DeepSeek r1
Created with DALL-E

DeepSeek r1

There are some misconceptions why DeepSeek r1 has had such an impact in the past few days. Here are some facts on the model in comparison to other models.

  • Censorship: all models are censored or finetuned unless you can host them locally and do some finetuning yourself
  • LLMs and reasoning models are not build to retrieve facts. You need additional tools such as RAG for that.
  • Inference and thus pricing of DeepSeek r1 is a lot (about 25x) cheaper than comparable models. This is achieved by being a lot more (energy)-efficient than comparable models
  • The weights are open source and can thus be hosted whereever you like.
  • You can finetune and adapt the models yourself if host the model independently.
  • The process of training has been publicly documented and you can be very sure that many labs are currently replicating and if successful verifying the process.
  • Some of the training processes are (at least in terms of previously publicly available information) innovative approaches that according to the researchers are a lot more efficient and successful.



Fayssal El Mofatiche, CAIA

Founder & CEO @ Flowistic| Business Angel | Senior FSI Professional | FinTech DevNight

1 个月

Good points! The critiques of the censorship of DeepSeek R1 are silly at best, as though other models are not censored and as though companies are not expected to abide by local laws everywhere in the world. But, even sillier because apparently these ppl do not understand what open source open weights means.

回复
Christian Hille

Founder and CEO of Caplign Wealth. Experienced Finance Professional and Investor

1 个月

Very good points and summary Christian Schuster

回复

要查看或添加评论,请登录

Christian Schuster的更多文章

  • KI im Asset Management

    KI im Asset Management

    Das Handelsblatt hat am Mittwoch einen Artikel zum Einsatz und der Anlage in KI ver?ffentlicht. Die Ergebnisse sind…

社区洞察

其他会员也浏览了