Kevin Roose: Anthropic "most AGI-pilled" lab, OpenAI close second

Roose frames his book as an act of preservation. He says the people building AI work for private companies rather than government, their Signal and Slack messages are set by lawyers to auto-delete, and the researchers are too busy to keep notes — so he used his access to capture stories before they vanished.

The interview traces the moment AGI moved from dream to mandate back to the transformer, the architecture Google developed in 2017 and published, which OpenAI then ran with. Google's internal argument, captured in dueling "Say AGI" and "Don't say AGI" memos, illustrates the split Roose describes, and he ties the ascendant pro-AGI faction partly to talent retention against OpenAI and Anthropic.

Roose also treats belief itself as a competitive ingredient. He recalls a year ago when OpenAI was criticized for spending so much on compute, and notes it is now the only major AI company not severely compute-constrained — suggesting the willingness to overbet, not just the technology, shaped the current field.