Ben Longmier
1.3K posts
Ben Longmier
@longmier
Sr Director, Satellite Engineering @Spacex
Bay Area Katılım Mart 2015
354 Takip Edilen15.7K Takipçiler
@Robotbeat There is dogma is every field that can be deleted with a few small tweaks to the basic assumptions.
English
This is actually true. *at the same temperature *, single-crystal gives you like 3-9x thermal fatigue life than polycrystal. But if you reduce temperature by 20-50°C, you can achieve the same lifetime. That’s like half a percentage point efficiency hit, 34.5% vs 35.0%.
lachlan@hyprturing
there is no way you actually need single crystal blades to make cost effective power production turbines. people love overcomplicating stuff
English
Ben Longmier retweetledi
Ben Longmier retweetledi
To clarify, the current terminals can support peak download speeds of 400-500 Mbps. The 220 Mbps number is referring to actual in-flight performance given that the satellite to user beams are shared amongst users. For larger commercial planes, we allocate more resources to ensure sufficient bandwidth.
Other providers quote peak speeds as real performance, which is only relevant if you have very few users. @starlink quotes real performance measured via automated network performance tests.
The latest aviation terminal supports two channels, so up to 1 Gbps peak speeds per terminal (the A380 design consists of 3 of these terminals). We will continue to increase peak bandwidth per terminal in future revisions.
English
What an amazing showcase of the excellence that is the @SpaceX team.
SpaceX@SpaceX
Three years since the first flight of Starship, the next generation is here. New ship. New booster. New engines. New pad and new test site. SpaceX engineers are working to solve one of the most difficult engineering challenges in history: developing a fully, rapidly reusable rocket
English
Radiator a really terrible name for the air cooled heat exchanger in a car.
English
Pretty cool.
I’m not sure what your goals are, but if you desire 1e14 #/CC plasma density, I have had much better results with simply a half twist antenna, and a capacitive matching network with a very short distance to the antenna.
The goal is to also launch the helicon wave into the downstream section of the cusped field beyond the permanent magnet.
PS: it is not published, but one can achieve 1e16 plasma density at a 1-2 T magnetic field. Kind of insane outside of the fusion world.
Godspeed!
English
Ah yeah, %GDP is a much better plot.
Holy smokes, yeah the railroads really were crazy. Opened up the Western states for business and living though.
Rockefeller was a ruthless mf, and then separate statement: his empire building and Standard Oil really lucked out on the timing of the railroad expansion.
English
As a percent of GDP, hyperscaling isn't that extreme. The railroads were goddamn crazy, though!
Fin Moorhouse@finmoorhouse
The hyperscalers have already outspent the most famous US megaprojects
English
@finmoorhouse Wow. The Manhattan Project was a lot of bang for the buck.
English
The hyperscalers have already outspent the most famous US megaprojects
English
SOMEONE BUILT A MAP THAT SHOWS EXACTLY WHERE EVERY POWER PLANT, TRANSMISSION LINE, SUBSTATION & DATA CENTER SITS ON THE US GRID
all on one interactive map. all free
you can see how the grid is laid out... where the datacenters cluster... which transmission corridors carry the load... where the high-capacity connection points are
opengridworks.com/power-plants
zoom into any region and the whole picture comes into focus why energy costs what it costs, why data centers go where they go, why some states are power exporters and others aren't
this is the kind of infrastructure visibility that used to require expensive industry reports
now it's one tab
English
@beffjezos Gradient ascent if you will…none of this AI gradient descent baloney, oh wait.
English
The purpose of life is to claw your way up the direction of the Kardashev gradient, while fighting off the parasitic Decels trying to bring down civilization
English
@lougrims @RocketLab Ha yeah, I know how it goes. It was a good promo.
The camera view flying through the plume was fun.
English
@longmier @RocketLab Thanks! I can only hope that we soon get into volumes where xenon (or even krypton!) are the limiting factor.
And that ppu may have been chosen to be extra shiny for the promo movie 😉.
English
Meet Gauss: Our new in-house designed and built electric propulsion system.
Electric propulsion satellite thrusters have historically proven extremely difficult to produce in high volumes, causing supply chain fragility for national security and commercial constellation operators alike. We’ve already established the Gauss production line, designed to produce more than 200 thrusters per year, ensuring we can supply thrusters on demand in large quantities.
English
A very smart friend asked if I’d consider helping look at AI. I’ve literally never read about it and know nothing. Anyone have some good/fun resources?
So far I’ve read The Road to Modern AI by @SchmidhuberAI
(Great context, and so many good references)
arxiv.org/pdf/2212.11279
Starting in on these videos next from @karpathy
youtube.com/playlist?list=…
I’m extremely dumb in all these topics, but it does strike me that the brute force method of fully dense LLMs being slammajammed through warehouses of GPUs that do matrix math layer by layer is a fools errand in terms of seconds of memory calls and Joules spent computing weights for mostly meaningless/empty nodes.
- 1-bit weights seem cute, but biology seems to play a few other clever tricks. There is more here.
- Mixture of experts seem cute, but basically a linear hack. Not meaningful enough.
- Batching is efficient, but only for multiple user requests and only in data centers.
There has to be a better way for a node in one layer to communicate to a node in a separate layer without waiting for the whole silly layer to finish its matrix math.
For sure evolving the software and hardware simultaneously is the only way.
@xai is doing this.
@GeminiApp is doing this.
Anyone else?
xAI@xai
Announcing Terafab: the next step towards becoming a galactic civilization x.com/i/broadcasts/1…
English
Average startup founder experience first thing in the morning while you're still booting up
David J Phillips@davj
Opens slack
English
This seems like a great idea. I am running it on my phone and it is super fast for such a small model!
Seems like this 1-bit model weight idea has been out since 2016.
arxiv.org/pdf/1511.00363
Just think of what else is lurking out in the literature, just around the corner !
PrismML@PrismML
Today, we are emerging from stealth and launching PrismML, an AI lab with Caltech origins that is centered on building the most concentrated form of intelligence. At PrismML, we believe that the next major leaps in AI will be driven by order-of-magnitude improvements in intelligence density, not just sheer parameter count. Our first proof point is the 1-bit Bonsai 8B, a 1-bit weight model that fits into 1.15 GBs of memory and delivers over 10x the intelligence density of its full-precision counterparts. It is 14x smaller, 8x faster, and 5x more energy efficient on edge hardware while remaining competitive with other models in its parameter-class. We are open-sourcing the model under Apache 2.0 license, along with Bonsai 4B and 1.7B models. When advanced models become small, fast, and efficient enough to run locally, the design space for AI changes immediately. We believe in a future of on-device agents, real-time robotics, offline intelligence and entirely new products that were previously impossible. We are excited to share our vision with you and keep working in the future to push the frontier of intelligence to the edge.
English