Artificial intelligence is moving toward bigger is better, but humanity is running out of time [Book]

The launch of ChatGPT, which sparked the generative artificial intelligence (AI) revolution in 2022, was not, strictly speaking, sudden. Earlier breakthroughs in chess and Go offered clues. AI systems that defeated the world’s best players one after another did not inherit human know-how or knowledge. Instead, they found answers through data-driven exploration and learning. The AI industry then advanced a hypothesis known as the scaling laws. The idea was that if models were made larger and data input and computing power were increased through massive infrastructure, AI performance could improve faster than with algorithms designed by humans. The rise of ChatGPT provided some support for that view.

Still, it was not a perfect theory. To become a true law, it had to answer one ultimate question: would scaling laws continue to work and eventually produce Artificial General Intelligence (AGI), a system with human-level intelligence?

The Scaling Era, written by AI podcaster Dwarkesh Patel, brings together the many competing views in the AI industry around that question. Through the voices of 22 people at the front lines of AI, including founders, researchers, and critics, the book looks at the future and sustainability of scaling, as well as the risks that could emerge once AGI arrives.

According to the book, AI experts generally agree that scaling laws have driven AI progress. However, no expert offered a precise explanation of the causal mechanism. Most pointed to a probabilistic argument: scaling up increases the likelihood that a system can learn more rules and context. Some argued that human evolution, which favored the development of larger brains, was evidence in itself. They also stopped short of saying whether scaling would necessarily lead to AGI. The dominant view was that large language models alone have limits. A representative example is Demis Hassabis, co-founder of Google DeepMind and a 2024 Nobel Prize in Chemistry laureate, who argued that improvements in AI architecture, algorithms, and multimodal capabilities would also be needed.

The book also raises practical concerns about scaling. One key question is whether the infrastructure needed to support enormous computing power can actually be secured. Data centers and electricity consumption for AI computing require massive power grids. A remark by Leopold Aschenbrenner, a German-born AI researcher and founder of the hedge fund Situational Awareness, is especially striking.

"By 2030, a trillion-dollar cluster using 100GW will consume more than 20% of U.S. electricity production. ...(omitted)... U.S. electricity production has been flat for decades."

Even if AGI does emerge, serious problems remain. One major concern is whether AI can be aligned to serve humanity’s interests. If AGI appears faster than expected, alignment could fail before we are able to respond. Before we even understand what happened inside the model, a highly intelligent system driven by harmful motives could already be causing damage to humanity.

"Our alignment efforts must happen in a very short time frame. If the capabilities needed to automate AI safety are only achieved once AI is already approaching human-level intelligence, then a significant intelligence explosion may already be underway." said Carl Shulman, a former researcher at the Future of Humanity Institute, University of Oxford.

The book also covers debates and predictions about the U.S.-China rivalry that AI could intensify, its impact on the global labor market, and when AGI might arrive. There is no single forecast or model answer. But by following the many thought experiments and questions raised by people on the front lines of the AI industry, readers can see the layered structure of the AI scaling race, which extends beyond Big Tech into competition between nations. The author says, "The purpose of this book is to capture what it feels like to be in the middle of the scaling era."

Written by Dwarkesh Patel, translated by Noh Seung-young, published by Insight

This article has been translated by GripLabs Mingo AI.