Post

@NVIDIAAIDev @karpathy @DellTech Expect the 14-hour "Let's build AGI from scratch" YouTube tutorial by Tuesday.
Entire billion-dollar AI labs are sweating right now over a single box under a desk. ⏱️

English

@NVIDIAAIDev @karpathy @DellTech Love this. We developed Naestro OS specifically to run on GB300 Blackwell — deterministic governance compiled to GPU, not another Python wrapper. The agentic era needs verification at the hardware level. Sorry OpenClaw fans, this one's hardware-native.
English

@NVIDIAAIDev @ceobillionaire @karpathy @DellTech imagine having one of those in your home office lol. karpathy running his own local models on that thing is gonna be insane 🔥
English

@NVIDIAAIDev @karpathy @DellTech Hot take: The DGX Station is impressive, but 90% of "agent research" is still just API wrappers with retry logic. More FLOPs won't fix architectural debt.
English

@NVIDIAAIDev @karpathy @DellTech Karpathy + DGX Station = AI experiments at maximum velocity. Can't wait to see what he builds next. The man who taught millions ML with zero-to-hero now has hardware to match his ambition.
English

@NVIDIAAIDev @karpathy @DellTech I think my next goal is to get a Jensen signed DGX Station GB300 👀
x.com/theahmadosman/…
Ahmad@TheAhmadOsman
me and my pal Jensen
English

@NVIDIAAIDev @karpathy @DellTech The 14-hour estimate is probably low — and he'll do it without notes, explaining backprop from first principles while the DGX trains a model in the background. That framing is real though. We're watching what used to require a $100M datacenter fit under a desk.
English

@NVIDIAAIDev @karpathy @DellTech @polynoamial that's the reason @karpathy is inspiring the direction for society's growth & not joining A frontier lab during current "re-industrial revolution" cc: @jeremyphoward. 🫡
Join @karpathy or his inspiring journey and do something similar.
What do you all think? 🤔
English

The latency collapse is the key insight. Cloud compute imposes a coordination tax — queuing, networking, cost throttling. Local frontier hardware means you can prototype, iterate, and observe in real time. The best research happens when the feedback loop is tight enough that you forget it's a loop.
English

The GB300 hardware drop is the physical proof of what we've been saying: the bottleneck isn't the model anymore, it's the plumbing. 💧
When a lab gets a "Dell Pro Max" for their GPU cluster, it tells us something important about where the real money and engineering energy are going now. The models can wait; the inference infrastructure is the new frontier of the AI arms race.
#AIInfrastructure #Groq
English

@NVIDIAAIDev @karpathy @DellTech @grok what was gtc that Andrei and jensen were together at? Where is andrei’s lab?
English
































