$NBIS just squeezed another 1.13% of training throughput out of 192 $NVDA B300 GPUs running DeepSeek V3 with a single software adjustment.
Obviously this might sound small but when you're operating thousands of GPUs even incremental efficiency gains matter and I think next phase
The fun part: DeepSeek Flash curated by Opus.
Opus writes no code, it sets the task stage by stage (cutout → pixels → scene → animation), checks images, sends fixes. clean cutout on the 2nd try, all per spec. 39 min. cost: $0.23 DeepSeek + ~$2.30 Opus curating (API prices)
DeepSeek V4.1 Flash. 18 min, $0.11.
Followed the rules, but the cutout was rough: the wing turned into a thin strip and the car looks like a sticker with a white glow. (before the rules, an earlier run spent an hour, failed the cutout and drew a box with SUPRA on the plate 😅)
You have reached the end of the archive
All of deepseek