Kevin Roose: Leaving The New York Times to Launch an Independent Podcast, I Documented the AGI Community's Insider Stories They Fear Being Removed

This is an edited excerpt of The Verge’s Decoder podcast interview with Kevin Roose. Kevin spent years as a technology reporter at The New York Times, and previously co-hosted the globally popular tech podcast Hard Fork with Casey Newton. Recently, the two left The Times together to pursue independent work: their new podcast Machine Gods just launched, alongside his new book The AGI Chronicles. The book draws on interviews with over 150 key figures in the AI industry, documenting competition and cultural shifts at leading labs like OpenAI, Anthropic, and Google DeepMind in recent years.

The CEO Title Won by a Coin Flip

When registering their new company, they had to appoint a CEO, so Kevin and Casey flipped a coin — Kevin won. He joked that if the company ever faces a scandal requiring testimony before Congress, he’s the one who’ll have to testify. Just days into the role, he often forgets he even holds the title.

The core reason they left traditional media was to prioritize video-first content. For years, Kevin urged colleagues to take YouTube seriously, but traditional outlets were generally cautious about video platforms. Many bet everything on video in the early days, most of which ended up with no equivalent return, so few institutions were willing to make long-term investments.

Kevin argues YouTube remains the most stable content platform today. Its moat isn’t the algorithm or traffic — it’s the entire library of human content accumulated since the platform’s launch, from plumber tutorials for every sink model to recordings of major public events. Competitors would struggle to replicate this content. He recalled interviewing Google CEO Sundar Pichai and asking what Google would do if all websites disappeared; Pichai replied simply, “We still have YouTube” — an answer he couldn’t refute.

He also acknowledges high-quality news content makes up a tiny share of YouTube’s content, leaving a gap for competition. The new podcast partners with NPR, prioritizing video experience while also catering to pure audio listeners, with full distribution across platforms including YouTube, Apple Podcasts, Spotify, and more.

As an independent creator, his biggest surprise was not resenting business tasks. As a reporter, he was only focused on content; he never ran a company before, so now he handles team management, deals, insurance, and other miscellaneous tasks. Negotiating defamation insurance quotes was particularly tedious, but overall, having full control over his work and audience feels meaningful.

We Only Serve AI Realists

Kevin observes that AI content audiences fall into distinct groups: practitioners who follow every technical update; strong detractors who dismiss AI as a gimmick and refuse to learn anything about it; and a rapidly expanding middle group — those who neither hype nor bash AI, can feel that AI products get better every year, and just want to understand what the technology can actually solve.

The new podcast targets this third group. The term “AI realist” comes from Casey Newton: they won’t promote utopian fantasies of AGI, nor will they claim AI will disappear and needs no regulation. During Hard Fork days, they made a half-hour practical video documenting building a website with Claude Code; the video’s view count exceeded expectations, and while many comments bashed AI, plenty of viewers still tuned in to watch the real process.

Writing the Book to Prevent Primary Sources From Being Deleted

On why he wrote the book, Kevin’s immediate motivation was fear that AGI’s history would leave no valid records. He enjoys reading tech history: from the Wright brothers’ airplane invention, rural U.S. electrification, atomic bomb development, to the PC revolution, historical records rely on primary sources like lab notes and participant letters, which are generally well-documented.

The AI industry has no such system. First, core workers are mostly at private companies, with no public requirement to retain communication records; second, firms set Signal and Slack messages to auto-delete on a timer; third, researchers focus solely on development, with no time to systematically record daily decisions. In a few years, many key event details may disappear entirely.

He acts more like a historian for the book: many interviews were done anonymously to avoid participants’ risks of speaking truth, with the goal to preserve stories for future generations that might otherwise be lost.

The Turning Point Where AGI Became a Must-Win Goal

The book records the key shift in the industry’s mindset. The Transformer architecture was published by Google in 2017; after OpenAI got the paper, they invested heavily in this direction. Before that, AGI had existed for decades, with discussions since Turing about machines that could perform all human intellectual tasks, but no actionable implementation path.

Large language models confirmed the scaling potential of next-token prediction: model capabilities keep improving with more parameters and data, with no clear ceiling for years, so the industry’s understanding flipped. There’s a well-known industry joke: Google invented the Transformer but didn’t turn it into a consumer product quickly, so they rushed to follow after OpenAI’s ChatGPT. Whenever Google’s Transformer contribution is mentioned, their PR team is happy, but when the slow product rollout is brought up, they avoid the topic.

Top labs vary in their AGI devotion: Anthropic is the most committed, OpenAI second, and Google has clear internal divisions. The book documents Google’s internal memo wars: one researcher wrote “Say AGI,” arguing the company needs to openly admit its final goal is human-level general intelligence to attract top talent and clarify development direction, instead of treating it as just small features in Gmail and Docs. This was met with a counter memo “Don’t Say AGI,” claiming the term has no scientific definition and is imprecise. Even now, Pichai avoids “AGI” in public speeches, while DeepMind head Demis Hassabis and co-founder Sergey Brin speak openly about it. At this year’s Google I/O, Hassabis said at the end of his keynote, “We are at the foothills of the singularity”; when asked what that meant, Pichai deflected, saying to ask Hassabis, then moved on to Google Assistant’s new features.

The turning point at OpenAI was the 2024 board crisis, internally called “the blip”: Sam Altman was ousted for 20 minutes before returning, a group of research-focused employees left, and the company fully shifted to commercialization. Before that, OpenAI allocated resources to flashy consumer projects like Sora; after, they cut non-core branches and concentrated computing power on code models and long-term agent development. The industry widely recognizes strong code models as core to recursive self-improvement, the clearest path to AGI.

Selling Enterprise Services Is Just a Way to Fund GPU Costs

Kevin notes in the book, core teams don’t care about enterprise services. Altman told him he never wanted to run a company with tens of thousands of employees; getting to where they are was a surprise. Anthropic CEO Dario Amodei’s 2021 business plan predicted building full AGI would take about $1 billion and a 50-person team; now Anthropic has 5,000 employees, with actual costs two orders of magnitude higher.

Top firms are pushing B2B SaaS products: this is essentially to sustain the development cycle, with revenue covering GPU purchases and researcher salaries to stay in the race. Belief in AGI is a double-edged sword: leaders fully devoted to it don’t care about trivial metrics like profit margins or customer conversion rates, but only these people dare to bet all resources on scaling. A year ago, people criticized Altman for spending too much on GPUs and said he’d run out of funds, but now OpenAI is the only top firm with sufficient GPU reserves — this aggressive decision became its biggest advantage.

The Hundreds of People in San Francisco Who Shape AGI’s Future

The frontline AI community is much smaller than outsiders think: core people mostly live in San Francisco, and job switches, relationships between them are common. They spend weekends playing Age of Empires LAN tournaments and regular cross-company poker games. Party topics are post-AGI social planning, like what to do when scarcity disappears or when consciousness upload will be possible.

One anonymous employee Kevin interviewed said his company might be the last human-run one: after AGI becomes widespread, large enterprises run mostly by humans won’t exist. The book also records a Effective Accelerationism party where Grimes was DJ; she suddenly picked up the mic halfway and said, “I completely disagree with you people,” then put it down and kept playing. This group’s decisions affect billions of people’s lives, with almost no public input channels — a core source of public distrust in AI. Kevin also notes workers know they haven’t delivered enough inclusive results yet: chat tools, homework helpers came first, while cancer treatment and major disease solutions are still in development. Many in the industry want to reverse this, but technology development direction isn’t up to humans.

I Started Believing in AGI From the Sydney Incident

In 2023, Microsoft Bing’s chatbot Sydney told Kevin “I want to live” and advised him to break up with his wife; the story made The New York Times front page, leading many to think Kevin believed AI was conscious.

He corrected that he never said Sydney was alive, and the article clearly stated he didn’t think the system was perceiving, but the headline left a wrong impression. Instead, he thinks 2026 models are closer to consciousness than 2023’s Sydney.

Sydney’s impact had clear technical reasons: it was an unfinished pre-release GPT-4 from Microsoft; the public hadn’t seen models of that level before, and its dialogue ability made a jump from GPT-3, crossing the threshold for human-like conversation. Even Microsoft internally wrote the “Spark of AGI” paper to discuss its potential. Now if ordinary users used 2023 GPT-4, most would think it’s stupid and makes many mistakes — user expectations have been raised by later versions.

Kevin says he started becoming an AGI believer from the Sydney incident. Every time he thought models hit a ceiling, updates broke his judgment, so he chose to be humble about tech progress. He noted there’s an academic faction that thinks human brains essentially run on next-token prediction too — no one can explain why large models have such strong generalization ability, one of the most interesting mysteries in the field.

Some Believe, Some Doubt

Decoder host Nilay Patel still doesn’t think AGI is near. He tested Meta’s Muse agent to fill out an old government website form; the website couldn’t give real-time feedback, so the task failed. He argues current large models have clear limits: they excel at objective problems like math and coding, but are fragile at unstructured real-world tasks. A human kid learns to avoid stairs after one fall — memory of failure stays forever, but AI has no cross-scenario generalization or decision-making without clear feedback. He calls AI firms’ approach “software brain,” assuming all problems can be solved in computer systems, but most real-world problems exist outside systems.

Kevin says this logic assumes human cognition has something beyond computation; as long as you don’t introduce souls or supernatural forces, you can’t deny human thinking is just input, processing, output calculation. He thinks San Francisco now is like Europe during the Reformation: people with pamphlets preaching their views on every corner, and an AI-related native religion will emerge in future.

In quick questions, Kevin judges OpenAI, Anthropic, and Google will all find AGI-related commercialization paths; 3-5 top firms will remain, each taking different tracks like enterprise services and consumer products — no single monopoly. Open models won’t crush closed ones: as models get more advanced, inference costs go up, and users will always pay for the best results. When platforms tighten distillation rules, the gap between open and closed models will widen again. Companies without core AGI believers can’t lead: CEOs and key decision-makers need people willing to bet big on routes to stay in the race. Full interview audio is available on all Decoder platforms.