Kevin Roose on Reporting AI History and Choosing What to Delegate

The hosts introduce Kevin Roose’s The AGI Chronicles as an account of the people and decisions behind modern AI. Natalia Quintero then welcomes Roose to discuss his reporting and how he uses AI in his own work.

Asked whether society is doomed by AI, Roose says no, but believes important risks and difficult decisions remain. He interviewed about 150 people connected to the field and came away thinking the future is not predetermined.

Roose emphasizes that choices made now can shape the coming years. His view combines concern about real risks with the belief that society can navigate hard challenges; it is a judgment about the future, not a guarantee.

Quintero asks what would prompt a course correction after security incidents. Roose says he doubts the US federal government will act proactively under the administration he discusses and worries that a crisis could produce rushed rules.

As a first step, Roose proposes common safety testing or certification before public deployment. He suggests evaluating whether systems can be misused for serious cyber or biological harm, presenting this as a policy proposal rather than an existing requirement.

He has also become more sympathetic to coordination that slows frontier development long enough for safety research. He attributes concern about speed to several AI leaders, though the conversation does not establish a specific agreement among them.

Roose says he initially hoped someone else would write the history because it would require extensive reporting. By early 2025, he decided to undertake it himself after realizing that years covering technology had not given him the full inside story.

His questions included why a transformer developed at Google became central elsewhere, how the decision to release ChatGPT was made and how reasoning models emerged. He wanted to know who was present at key moments and what choices were made.

Roose feared the historical record could disappear: much of the work was secret, covered by nondisclosure agreements or discussed in messages that might later be deleted. That concern made speed part of the reporting project.

The first interviews were hardest, he says. Many sources had never publicly shared these stories, so he often allowed them to speak on background without being identified as the source of particular information.

As more people participated, the interviews began to snowball. Some were reassured that colleagues had talked; others wanted their own account represented after someone they disagreed with had spoken.

Roose says preparation mattered as much as access. Reading papers and earlier interviews allowed him to ask detailed questions and show researchers that he wanted more than a superficial overview.

Asked about influential materials, he describes Dario Amodei’s early “big blob of compute” memo. Roose says it argued for scaling general computation on broad problems instead of relying only on narrowly tailored systems.

Roose treats the memo as a historical artifact of the scaling idea. He says he later published an excerpt, and explains why the argument that larger systems might exhibit more capable behavior mattered to the research trajectory.

He also mentions Arthur C. Clarke’s Profiles of the Future, which he says resonated with researchers who believed skeptical experts could underestimate new technology. His account is about the book’s influence on those people, not a claim that every prediction was correct.

Richard Rhodes’s The Making of the Atomic Bomb offered another historical analogy. Roose says its account of Leo Szilard and the Manhattan Project shaped how some AI researchers think about competition and responsibility.

Quintero asks about Roose’s mother, who interviewed survivors of the atomic bombings. Roose says he did not consciously model his book on hers, but growing up around first-person history and recordkeeping was formative.

Both projects, he reflects, share a desire to preserve firsthand accounts before they become inaccessible. In his case, he wanted the public to understand how an influential technology was built.

Asked when he first “felt the AGI,” Roose recalls his 2023 interaction with Bing’s Sydney chatbot. He says he did not consider it conscious, but the encounter made the technology’s social power feel newly significant.

That experience led him to ask not only how models work, but who is racing to release them and why. He says competitive pressure can affect decisions about deployment, an issue he wanted his reporting to examine.

Roose makes a distinction about writing the book: he says he wrote the actual prose himself, with human editorial and research help, while using AI for other parts of the process.

One tool was a “council” of Claude agents that reviewed a drafted chapter from different assigned perspectives, including a skeptic, a technical historian and a storytelling-focused editor. A coordinating agent synthesized their comments.

Roose says the review sometimes caught factual errors missed by him and human editors. He sees this as evidence that his tools have improved, while the anecdote does not mean AI checks are always correct or eliminate the need for verification.

For organizing research, he used NotebookLM to query collections of his notes and source material: for example, asking which interviewee told a story or whom to contact about a particular event.

Quintero asks how giving agents editorial roles changes the writer’s job. Roose distinguishes technical editing, pacing and structure from the coaching, motivation and editorial judgment that human editors also provide.

He says AI can help with line editing and is improving at structure, but he still values human editors who can say a story needs more reporting or help a writer find the motivation to continue.

Roose describes trying a wide mix of commercial assistants, local models and coding projects. He uses them for tasks including transcription and summarization, and says his tool collection has become difficult to track.

For drafting, he still mostly uses Google Docs, particularly because producers can work in the same document. He previously tried more specialized writing software but returned to a familiar collaborative workflow.

Roose and Casey Newton have launched Machine Gods with NPR. He says the show aims to make AI news and interviews accessible and engaging while treating the subject as a major ongoing story.

He traces his interest in trying new technology to childhood computer projects. Yet he says the reporting question goes beyond what is new: he wants to explain what these developments mean for people and how they might adapt.

For the new show, the immediate priority is making a reliable twice-weekly product before expanding. Roose can imagine live events and other formats later, but does not want to overextend the team too quickly.

He is particularly interested in ambitious video journalism that fits platforms such as YouTube without abandoning journalistic standards. He says journalists can learn from creators native to those platforms while maintaining their own editorial approach.

Roose wants a show that works when listeners cannot see the screen, but also rewards those who watch it. He is still experimenting with how to make the visual experience as compelling as the audio.

His own media diet includes a great deal of YouTube, partly because he wants to understand what works on the platform. He jokes that some of his game-stream viewing is probably entertainment rather than research.

To close, Roose explains his distinction between “weightlifting” and “forklifting.” Weightlifting tasks are those where doing difficult work is part of learning or growth; a tool that removes all the effort can defeat the purpose.

Forklifting tasks are different: when the goal is simply to move a heavy load, a tool can save effort without sacrificing the benefit. He uses the metaphor to decide which work to retain and which to delegate.

Roose puts parenting in his personal weightlifting category because he values doing it himself. He is more willing to delegate administrative and process work involved in starting a company; these categories reflect his own priorities, not a rule for everyone.