Kevin Roose on The AGI Chronicles: Three Labs Are Leading the Race to AGI

Kevin Roose, a technology journalist and author, said on Bloomberg This Weekend that a relatively small group of researchers and executives at three companies, OpenAI, Anthropic and Google DeepMind, has defined the push toward artificial general intelligence. The occasion was his new book, The AGI Chronicles. With hosts David Gura and Christina Ruffini, he discussed two questions: who really controls the technology, and how to regulate systems that developers want to take beyond human level.

Three labs and a narrow circle of people

Roose's thesis is simple and uncomfortable. The race to AGI looks like a mass movement of the whole industry, but a handful of people decide its pace and direction. Only a few players have the resources to train frontier models, so the membership of this group barely changes.

The companies' histories are intertwined. OpenAI appeared in 2015. Anthropic was founded in 2021 by people who came from OpenAI. DeepMind has existed since 2010; Google bought it in 2014 and merged it with Google Brain in 2023 to form Google DeepMind. Many key figures worked side by side for years, argued and parted ways. So "a small group" is not a metaphor here but a literal description of a circle of people who know each other personally.

What we know from the interview

The segment itself is short, and Bloomberg's piece does not include a detailed summary of the book. The documented facts are these:

• Author: Kevin Roose, technology journalist.
• Book: The AGI Chronicles, new.
• Venue: Bloomberg This Weekend, hosts David Gura and Christina Ruffini.
• Companies at the center of the thesis: OpenAI, Anthropic, Google DeepMind.
• Main question: who controls the technology and what risks it carries.
• Second topic: the difficulty of regulating increasingly capable systems.

Why regulation can't keep up

Regulators face two problems. First, the target is moving. Systems change faster than laws are written and passed, and definitions such as "frontier model" go stale within a couple of releases. Second, expertise is concentrated inside the same three or four companies that are supposed to be regulated. For oversight, that is a structural weakness.

The record of recent years bears this out. The European AI Act entered into force in August 2024, but its requirements are being phased in. California's SB 1047, a bill on the safety of frontier models, was vetoed by the governor in September 2024. These examples do not come from the interview. They come from the broader context, and they show the same gap Roose describes.

How businesses and developers should read such claims

The book and the interview offer no new technical data, but they provide a useful frame for practical decisions. If the development of frontier AI depends on a narrow circle of companies, then your dependence on their APIs is concentrated in the same hands. It therefore makes sense to build provider switching into your architecture and not tie a product to a single model.

It is also worth following the labs' public safety policies. These documents show the conditions under which companies are prepared to slow down or halt releases. For teams that embed agents and coding assistants into their workflows, this is a risk factor on par with token prices or rate limits. The full list of Roose's arguments will become clear only after reading the book itself, but the general conclusion is already visible: the question of who controls AGI has become a separate risk for the market.