Replying to Avatar trosso19

Right which is why the paper that came out at the same time describes all their training methods. They basically handed the whole world a "here's what you can do if you want to be as advanced as openAI's o1-pro models". There is no moat anymore for proprietary AI models or services.
Avatar
bitcoinlimit 11mo ago

But it all boils down to the hardware. You need h100s or similar high performance gpus to train and deploy models at that scale, which is still a significant barrier for most companies regardless of their methodologies or frameworks. It’s not the lack of knowledge for the most part but the hardware

Reply to this note

Please Login to reply.

Discussion

Avatar
trosso19 11mo ago

True, and there are still lots of rumors going around about the actual training hardware for deepseek, but I still think open sourcing a SOTA model is a huge step forward. One of the things the community has been best at is reducing the compute cost for various models.