@elonmusk At 220k GPUs your real risk isn't speed, it's silent data corruption with no error flag. the handwritten C stack with no goldenreference check trains on bad gradients for hours before anyone catches it. Cheapest insurance you can buy.
@elonmusk Hardcode to 220k maxes peak when metric is (peak×uptime). You lose nodes nonstop at that scale. rigid topology gives back in restarts what it won in peak. Codegen the sched from a topology descriptor. Same speed, but you re-map around dead racks live and retarget Rubin for free.
@elonmusk This is what war actually looks like, and to die, up close and personal. Having the lucky fortune to live in the USA across th world don't negate the suffering and deaths that could end today.
Britain’s PM strutted out today, chest puffed, promising boots stomping and jets screaming, war-ready swagger on full display. Then, the gut punch: ‘Only if the US tags in.’ Translation? UK’s all hot air without America’s muscle. This isn’t leadership, it’s a lapdog whimpering for its master. No, we are not starting World War 3. We’ve been screaming this for years: Europe’s a paper tiger, and the US is still the only one with teeth. Share this if you’re done with the charade.