chess had succumbed decades earlier to the cold and inelegant Deep Blue, which had used hard-­coded rules assembled by grandmasters and programmers to topple world champion Garry Kasparov in 1997. But Go seemed too beautiful and expansive to submit to a similarly soulless machine. Top Go players make decisions based not on calculation but on intuition. At the time of Kasparov’s defeat, experts estimated it would take a century of further progress before machines could win at Go.

In 2016, a Go-­playing neural network developed by Google DeepMind, where Tantum had just started working, defeated world champion Lee Sedol. DeepMind trained AlphaGo on millions of games played by human professionals and amateurs. The system learned the patterns that defined exceptional play, then squared off against itself millions of times, experimenting with new tactics, many of which humans had never explored. With each game, it built beyond its human foundation, eventually becoming the best Go player the world had ever seen.

DeepMind later excised the step in which AlphaGo learned from human games. That new approach yielded an even more powerful system, one that could beat any human or machine not just at Go but also at chess and Shogi (a popular Japanese board game). As David Silver, the project’s lead scientist, said, by not using “human expertise in any fashion, we’ve actually removed the constraints of human knowledge.”

But doubts persisted. AI systems still could not process language well or generalize across tasks. Different problems still required different kinds of neural networks. To most observers, systems such as AlphaGo remained merely fascinating gimmicks, and any prospect of significant national security relevance for AI still felt very far away. But believers saw a clear path: Find better algorithms, add more compute and data, and scale up.

"Sobre este título" puede pertenecer a otra edición de este libro.