Empire of AI — a book review. I recently finished the book Empire of…

In an age where AI development is moving at a breakneck speed and many AI companies are wielding unchecked power and unprecedented investments, the work of a journalist like Hao becomes very important. We are used to our elected representatives being scrutinized, but this is far from true for big tech executives who potentially hold more power and manage bigger funds.

Hao’s focus in the book is OpenAI, but she also writes about the broader AI and tech landscape. In the book, there is a timeline of the history of AI. Even though it is today being portrayed as a breakthrough technology, the research field of AI is not new. It was founded as an academic discipline 1956, and there had already been research done before that. The Turing test was described in a paper published in 1950. Research within AI was long a topic of little interest. Even though the first chatbot (Eliza) was created in 1966, and neural networks originated in the 60’s, there was too little computing power or data to build anything meaningful. In the 1980’s AI saw a boom built on expert systems. This led to increased funding that dried up again in the 90s. In the 2000’s the discipline of machine learning grew, and neural networks were dusted off to become the major technology utilised. With enough data (the whole internet) and enough compute (cloud and GPU clusters), companies like OpenAI could take off.

OpenAI started as a foundation to compete with Google because of the fear of what would happen if Google reached AGI. Focus was on safety and transparency. The argument was that “we have to make progress in this area before anyone else does”. The question is whether this could have happened anywhere other than in Silicon Valley? This was a unique concentration of talent, money, and power that could undertake such an endeavor. Arguably, many other companies and states are playing catch-up, driven by the same fear. This time, the threat is OpenAI itself. Through Hao’s portrait of how OpenAI came to be, it is interesting to see how much power within tech is centered around a few individuals, and how many decisions they make that affect the whole world. Yet their power is unchecked, and decision-making is far from transparent.

One overarching theme from the history of OpenAI is the battle between AI safety and commercialisation. It has been a constant struggle throughout the company’s history. There have been plenty of promises made in the name of AI safety, but commercialization has always won the battle. Actions taken by the company have prioritized a commercial strategy. ChatGPT was meant to be a research preview, and OpenAI was surprised by the product market fit. The technology was already old and available to the public through an API. But the launching of a chat product took the world by storm, melting OpenAI’s GPU clusters. Even though they started with transparency in their research, OpenAI held back research results in the name of safety, only to build a product around the results and launch it to the public. Today, it is clear that OpenAI has a very commercial strategy of launching products with little focus on safety and testing.

One thing that surprised me in the book was the amount of manual labour that goes into training an AI model. The whole concept of reinforcement learning is nothing but humans teaching a model what is good and what is bad output. This work is done by “data annotators” in countries like Kenya, Venezuela, and other places where people are ready to do gig work for very little money. They spend time looking at pictures, reading texts, and annotating data. Often, they have to rate texts and images that describe violent, sexual acts, leaving the workers traumatized. With this as a background, the AI companies’ promises of reasoning and intelligence fall flat. If every LLM needs this much human training (everything from learning it to telling jokes to writing poems), is it intelligent? And can we scale it?

The question of scaling is repeated throughout the book. What led OpenAI to where it is today was the possibility to scale in compute and scale in data. But this feat is becoming increasingly hard. LLMs are already trained on the whole internet. AI companies are already spending unprecedented amounts of money on infrastructure. Infrastructure that is using an unprecedented amount of resources (electricity and water), leaving communities depleted. The scaling is already unsustainable, so what else is there to scale on? According to Hao, much of the success of OpenAI comes from scaling, not from any new research. As the focus has been on commercialisation, resources for researching new areas and think long term has been held back. As AI companies are now turning to tricks like synthetic data and optimizing the use of GPUs the question is if this will be enough for new breakthroughs. As the AI companies (and some states) are now locked in a race towards AGI, they keep crossing lines and do scales that were hard to imagine only a few years ago, without much thought to the consequences.

Hao describes the cost of using AI. The cost in resources, the cost of human sanity, the cost in investments. She never asks the question directly, but it is in the background — is it worth it? Is what we are getting out of AI worth the costs? There is a hope that with AI (or perhaps AGI), we can automate a lot of general tasks. But as we pour money and resources into these companies, they are far from being profitable. The question is how many of them will survive in the long term, and how many will create something that lives up to the cost. Considering the already unsustainable pace, and what AI is being used for, I start to question wether we are using this invention for the right purpose. Is it worth having an AI summarize long articles, written by another AI, if it causes whole communities to have mental breakdowns and uses up the drinking water of a whole city?

So far, most AI companies have operated in a “move fast and break things” mode. With little respect for things like regulation, IP rights, minorities or underdeveloped communities. They keep amassing investments, and hiding the costs (both human and monetary) from it’s users. Looking back at the results big tech gave us with this method in the beginning of 2000, there was a lot of invention, but also a lot of gruesome sideeffects. Now the scale is even bigger.

Hao also describes how you can do good things, in a sustainable way, with the technology. In the book, Hao describes a project in New Zealand to preserve the Maori language. Similarly, I’m impressed with how the Danish Alexandra Institute has collected Danish dialects into a speech-to-text model. It is a nice counterweight to how innovation can be used for good, when data is donated and the outcome is meaningful.

I would recommend anyone working with AI to read (or listen to) this book. Even if not taking facts stated in the book at face value, it is important to understand how the speed of development is unsustainable, and the sideffects and costs that come with it. Bigger is not always better, and having a balanced, informed, view of current development always helps making strategic decisions.