'Scaling Law' Behind AI Boom Reaches Critical Turning Point

AI Empire: Power, Capital, Labor by Karen Hao (Published by Saengak-ui Him)

ChatGPT, the generative artificial intelligence launched in November 2022, has brought revolutionary changes to the world. Using generative AI to search for information, draft documents, and develop business plans has now become routine. AI has entered a phase of fundamentally transforming core societal functions including healthcare, education, finance, law, media, and culture. OpenAI, at the forefront of these changes, is the world's most closely watched company.

OpenAI originally started as a nonprofit organization. It was founded with the goal of developing unprecedentedly powerful AI for the benefit of all humanity, not for economic gain. Today, however, OpenAI has pivoted toward commercialization in pursuit of profits. Serious side effects have emerged from the generative AI development process, including unauthorized use of copyrighted data, labor exploitation, and environmental pollution.

The author, a former Wall Street Journal reporter, exposes the hidden underside of OpenAI and ChatGPT's success story. Based on interviews with approximately 260 industry figures including about 90 current and former OpenAI employees, along with internal documents and correspondence, the book vividly reveals the AI industry's true face.

Among OpenAI's various harmful effects, the author focuses on the elimination of diversity in AI research. OpenAI co-founder Ilya Sutskever concluded that dramatically improving AI performance required massive increases in computational resources. The scale of computational resources is determined by AI chip processing power and the number of chips. OpenAI discovered that computational resource volume had been doubling every three to four months over six years—far faster than Moore's Law, which states semiconductor processing power doubles every 18 months. OpenAI called this the "scaling law."

However, the author points out that the scaling law, like Moore's Law, is merely an empirical rule rather than a law of physics. Moore's Law was a hypothesis that Intel founder Gordon Moore included in a 1965 article for Electronics magazine. The author argues it represented the pace of advancement Moore believed his company could achieve and chose to pursue. As the semiconductor industry followed Moore's choice, Moore's Law became a self-fulfilling prophecy—and the same applies to the scaling law. "Their belief in scaling is now accepted like doctrine across the technology industry," the author notes. She also argues the scaling law is no longer sustainable as the resources and data required for AI development grow exponentially.

The author compares OpenAI to an "empire." To develop AI aligned with their vision, they are seizing resources: works by artists and writers, data recording countless people's experiences and observations, and land, electricity, and water needed to operate massive data centers and supercomputers. This, the author diagnoses, is the essence of AI power holders. The book also addresses how, as low-quality data usage increases for generative AI model development, workers in developing countries earn an average of less than $2 per hour filtering harmful content.

The author's conclusion is that AI need not look the way it does today. While ChatGPT, large language models, and generative AI applications currently dominate attention, this represents just one of many possible manifestations of AI. The author particularly emphasizes that we must not let a handful of Big Tech companies like OpenAI determine AI's future—we must reclaim control over AI. To this end, policymakers should strengthen regulations preventing AI companies from using people's data and work without permission, while research funding institutions should promote diversity in AI research to explore fundamentally different alternatives.

Governments and corporations worldwide, including South Korea, are nurturing AI as a future growth engine while following OpenAI's strategy. But as the author argues, future AI technology is not predetermined. The book merits attention for stripping away illusions created by excessive expectations about AI and presenting a balanced perspective.

672 pages, 28,800 won.