The Inside Story of the 50 Who Built Modern AI With Kevin Roose
Kevin Roose argues that AI's future is not predetermined, explains how he reconstructed its secret history, and shares how he uses AI as a tool while keeping the struggle that makes him better.
Watch on YouTube ↗Charlie's summary
- ▶ 4:07
Regulation likely needs a crisis, but Kevin wants mandatory safety testing and a way to slow the frontier first
Kevin doesn't expect federal regulation from the Trump administration unless something bad happens, as after 9/11. He hopes to avoid that. His first step is an industry certification: test models before release for autonomous cyber attacks or bioweapon misuse, as pharma tests drugs. That is “basic table stakes.” He has also come to believe slowing the frontier is possible, because Dario, Demis and Sam say it is moving too fast.
- ▶ 7:44
The history of AI was at risk of vanishing, so Kevin set out to record it quickly
Kevin says he knew the story only partly. He didn't know why the transformer took off at OpenAI, or how the ChatGPT release and reasoning models came about. Inside the labs everything is under NDA, and Slack and Signal messages delete themselves. He feared few records would remain to answer “how did we get here?” So he treated it as job one to get the story, and job two to do it fast.
- ▶ 11:22
Getting 150 insiders to talk: background interviews, snowballing, and having read their papers
Kevin says the first 25 interviews were hardest. He offered to speak on background and not name sources. After 25 to 50 it snowballed, as sources talked to each other. Some spoke because a rival had, wanting their own side told. The biggest tactic was sincerity: he read researchers' papers and cited other interviews, which showed he wanted depth, not a surface overview.
- ▶ 15:29
Dario's “big blob of compute” memo: scale general models instead of building narrow ones
Kevin singles out a 2017 memo Dario Amodei wrote at OpenAI. At the time, research went into narrow systems, like DeepMind's Go player. Dario argued for assembling “big blobs of compute” and pointing them at general problems like predicting the next word, trusting intelligence would emerge. Everyone at OpenAI remembers it reshaping their road map. Kevin says you can see the light bulb go off in real time. He published the first 10 pages on his Substack.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.