6 ms·
> We match DeepSeek V4 Pro Base using ~50x fewer FLOPs – that’s around half of GPT3’s pretraining compute, or ~$0.5M on GB200. If this holds up that's a really
by simonw 6d ago
> We match DeepSeek V4 Pro Base using ~50x fewer FLOPs – that’s around half of GPT3’s pretraining compute, or ~$0.5M on GB200.
If this holds up that's a really big deal.
- wayfwdmachine 6d agoHuge if true. As it were.