Empire of AI

Book Review

Empire of AI by Karen Hao, 2025, Penguin Press

This is an important book that should be read by everyone working with or using AI. It paints the big picture of what is going on, at least through January 2025. To keep up it is necessary to check the headlines every day. This book will help you figure out where you can use AI or where it will harm you or business. Both outcomes are likely true. The author has significant technical training, but is not an AI insider, just a very astute observer.

A significant theme in Empire of AI is the story of Sam Altman and his pursuit of AGI, Artificial General Intelligence. This led to the founding of OpenAI and ChatGPT. Arguably these accomplishments have spurred today’s tremendous interest in AI. This activity eclipses anything I have seen in tech. Some of the sub-themes in the book include: is OpenAI pursuing AGI for good of humanity, or simply to enhance its own power and influence? What has happened at OpenAI to see the departure of many top AI scientists, memorably to Anthropic and other startups? Just normal growing pains or something within the organization itself? Will OpenAI keep its edge or will others move ahead? Today’s headlines talk about Meta’s recruiting of AI scientists around the world. Where will they end up?

Another sub-theme in Empire of AI is the downside associated with the technology. One is excessive power consumption associated with both model training and inference. Research papers have documented this power consumption. AI vendors claim their models use less power, but hard data is hard to come by. So we know that AI will consume huge amounts of power, but we don’t know exactly how much. Another downside, not reported widely, is the extensive labor force needed to train many models. These use RLHF, or Reinforcement Learning with Human Feedback. An LLM needs someone to evaluate its responses and people hired to do this task have often been from developing countries and are paid very low wages. Meta recently announced that it would partially acquire Scale, one of the companies offering training services, for “only” $29 billion. Since the “intelligence” of an AI model is in its training parameters and not its algorithm, you get the picture of the importance of human in the loop training.

A major sub-theme in the book was the conflict between AI safety adherents and commercialization interests. The so-called “Boomers” and “Doomers.”  OpenAI itself was founded to benefit humanity. However, it moved more towards commercialization and is still moving in that direction. Original team members were focused on safety issues and research. As commercialization took hold with ChatGPT and the GPT API, safety teams moved into a backseat role. Models now are rolled out and user feedback is recorded. This seems much like other commercial software development. To me, we are far from AI taking over the world; taking over jobs is much more likely.

One more major sub-theme was Sam Altman himself, his personality, his family, and his quirks. These I take as opinions of the author.

Reading this book will put you only 6 months behind the technology. The rest is up to you;  enjoy the ride.