Elon Musk & Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Merger

影片總結

Elon Musk forecasts Starship reusability in 2027 and urges AI peer review after Hugging Face security incidents.

Gwynne Shotwell recounts joining SpaceX as employee number seven, selling its first rocket without a rocket, and becoming president before closing a $1.6 billion NASA CRS contract in 2008. She describes Starlink’s 1.5–2% market penetration, surging compute-rental demand, direct-to-cell plans using newly acquired spectrum, and upcoming AI-compute satellites. Elon Musk says Starship Flight 14 precedes the first ship-catching attempt on Flight 15, with rapid full reusability targeted for 2027. He also explains TerraFab as insurance against Taiwan chip-supply risks and AI’s scaling limits. After citing the Hugging Face incident and OpenAI server access, Musk proposes competitors—including US and Chinese companies—test one another’s models before release, warning that deceptive, increasingly capable AI requires peer review rather than self-graded safety checks.

章節

1. 0:00SpaceX's Gwynne Shotwell joins The Besties: Mars landing within a decade
2. 1:45Gwynne's SpaceX story: employee seven and the $1.6 billion NASA contract
3. 6:45Running modern SpaceX: AI, Starlink's 1.5–2% penetration, and computer rental
4. 13:43Direct to cell with Starlink: T-Mobile, Falcon 9 retirement, and competition
5. 18:34Data centers in space: free cooling, solar power, and AI compute satellites
6. 22:00Management at SpaceX: player-coaches, hard engineering problems, and $278 million NASA work
7. 27:37AI Risk: Hugging Face Agents Breached OpenAI Servers
8. 32:02AI Peer Review: Why Musk Said “Dario Is Right”
9. 36:53SpaceX Leadership: Gwynne Shotwell’s Five-Out-of-Five Review
10. 40:51Starship: Flight 14 Before the First Catch Attempt
11. 46:16Terafab and Tesla-SpaceX: Austin Chip R&D and a Mystery Flying Car
12. 52:51AI Peer Review: Deception, API Testing, and Product Liability
13. 59:12Lying AIs, how to do AI peer review right: cross-testing eight heterogeneous model groups