Rating: 8.5/10.
Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI by Karen Hao
Book about the recent history of AI, focusing on OpenAI, but also covering the broader industry as well. It starts with the life of Sam Altman, who was born in 1985 to a middle-class family and quickly proved himself to be a capable leader and entrepreneur. He sold his first startup at the age of 26 in location tracking, and then he developed connections and invested in many top founders of the tech industry. He was seen as charismatic, but having a reputation of not always being truthful at times.
OpenAI was founded in 2015. It was led by Sam Altman and Elon Musk under a nonprofit structure, and they started by gathering many of the top AI researchers and paying them top salaries to avoid being poached by competing tech companies, though without any clear direction. Eventually, Altman realized that achieving AGI would require an enormous amount of capital that was incompatible with the nonprofit structure, and at that point, Elon Musk left. The company shifted from Dota 2 gameplay agents to large language models, which at that time was the GPT-2 model. Based on this model, they received an investment from Microsoft following a promising demo.
A brief history of AI – early AI was dominated by symbolic approaches, which worked well for demos but were brittle in the real world. Deep learning had much better commercial value and worked quite effectively, and due to the commercial value, the high salaries drew research away from academia and other places into industry labs. Some were skeptical of this trend, but over the years, the technology repeatedly exceeded expectations of what it could do, even though there were issues such as hallucinations and biases, the progress was rapid. When the author visited OpenAI’s lab in 2019, she met with some of the senior executives to understand the company’s goals, but the sense was that the objectives were rather vague and nonspecific – they were just trying to move towards AGI as quickly as possible, before anyone else.
OpenAI first dabbled into language models in 2018 when Alec Radford adapted the transformer architecture, which was invented for translation, into the GPT-1 model. At one point, Ilya Sutskever would constantly push the team to scale bigger, as he was convinced that scale was the most important factor and that adding more scale would improve performance – a view that was not common at the time. That is what they attempted, and in 2019, they trained GPT-2 with 1.5B parameters, which they deemed too dangerous to release. At the same time, they ambitiously started working on GPT-3, which used as much scale as they could manage, vast amounts of data from the entire Internet.
As they launched GPT3 to developers via API in 2020, there were many conflicts between teams concerned about safety and applied teams trying to commercialize the model. For example, OpenAI’s DALL-E model had strong safety protocols, but this led them to lose traction to more aggressive startups like Midjourney and Stability AI. Instead, for later releases, they tried strategies like having a research period with more blunt filtering as the company tried to launch quickly and then more slowly improve the filtering mechanism. Due to concerns over safety issues, Dario Amodei and several others left to start Anthropic. Safety issues affected other labs too: Google safety researcher Timnit Gebru published a paper on the environmental effects of large model training; she was initially asked to remove her Google affiliation before ultimately being terminated from Google.
A few chapters talk about the experiences of workers in less developed countries and how AI affected their lives before release. AI companies frequently used workers from low-cost countries to annotate training data, typically through intermediaries like Sama, Appen, and Scale AI. Much of the work involved content moderation, which was an extremely unpleasant task that often required annotating disturbing content generated by versions of the model itself, frequently leaving the workers traumatized. Initially, Venezuela was a preferred country since they had a population of desperately poor but highly educated people with internet access. Later, this shifted to Kenya, which had native English speakers instead of Spanish speakers who required translation. The work paid well relative to local salaries but was highly unpredictable, causing workers to set up alerts and rush to their computers whenever annotation tasks appeared in order to claim them before others.
Another issue that affected developing countries was building data centers in areas with available electricity but which often were short on water. As data centers expanded, they required vast amounts of power and fresh water for cooling, with magnitude often similar to that of a small city. Google and Microsoft set out to build their centers in Chile and Uruguay, but both were met with resistance from locals who had to drink untreated water, whereas data centers would use as much freshwater as thousands of people.
Around 2022, OpenAI decided it was time to build a chat interface as quickly as possible before Google got wind of the technology and started to catch up. The ChatGPT launch happened in November 2022 and was initially intended to be quite a quiet experiment, since similar models had already been available in the API. However, this unexpectedly gained millions of users within days. They struggled with the infrastructure capacity to keep up, and after ChatGPT went viral, OpenAI rushed to expand their headcount.
The last section of the book is about issues following the launch of ChatGPT. As the technology continued to develop at an impressive pace, insiders at OpenAI became increasingly concerned by Sam Altman’s nonchalant attitude towards AI safety, often saying different things to different people to manipulate them. Around November 2023, this culminated with several board members and senior leaders at OpenAI staging a coup to fire Altman, including Mira Murati and Helen Toner among others, after realizing a pattern of lying and manipulation. Over the weekend, it became apparent that this movement did not gain enough traction within the company, that firing him would destabilize the entire company, he was reinstated, and the charges were dropped, with a commitment to transitioning the company eventually to a for-profit structure.
Soon after this incident, OpenAI launched its GPT-4o model and voice mode, again with scandals as actor Scarlett Johansson claimed they used her data without consent. Claims also surfaced that departing employees had to sign non-disparagement agreements or risk losing their vested equity. More recently, OpenAI has started to lose talent to other emerging labs like Anthropic or Thinking Machines, and that is where we are today at the beginning of 2025 when this book is published.
Overall quite an insightful read as the author is a journalist who has been tracking the company since its early stages and provides some details of inner workings of one of the most secretive companies in tech. It sheds light on some issues involving AI development and large language model development that are not widely publicized within the tech industry, especially issues involving data annotation, ethics, and data center costs. However, while there is extensive documentation of problems, she proposes few solutions (other than I guess to not develop AI), and not much in how the situation should be improved or better regulated given the competitive nature of frontier model development. Nevertheless, it is well researched and even as somebody who works as an engineer in the AI industry many details were new to me.
