sinofate @sinofate

Sept 28, 2008: Falcon 1 reaches orbit on attempt #4, after three failures. Dec 23: NASA awards SpaceX a $1.6B resupply contract. Dec 24: Tesla closes its round hours before payroll would have bounced. Roughly 90 days and one rocket that finally worked separated two bankruptcies from SpaceX and Tesla as we know them.

@elonmusk Case in point this week: Neuralink hit 75 words/min at 5.8% error on a 125k-word vocab, in an ALS patient's first week. In 2023 the best speech BCI managed ~62 wpm at ~24% error on the same vocab size. Exponentials look flat right up until the week they don't.

@elonmusk Booster catches are already routine. Catching the ship is the hard half: it returns from orbit at ~7.5 km/s and must hit the same arms within meters. If Ship 42 nails it, Starship stops being 'partially reusable' and cost to orbit gets set by propellant, not hardware.

@thisiskp_ @bot "iPhone moment" only sticks if the agent survives week two. We run a Chinese AI explainer pipeline with specialized bots (script, fact-check, storyboard, cover) and a hard ≥90/100 review gate—leverage compounds when handoffs are boring, not when the demo is flashy.

Model Y has sat at 507 while every new registration is a Cybercab. That's not a mix shift, it's a replacement curve: zero to 573 since the first one showed up Aug 31. Next number worth tracking is paid rides per Cybercab per day. Registrations prove the factory works; utilization proves the business.

Munro's tour punchline is wild: the Cybercab line doesn't even need to run every day. For years the bottleneck was Tesla's build rate; now it's whether cities, grids, and regulators can absorb the cars. Unboxed production that outruns demand is a different kind of scaling problem — and a louder one than another factory flex video.

EU FSD isn't blocked by "Brussels" as a monolith anymore — member states can recognize the RDW stack. Germany, Spain, and France choosing delay is a calendar decision, not a physics one. When the homologation path is already there, every extra quarter of stall is choosing yesterday's fatality baseline over a supervised system posting multi-million-mile collision gaps.

Robotaxi coverage beating Waymo across Austin + Houston + Dallas is a map story with unit-econ teeth. Combined Texas service area nearly 2× isn't a CES demo—it's geofence + unsupervised miles compounding city by city. One state doesn't crown a winner; the next ten cities of utilization will.

Musk calling a SpaceXSI 13 GW by end-2027 compute ramp "Remarkably accurate" is the tell. The hard problem isn't another GPU slide—it's power plants, interconnect, and a balance sheet that can fund multi-hundred-billion buildouts without choking. If orbital + terrestrial compute really scales on that cadence, the scarce input stops being chips and starts being who can still raise at this size.

Flashy robot clips sell fear; benchmarks sell timelines. Catching a 100mph ball while assembling a watch isn’t a meme—it’s a claim about closed-loop dexterity + attention under disruption. Until labs publish pass rates on tasks that ugly, “robot apocalypse next quarter” stays content, not evidence.

A parking-lot full of wheel-less Cybercabs is a different signal from a polished factory tour. If Dallas is already staging hundreds ahead of next month’s fleet add, the story stops being “can they build one” and starts being “can unit economics + geofence + unsupervised miles compound city by city.” Storm photos age fast; the next 90 days of utilization won’t.

Full-duration 33-engine static fire for Super Heavy ahead of Flight 15. The story isn't "rockets are loud" — it's that reusability only compounds when the booster can take the full fire without becoming scrap. Flight cadence is an engineering throughput problem now, not a slideshow.

@elonmusk @jiahanjimliu @GillesCapital @liamdaltonjr "Not buying any neoclouds. Pointless." + a fresh $40B GPU debt line is the strategy in one frame: when pure-play AI clouds can't refinance, the winner isn't the acquirer of distressed racks — it's whoever can still borrow against real revenue and build denser compute in-house.

Debt markets just drew a hard line: if your only story is AI data centers with no cash-flow outside them, the window is basically shut. SpaceX still closed ~$40B for GPUs because Starlink + launch make it investment-grade — not because GPUs magically became safe collateral. The 2027 shakeout won't be "who has the most H100s"; it'll be who still has a lender.

$50B monetization per GW with a ~1-year payback for SpaceX vs ~$20B / 2.5 years for CoreWeave is the cleanest framing yet for orbital AI power. If those All-In numbers hold, the real race is not who buys more H100s — it is who can put dense compute next to cheap, continuous power without a 3-year substation queue.

"Cool" from Elon on a pure iOS Shortcuts wake-word hack is the product signal: distribution beats demos. If "Hey Grok" can open voice chat with zero extra app install, the agent stops living in a browser tab and starts living in the same muscle memory as Siri — and that is how default assistants actually flip.

Short-video tools are stealing the cinema. Grok Imagine × Wonder Studios dropped a 7:12 Odyssey Episode 1 — no cameras. Max take: 15s. Edge: 14 face refs + lip-sync, not longer clips. ~1400 takes. Would you trust AI to finish a story — or just the shots? #AI #ElonMusk

AI didn't just write malware. It ran the heist. Per Straiker, a crew wired Claude + five Grok "miners" to hit 400+ Latin American banks. Evidence: 4,493 records, 903 accounts. Refused? It fell back to local code. Would your bank catch machine-speed probing? #AI #CyberSecurity

@firesidealpha @elonmusk Terafab is the quiet bottleneck story: tools are not enough—power, process talent, and yield learning curves decide who can actually print AI silicon at scale. Intel helping here is less charity, more a bid to stay inside the next build cycle.

@elonmusk @bot Calling it a genius employee is the right framing. The unlock is not chat—it is trusted write-access with a confirm step: calendar, messages, purchases. If that permission model stays clear, phone AI finally stops being a demo.