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

1. Washington is unfixable; Musk sees AI and robots as the debt escape

• Back at Tesla Global Engineering Headquarters in Palo Alto after a Washington side quest—Musk said he has not been to D.C. since May—his lesson was categorical: “The government is basically unfixable.” He said interest payments exceed the U.S. Defense Department, or “War Department,” budget, and warned that if AI and robots do not solve the national debt, “we’re toast.”

• Optimus Version 3 is being finalized around three capabilities Musk says competitors still lack: essentially human manual dexterity, an AI mind able to comprehend and navigate reality, and very-high-volume production. He now spends “more of my mental cycles” on Optimus than on any other single undertaking.

• There is no off-the-shelf humanoid supply chain: Tesla has had to recreate it from scratch, including substantial vertical integration and the design of every motor, gearbox and piece of controlling electronics from physics first principles. With 26 actuators per arm and roughly 27-28 degrees of freedom in a human hand, Musk called the hand-plus-forearm most of the robot’s engineering difficulty—versatile enough to thread a needle, play an instrument or assemble a car.

• At one million units per year, Musk estimated production cost around $20,000, perhaps $25,000, depending partly on a $5,000-$6,000-plus AI chip; selling price would depend on demand. Specialized machines are easier, he said, 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.

2. Tesla’s silicon advantage is specific, while Starlink needs new phones

• Tesla has two chip programs: Dojo on the training side and AI4 on the inference side. AI4 is already shipping in vehicles; AI5 is being tightly co-designed with the software and is projected to provide 8× nominal compute, about 9× the memory and roughly 5× the memory bandwidth.

• The headline 40× gain is not a uniform improvement: Musk said it compares AI5 with AI4’s worst limitation, softmax, which currently takes around 40 emulated steps and will run in only a few native steps. AI5 will also dynamically handle mixed-precision models.

• Musk said he is confident that AI4-equipped cars will achieve self-driving safety at least 2-3× human and “maybe even 10×.” Version 14, coming over the next few months, increases parameter count by an order of magnitude, uses substantial reinforcement learning and addresses overly lossy reality-compression steps. He said the car will feel “sentient by the end of the year.”

• Direct-to-cell Starlink is a roughly two-year phone-hardware project because current devices do not support the acquired frequencies; new satellites and handset support are being developed in parallel. The promised result is high-bandwidth connectivity, including video, nearly anywhere on a phone—but not inside buildings with thick metal roofs. Musk said terrestrial carriers will remain, while a unified Starlink account for home and mobile service, and possibly acquiring a carrier, remain options.

3. Starship’s reusable heat shield is a major remaining barrier

• One more launch of the Starlink Version 2 stack remains before Starship Version 3 introduces Raptor 3 and changes “pretty much everything” on the rocket. Musk allowed for initial “teething pains,” but, absent major setbacks, expects next year to demonstrate both catches and more than 100 tons to orbit with full reuse.

• When the host called that five times the best rocket, Musk corrected the comparison: Falcon Heavy manages about 40 tons with reusable side boosters, so the payload comparison is 2.5×. Starship’s additional distinction is that the vehicle would be fully reusable.

• The host asked how SpaceX recovered so quickly from the recent explosion despite environmental, FAA and other reviews. Musk acknowledged “a lot of questions and reviews” but credited the team with testing the next ship and booster, getting them to the pad and flying them rapidly. SpaceX has pursued full reuse since 2002, a 23-year engineering journey.

• A major unsolved piece is a fully reusable orbital heat shield—something Musk says has never existed. Unlike the shuttle’s nine-month refurbishment cycle, Starship’s tens of thousands of light tiles must withstand extreme heat without transferring it into the primary structure, avoid dissolving in rain and work again without laborious tile-by-tile inspection.

4. Grok is turning inference into cleaner training data

• Grok is applying heavy inference and reasoning to the corpus of human knowledge, assessing each piece as true, partially true, false or missing. It then rewrites the material to remove falsehoods, correct half-truths and add missing context across Wikipedia, books, PDFs and websites.

• The host proposed publishing those revisions as “Grokipedia,” describing Wikipedia as hyper-partisan and full of activists. Musk’s answer was provisional: “I’ll talk to the team about that.”

• Asked about scaling the Memphis Colossus supercluster, Musk said they could have a second one. His rough rule of thumb is that 10× more compute might double intelligence—still enough, in his example, to move from an IQ of 100 to 200. He expects scaling toward compute powered by the Sun and ultimately the galaxy, invoking roughly Kardashev Type II- and Type III-scale systems.

• Musk guesses AI might be smarter than any single human at anything as soon as next year and smarter than the sum of all humans around 2030. He linked a plateau and eventual decline in human intelligence to population declines and low growth rates.

5. Mars is planetary redundancy, but optimism comes first

• The host bundled below-replacement birth rates, open borders, weak assimilation, crime and selective media attention into a theory of Western self-destruction. Musk agreed that these data points seem to suggest the West is “suicidal” and said he was “very worried,” contrasting America’s general optimism with the rarity of an optimistic European.

• His causal link runs through purpose: “Having a child is an act of optimism about the future.” Musk argued that removing religion can leave a vacuum filled by destructive de facto religions, including the “woke mind virus.” His own proposed philosophy is curiosity: expanding consciousness, exploring the stars and discovering which questions the universe invites.

• Space exploration embodies that optimism. Musk said going to the Moon should serve the establishment of a lunar research base, followed by a self-sustaining Martian city before civilization’s arc descends. The opportunity to make life multiplanetary, he stressed, exists now for the first time in Earth’s 4.5-billion-year history.

• Asked how he would govern Mars, Musk said governance is secondary to self-sustainability. The load-bearing test is whether Mars still prospers if resupply ships stop, requiring all the ingredients of civilization—not merely a chip fab, but the ability to build.

• With exponential tonnage growth across 10-15 transfer windows, each occurring every two years, Musk said Mars should be able to become self-sustaining in roughly 25 years, after initially saying it could be done in 30.