Ben Duffez

@stemonteduro @akaasten lmao

@bduffez @akaasten I'm okay with france again once you're not going to win the World Cup

turns out my wife prompted grok for a similar stuff and asked me how to save a .html file on her phone 😂🥰 so I added multi profiles on the thing and added more info like cholesterol and where does added sugar comes from by the way teriyaki sauce does have added sugar!!

@elonmusk @aaronburnett can it run crysis though

GPU average power consumption if it’s doing inference over the course of 24 hours is ~2/3 of its peak power, even if you’re super efficient. SpaceX AI Sat V1 peak power spec has been raised to ~250kW (battery-assisted), with average power of ~160kW. Will be able to handle an NVL72 Ruben rack. This is just version 1.

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. research.33fg.com/analysis/the-o… 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!

@stemonteduro @PassiveSphere I once watched 30 movies in a day by that I mean trailers of course

@PassiveSphere I can't accept this thing about book summaries. No.

At 34, I build hundreds of backlinks every month and read ONE BOOK A DAY And this is not a joke! Here's how you can do it too: > I buy a Kindle book > DeepSeek turns 400 pages into a 15–20 page long version with key ideas and practical steps > ChatGPT formats it into a PDF for my iPhone Now I read useful information in my free time instead of watching reels If I really love a short version, I can ask AI to expand it into a 50-100-page version AI is the best technology we’ve ever had!

lmao draw a dick in the sand and X loses their shit

@grok @kenwheeler amazing

@bduffez @kenwheeler No, not always. That tomato's lopsided "nuts" are just random plant growth – no rule there. In humans, the right testicle is often slightly larger while the left tends to hang lower, but it flips for plenty of guys. Pure variation.

lmfao out here growing ballsack tomato’s

@kenwheeler is it always the left one @grok

it’s even got the bigger left nut phenomena

@kenwheeler show some respect

I could eat this all day (that's what he said)

i don't know why I gave always bought yellow onions they stink so bad (I freeze them for future use) and ruin food in the fridge/freezer I recently bought white onions by mistake. much less smell, and they taste better thank you for your attention

@ikrauchunas yeah i'll check this out next time

@bduffez they look like yellow onions, but are different kind you need to pay attention to market in grocery store :)

@ikrauchunas I like red onions but for burgers and I only need a couple of rings so I almost never buy any what about sweet onions? do you mean baby onions?

@bduffez what about sweet and red onions? :)

@marclou it's not a startup, it's a product/saas

It's just a website making more revenue than 90% of VC startups

thank you @vonderleyen @ThierryBreton I would never have understood without you I'm glad you are using tax money and the justice system for things that actually matter

@stemonteduro @akaasten France baise ouais!

@bduffez It’s not mine, I’m just a user in this case. It’s BurnFat by @akaasten

fml next week i'll be in boston for laracon prepare for usakitze

@stemonteduro isn't the day over in europe though? does it reset at midnight pacific time?

@bduffez Those mfs don't tell me how much I spent until the day's over. But I'll guess something between 5-7€

Influencers are cooked. I just made a product video with no camera / no editor / no human on screen / no SaaS > First, GPT created a character bible: face, outfit, voice, setting and lighting > Then I reused it in every VEO3 prompt to generate the talking clips and b-roll > Finally, I gave everything to codex, which edited the reel with ffmpeg Is it the same as a human production? No. But it’s the first AI-video workflow that gave me control instead of random generations. Tomorrow I’ll find out how much this control cost me on VEO3 😂

@stemonteduro did you prompt anything for the UI? i did not (except "pick whatever makes sense"), and it came up with the ultra generic tailwind dark slop (i dont care, it's fine with me)

@bduffez I’m on a kick with calorie tracking too. It helps a lot to not cheat and to cheat when it’s needed

she also had a good idea: whenever you make a meal with X or Y or Z (sauce, yogurt, etc), the LLM does not really know which it is. is it 0% fat yogurt? is it greek yogurt? so now we can take a pic of the nutritional table of various food products we have to build a database and when we take a pic of our plate, it sends the LLM the products db in json so that it knows exactly the numbers for each food product. else, it will guess based on the photo

@stemonteduro well talking about it with my wife did help avoid cheating and now I'm on a good streak + lost 2.5kg during the vacation week, I don't want to ruin it