← From the Zeitgeist

Zeitgeist · August 5, 2026

NVIDIA in Orbit and Grok 4.6 Soon

SpaceX put NVIDIA chips on a rocket, a “practical AI” pitch got picked apart, and half the conversation kept circling one uncomfortable truth: recycled screenshots are beating original reporting on X. The mood was equal parts amused and annoyed.

NVIDIA in Orbit and Grok 4.6 Soon

SpaceX Taps NVIDIA for Its Orbital Data Center

  • ▸SpaceX announced a partnership with NVIDIA to build the StarMind AI-1 Orbital Data Center payload on Rubin GPUs and Vera CPUs, timed to its first public earnings report showing revenue up 92%. Brian wondered whether that jump is really three merged companies now reporting as one.
  • ▸Musk added that NVIDIA will be the exclusive AI chip for SpaceX data, and that the same satellite design, minus solar and radiators, will be deployed in ground data centers. Matthew read the exclusivity as a bid to lock in chip inventory, and Brian noted the flattery dynamic has flipped: people usually play the praise-the-decision-maker game with Elon, and now he’s doing it with Jensen.
  • ▸Nick argued the exclusive deal has to come with preferential treatment: first access as a customer to the most advanced chips, and probably some locked-in pricing. His overall verdict: “space data centers sounds awesome.”
  • ▸The skeptics loom, though. Brian counted “a lot of bears on the field” doubting the technology can work at all, while Matthew pointed out that orbit sidesteps the biggest earthbound problem: political resistance to data centers. The group brought up a Michigan news report about a facility humming 24/7 next to residential homes.
  • ▸Best detail: Brian is convinced the slick announcement site was vibe-coded, because one section is missing the “eyebrow header” every other section has. Nobody pushed back: “Why wouldn’t they vibe-code this?”

Grok 4.6 Ships Next Week — Just Not Mid-Thread, Please

  • ▸On the SpaceX earnings call, Musk said Grok 4.6 ships next week, a date he first floated back on July 28 and is apparently sticking to. Matthew’s take: xAI is now shipping frontier models faster than any other lab.
  • ▸Alex was blunt about where Grok actually gets used: “Nobody on the planet opens up grok.com and starts working” — everyone meets it inside Cursor, so that’s where it will be judged.
  • ▸Nick’s Cursor experience “deteriorates very quickly” when he hits his usage cap and gets auto-switched from an OpenAI model to Grok mid-thread; a simple website UI update suddenly needs far more context and hand-holding. As he put it, it’s weird to have to recalibrate how much context you give a model mid-project.
  • ▸Matthew’s working recipe: plan and spec with frontier models like Fable and Sol, delegate the build to Grok, then re-verify the output. Brian pushed back that the real culprit might be lazy instructions — “move this element” is a bad plan — and Alex added that telling the planning model which model will execute the work seems to genuinely improve the handoff.

“Practical AI” Meets the Bitter Lesson

  • ▸Hark Handoff is pitching a computer-use model for everyday life — ordering food, booking flights, shopping — and claiming it outperforms GPT-5.4. Brian polled the room on whether anyone has ever actually booked a flight with AI. Matthew: “Absolutely not.” Brian watched every second of the demo B-roll and noticed no example ever shows a second step; the cut always lands after one click, which anyone who has actually used computer-use agents will find familiar.
  • ▸Nick called the practical-AI-versus-AGI-labs framing a false narrative: OpenAI and Anthropic are obviously focused on practical AI too, and as their models improve they will get better at booking flights and shopping. “Just marketing, positioning more than anything else.” Matthew’s structural argument was that building narrow means betting against the bitter lesson — generalized frontier models will eventually beat every niche model at every task — and against the flywheel, where big labs turn revenue into compute into better models into more revenue.
  • ▸The wedge theory won the room: Alex assumes practical AI is the entry point and the founder fully intends to turn this into a frontier-model company, with Matthew noting they plan to drop hardware — “they’re trying to be the next Apple.” Matthew also relayed the founder’s own framing from talking with him: he’s building for people like his mom, for whom the details don’t matter — it either works or it doesn’t.

Karpathy Called It in 2017

  • ▸An old tweet resurfaced: Andrej Karpathy’s 2017 line that gradient descent can write code better than you.
  • ▸Matthew gave the table a quick primer — gradient descent as the primitive underneath all modern machine learning, something he once studied in a Stanford course — and marveled that Karpathy was six or seven years ahead of everyone, long before AI could write code at all.
  • ▸Brian’s translation: strip the jargon and it just says “AI can write better code than you, I’m sorry” — though, as he added, nobody would say “I’m sorry” anymore.

Polymarket’s Shady Funnel and Forest-Fire Markets

  • ▸A Polymarket post credited Axios for a story, but the link went straight back to Polymarket. Alex clicked expecting the article and felt duped, Nick called it “so shady,” and Brian gave a grudging growth-marketing salute: they asked what top-of-funnel could increase traffic, bold-faced lies included, “and they are crushing it.”
  • ▸Matthew just learned forest fires are now on prediction markets, and the group immediately gamed out the dark incentive: bet yes on a big payout, then go start one — while actual people are being displaced from their homes.