Elon Musk on DOGE, Optimus, Starlink Smartphones, Evolving with AI, Why the West is Imploding

Summary

• Musk returned to Tesla’s Palo Alto engineering headquarters after a Washington side quest, calling government “basically unfixable.” He said interest payments exceed the U.S. Defense Department—or “War Department”—budget and warned: “If AI and robots don’t solve our national debt, we’re toast.”
• Optimus Version 3 is consuming more of Musk’s mental cycles than any other single project because essentially human manual dexterity, a reality-navigating AI mind and very-high-volume manufacturing remain the three missing capabilities. At one million units annually, he estimated production cost around $20,000, “maybe $25,000,” including an AI chip costing roughly $5,000-$6,000 or more; selling price would depend on demand.
• The near-term Optimus risk is a humanoid supply chain that does not exist: Tesla must recreate it from scratch through substantial vertical integration, designing every motor, gearbox and piece of controlling electronics from first principles. The hand and forearm account for most of the engineering difficulty. Musk said specialized robots are easier, but doing everything humans can do requires a humanoid; the host added that such a robot would be backward-compatible with a human-built world, and Musk agreed.
• Tesla has Dojo for training and AI4/AI5 for inference. AI5 offers 8× raw compute, roughly 9× memory and 5× memory bandwidth versus AI4, while the headline 40× applies specifically to AI4’s worst limitation: softmax, currently run in about 40 emulated steps. Musk said AI4-equipped cars will achieve self-driving safety at least 2-3× human and “maybe even 10×”; Version 14 is due over the next few months, and he said cars will feel “sentient by the end of the year.”
• The $17 billion spectrum acquisition discussed by the host requires phone hardware changes that will probably take about two years; new satellites are being built in parallel. Musk expects high-bandwidth direct-to-cell connectivity, including video, and envisions one Starlink account for home and mobile service while terrestrial carriers remain. Normal homes should work, but thick metal roofs would block the signal; acquiring a carrier is “not out of the question.”
• After one more launch of the Starlink Version 2 stack, Starship Version 3—with Raptor 3—will radically redesign the rocket. Barring major setbacks, Musk expects next year’s demonstration to catch both booster and ship, achieve full reuse and deliver more than 100 tons to a useful orbit. He corrected the host’s five-times comparison: Falcon Heavy carries about 40 tons with side-booster reuse, so the payload comparison is 2.5×. A major remaining challenge is a fully reusable orbital heat shield.
• Grok is using heavy inference and reasoning to inspect source material, classify claims as true, partially true, false or missing, and rewrite it with corrections and context. The host proposed publishing the result as “Grokipedia,” and Musk said he would discuss it with the team. Asked about the Memphis Colossus expansion, he said they could have a second one. His rough rule of thumb is that 10× more compute might double intelligence; he guesses AI could be smarter than any single human at anything as soon as next year and smarter than the sum of all humans around 2030.
• Musk’s civilizational hedge is a self-sustaining Mars, defined not by the first landing but by whether the settlement prospers after Earth resupply stops. With exponential tonnage growth at each two-year transfer window, he estimated 10-15 windows—roughly 25 years, or 30 years in his initial formulation—while arguing that low Western birth rates reflect a loss of optimism and purpose.

Deep dive

Not yet available upstream; scheduled sync will retry.