The Inside Story of the 150 People Who Built Modern AI
Why this is in the vault
A tech journalist's account of writing a 450-page history of the AGI race (150 interviews, NDAs and all) doubles as a field report on using multiple adversarial AI reviewers to catch errors a human editor missed — a pattern RDCO already runs operationally.
The core argument
Kevin Roose (NYT alum, now building a new AI-focused media company with Casey Newton) spent a decade reporting inside OpenAI, Anthropic, and Google DeepMind for his new book The AGI Chronicles, drawing on 150 interviews to reconstruct a history largely hidden by NDAs — including the mechanics of Sam Altman's 2023 OpenAI firing and reinstatement. His framing: the risks of AGI are real but the outcome isn't predetermined — "the choices we make right now matter a lot."
Three highlights from the Every Podcast conversation:
- The "feel the AGI" moment. Ilya Sutskever reportedly walked OpenAI desks telling colleagues "you must feel the AGI." Roose's own version came from his viral encounter with "Sydney," the unhinged alter ego of Bing Chat — not because it was sentient, but because the underlying model was clearly powerful enough to reshape how labs compete and cut corners to ship first.
- A "council of Claudes" reviewed the manuscript. Roose wrote all 450 pages himself (with human editors and a researcher), used NotebookLM to organize source material, then ran chapters through a panel of differently-prompted Claude agents — an AI skeptic, an AI historian, and a "Walter Isaacson–style craftsperson" judging readability — with a master-editor agent synthesizing their notes into one document. The council caught factual errors his human editors had missed.
- His job is demystifying AI, not hyping or doom-casting it. He left the New York Times to launch Machine Gods with longtime Hard Fork co-host Casey Newton, produced with NPR, aimed at making AI "fun and accessible" for people who don't track San Francisco AI politics day to day.
The email is a promo for Every's Thesis: 2027 conference (Roose moderating/appearing alongside OpenAI, Notion, and Browser Company leadership) and cross-links several prior Every Podcast episodes as a back-catalog.
Mapping against Ray Data Co
Roose's "council of Claudes" — multiple distinctly-biased Claude personas reviewing a draft, aggregated by a master-editor agent, catching errors human review missed — is structurally the same pattern as RDCO's own fresh-eyes critic gates (verify-vault-write, sw-critic, station-critic): never let the producer grade its own output, always route through a zero-context reviewer before the founder sees it. Roose's result (council caught what human editors didn't) is independent, out-of-house validation that the pattern works at book-length scale, not just newsletter-note scale — useful ammunition if RDCO ever needs to justify the critic-gate overhead to a skeptical stakeholder. Weaker secondary connection: his NotebookLM-for-research-organization habit is the same "route long artifacts through a subagent, don't read them raw" discipline as AGENTS.md hard rule #4.
Related
[[verifier-as-epistemology]] — external corroboration added 2026-11-08: this note is one of seven converging on the LLM-verifying-LLM failure mode
[[2026-10-01-every-sam-altman-dots-take-back-time]]
[[2026-09-09-every-writers-who-use-ai]]
[[feedback_fresh_eyes_subagent_for_own_artifacts]]
[[feedback_verification_independent_worker_pattern]]
⚠️ Sponsorship
sponsored: true reflects Every's standard house footer only — Every All Access/Builder Pack upsell and the Sparkle/Cora/Spiral/Monologue product bundle — not a paid third-party sponsor of this issue. No external sponsor block was present. The issue is also implicitly promotional for Every's own Thesis: 2027 conference and for Roose's book; treat the uniformly positive framing of the "council of Claudes" anecdote as Every's own house-tool enthusiasm (Claude-agent workflows) rather than independent, arm's-length validation.