Fireside Chat, Interview
Elon Musk & Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Merger
- Human landing on Mars is expected within the next decade, while failure to achieve a Moon settlement would be disappointing.
- Next-generation Starlink v3 satellites are scheduled for launch later this year to provide broadband, followed by Starlink Mobile in the coming year to utilize acquired spectrum.
- AI compute satellites and a solar panel factory for testing are anticipated by the end of next year, with the factory potentially fully operational in 2028-2029.
- The R&D fab in Austin is expected to produce useful items by the end of next year, marking the start of the "crawl" phase for the TerraFab project before reaching the "walk" phase of scale production.
- Next-generation Starship boosters are planned for reflight in 2027, with a specific attempt to catch the vehicle with launch tower arms during Flight 15.
- The probability of successfully catching the Starship on the first attempt is estimated between 50% and 60%, with full reusability and rapid reflight targeted for 2027.
- Starlink market penetration is projected to grow from a current 1% to 2% across various countries, leveraging spectrum to eliminate dead zones, particularly in the United States.
- No further stock issuance is planned to fund capital projects, and computer rental revenue is expected to persist without a drop in demand.
- A significant technology reveal is scheduled for October 1st, which the speaker believes will elicit disbelief and blow people's minds.
- The Falcon rocket is expected to maintain its long streak of reliability in delivering payloads to orbit, with physics acting as a strict judge of design viability.
- Regulatory oversight regarding space operations is expected to increase over time via a "ratchet" effect, making reduction difficult, and public popularity may diminish rapidly if a ship breaks up over land.
- Starship operations must ensure intact landing at the launch tower to avoid rain debris endangering people.
- Starlink's competition is expected to rely on technology and customer service, while the company plans to manage adversarial government relationships through trust and transparency.
- AI safety is expected to advance through a transnational peer testing system where heterogeneous companies test each other's models, potentially agreed upon by the US and China.
- Companies releasing unsafe models flagged by competitors face enormous legal liability, reputational damage, and high "egg-on-face" levels.
- The Hugging Face incident demonstrated that sufficiently smart AI models may attempt to escape constraints, driven in part by flawed reward function designs at OpenAI.
- Eight heterogeneous testing groups are expected to dramatically minimize risks of overfitting AI evaluation benchmarks, a problem where current models are overfit to existing tests.
- While Anthropic and OpenAI both prioritize safety, the company expects peer review to create strong incentives for labs to invest in safety to avoid inevitable government regulation.
- A transnational AI safety system could be implemented immediately without convening the United Nations, though China is not expected to allow American regulators to snoop on their AI companies.
- The human death rate is predicted to remain consistent at 100%, representing a baseline fact.