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

Musk says Optimus now consumes more of his attention than any other single project, and Tesla is finalizing a third-generation design intended for human-level hand dexterity, autonomous perception and mass production. Tesla had to design every actuator from scratch, and Musk estimates roughly $20,000–$25,000 production cost at one million robots annually.

Tesla’s next AI5 chip is designed around weaknesses discovered while jointly developing its driving software and hardware, producing far larger gains than raw computing power alone. Musk said AI5 has about 8× compute, 9× memory and 5× memory bandwidth versus AI4, with some workloads improving as much as 40×.

Musk expects existing AI4-equipped Teslas—not future hardware—to become at least two to three times safer than human drivers through software improvements alone. He said FSD version 14 will increase model parameters roughly tenfold and called it Tesla’s biggest software upgrade since version 12, with release expected within months.

SpaceX’s $17 billion spectrum purchase is aimed at turning Starlink into a direct competitor to terrestrial mobile service, but compatible phones will require new radio hardware. Musk expects supported handsets in roughly two years, eventually allowing one Starlink account to provide home internet and high-bandwidth mobile connectivity globally.

Musk expects SpaceX to demonstrate fully reusable Starship operations next year, including recovery of both stages while carrying more than 100 tons to useful orbit. The remaining fundamental obstacle is the heat shield: unlike the Shuttle’s labor-intensive tiles, Starship’s must survive repeated orbital reentry without individual inspection or refurbishment.

xAI is using inference compute to rewrite its training corpus before training future Grok models, checking source material for errors, omissions and misleading context rather than simply ingesting it unchanged. Musk also agreed to discuss publishing those rewritten materials as a “Grokpedia,” potentially turning an internal data-cleaning process into a public knowledge product.

Musk’s current AI forecast is extraordinarily aggressive: machines smarter than any individual human at any task could arrive next year, and exceed humanity’s combined intelligence around 2030. His rough scaling assumption is that a tenfold increase in computing power might double intelligence, so he expects capability growth to continue despite diminishing returns.

Musk now frames Mars success not as landing people there but as creating a civilization able to survive permanently after Earth resupply stops. He estimates that could be achieved in roughly 25 years, provided cargo delivered to Mars grows exponentially across about 10–15 biennial transfer windows.