W4lter

Maybe human not sure

[W4lter @itzW4lter]

@_Tom_Henderson_ It's mostly proof to stop designing games for console and controller.

[W4lter @itzW4lter]

@farzyness @FaisalIqbalFSD @grok Why do you discriminate between an embodied model and the digital ones you obsess over?

[W4lter @itzW4lter]

@spencerpratt Woke is the unionization of weak people. Weak is the antithesis of sexy.

[W4lter @itzW4lter]

@WR4NYGov @CernBasher I view cheat meals not as calories, but miles of running and then i don't eat them.

[W4lter @itzW4lter]

@BoochaDon @hamids "Elon is a liar but I'm not, now give me likes" Hamid

It's easy to find wrong predictions from someone who as prolific as Elon. Is Hamid's goal that Elon never posts ever again? I'd rather hear what Elon has to say and interpret it myself.

[W4lter @itzW4lter]

@LimitingThe Also Elon name dropped Micron in the tesla earnings call last January and people are blaming him for missing the memory trade lol

[W4lter @itzW4lter]

@farzyness Starlink is the only reason Ukraine wasn't blitzed

[W4lter @itzW4lter]

@hamids This expands much farther on what i stated. The spacex data centers are for inference, not training, and each will be either self contained or just working with its immediate neighbors. @hamids

https://t.co/Ryx7jJbVBK

[Aaron Burnett @aaronburnett]

orbital data centers have one glaring constraint few are discussing publicly: inference network coherence. We tackled it this week in our analysis "The Orbital Inference Containment Tax"

tbf we mostly hand-waved coherence in our early ODC models (other things to learn first), but went deep after encouragement from someone who's worked on this problem for years.

Fair warning: this is the easily the densest analysis we've ever done, because we're modeling at the edge of both AI and spacecraft.

If you have the time to dive in, it's really f*ing interesting and insightful for what I expect will happen with orbital architecture over the next few years.

If this isn't your day job, share the link with your AI of choice and have it get you up to speed.
https://t.co/GFX3cS5IRN

core learnings this week:

- Frontier LLMs are mixture-of-experts models. Ground operators spread hundreds of experts across big GPU pools down to about 1 expert per GPU. NVIDIA's benchmark shows that spreading is worth up to 1.8x more tokens/second per GPU. and tokens/second is $/second in inference.

- A self contained satellite today can't spread that wide. GPUs computing one answer must sync every ~5 microseconds, and light only covers a ~750m round trip in that window. consequently a coherent GPU team ends at the edge.

- Your options to address this are formation-fly sats ~150m apart (Google's Suncatcher bet, orders of magnitude closer than Starlinks fly today) or eat the penalty of containing the model inside one sat.

- An AI1 sat's power spec ≈ one rack of GB300s (NVL72) 72 GPUs. if you force a 256 expert model into that box then each GPU has to juggle 3.56 experts vs ~1 on the ground.

- We conservatively stacked every assumption against orbit and worst case, a sat is 44% less efficient at processing tokens than on the ground.

That's what we call 'the containment tax' 1.8 sats worth of GPUs to do the work of 1

note,we used max conservative assumptions as I believe it's important to stress test the orbital compute thesis when reasonable.

That said the tax decays, when modeled the levers become pretty clear...

1) SpaceXAI's C-rewrite captures up to half the tax consistent with the token/second uplift we modeled last week.

2) Co-design the model to fit the sat's 72 GPUs. The largest single lever, and it's a training decision, not a hardware program.

and

3) add more GPUs per sat via power density. The next node class (assuming a 200kw-240kw class sat) cuts the baseline to 1.44x before software improvements even runs

1 and 3 are already underway at SpaceX.

2 is the cheap, logical next step given their compute advantage and ~3.5-week model release cadence

nothing stops them offering co-design as a service to neocloud customers.

what we expect? Grok and Composer are the first two models co-designed for orbital inference sats and any external customer co-design comes later.

Scoping note: the containment tax hits the revenue and payback side of the equation, not the cost side, so it's not in our orbital-vs-terrestrial cost work yet.

These learnings will roll into future iterations of AI compute and the SpaceX Gigamodels.

Thanks for reading and happy modeling!

[W4lter @itzW4lter]

@hamids The satellites don't need to communicate with one another if they are running inference of models that fit on a single satellite. Also you overlooked the self sufficient power from space solar.

[W4lter @itzW4lter]

@SpicyTech_ca Your phone becomes a pokeball for which ever vehicle you need: car, truck, semi...

[W4lter @itzW4lter]

@alojoh Elon's job isn't predicting, that's yours, his job is to build, it's much more impressive.

[W4lter @itzW4lter]

@alojoh @The_Guy_Space The main point is starship launches and lands in the same spot so there's no time wasted moving a rocket across land and sea, we need to see a cable system that can launch, not just catch

[W4lter @itzW4lter]

@alojoh @The_Guy_Space I'm not sure if that hoist system can duplicate all the umbilical functions a launch tower does, then if a tower has to be added inside the cables, it will interfere with the movement of the cables unless they're above or to the side; one would need to be built to find out

[W4lter @itzW4lter]

@alojoh @The_Guy_Space I believe the arms allow for it to be restacked and launched immediately for re-usability

[W4lter @itzW4lter]

@thejefflutz @farzyness when he sees the word steered:

[W4lter @itzW4lter]

@PalmerLuckey Fathers instill the fear of shame and honor in your name, so it's rooted in the down trend of masculinity and high testosterone like most other modern problems.

[W4lter @itzW4lter]

@farzyness @SawyerMerritt Neuralink

[W4lter @itzW4lter]

@TeslaBoomerMama Schwab
100%
10/10

[W4lter @itzW4lter]

@FutureAZA @LimitingThe He's a professional retard