Letโs not fall for the misinformation about ๐๐๐๐ฉ๐๐๐๐ค ๐๐! Letโs set the record straight:โฃ โฃ
1. Training didnโt just cost ~$๐๐ ๐งThe $๐.๐๐ figure covers base model compute onlyโno ablations, smaller runs, or data generation included.โฃ
2. Itโs not a side project ๐โโ๏ธDeepSeek is owned by ๐๐ข๐ ๐ก-๐
๐ฅ๐ฒ๐๐ซ, a Chinese hedge fund managing $๐๐+ with a team of math, physics, and informatics Olympians.โฃ
3. They donโt have โa few GPUsโโthey have ๐๐,๐๐๐ ๐โโ๏ธ
4. The real ๐๐๐๐ฉ๐๐๐๐ค ๐๐ is a ๐๐๐๐ ๐๐จ๐ model requiring ๐๐๐ฑ ๐๐๐๐ ๐๐๐๐ฌ (๐๐๐๐๐ฌ) to run ๐ซ
5. The smaller โdistilledโ versions (e.g., 1.5B) are not R1; ๐คญ theyโre just fine-tuned ๐๐ฐ๐๐ง/๐๐ฅ๐๐ฆ๐ models. Yes, they can run locally, but theyโre nowhere near R1-level performance.โฃ
6. Hosted versions on their website may use your data to train new models ๐คฏ(check the ToS).โฃ โฃ
7. The exciting part? DeepSeek just announced ๐๐๐ง๐ฎ๐ฌ-๐๐ซ๐จ-๐๐, an open-source model that generates images and outperforms OpenAIโs ๐๐๐๐-๐ ๐ and ๐๐ญ๐๐๐ฅ๐ ๐๐ข๐๐๐ฎ๐ฌ๐ข๐จ๐ง across benchmarks ๐ฅบThe AI competition is heating up!โฃ โฃ
The good news? has been contributing to open-source and science for 2+ years ๐ซก Hugging Face is even building a fully open pipeline. The future looks bright for everyone!
