Hey buddies,
Itās been unattainable to flee the avalanche of headlines about DeepSeek, however I seen thereās an excessive amount of noise and deceptive info (like DeepSeek R1 solely costing $5.6 million to coach). So I spent final night time studying all these totally different sources to sew collectively a transparent rationalization of why that is essential information.
To make it extra digestible, I made this visible explainer due to all of the optimistic suggestions I acquired from you final week.
Watch on: TikTok | Instagram | YouTube
DeepSeek shocked the world for 4 causes:
-
Coaching Value: Totally different retailers and content material creators reported that DeepSeek R1 price $5.6 million to coach. That is incorrect. The bottom mannequin, DeepSeek V3, price $5.6 million to coach in its last run which excludes the totally different experiments they did main as much as the ultimate consequence. The price to coach R1 was seemingly extra and we donāt understand how a lot. We donāt understand how a lot it price OpenAI to coach its o1 mannequin both. The one estimate now we have is that GPT-4 price greater than $100 million to coach.
-
Coaching Methodology: DeepSeek used Reinforcement Studying (letting the mannequin be taught and enhance primarily based on rewards) versus supervised fine-tuning (feeding particular examples it ought to be taught from) like OpenAI did with o1. Right hereās a visible explainer of various machine studying strategies in case you want a refresher.
-
Utilization Value: The largest shock mayāve been that DeepSeek R1 is open-source and free for customers to make use of. Itās additionally 97% cheaper for builders and companies who wish to use their API inside their very own functions.
-
{Hardware}: Because of the USās export controls, NVIDIA can solely promote H800 GPUs to China that are a modified (and weaker) model of the H100s that every one American corporations use. Coaching a reasoning mannequin on much less environment friendly {hardware} places a giant crimson query mark on the necessity for large {hardware} investments and despatched NVIDIAās inventory down greater than 15%.
All these elements mixed have put into query Americaās perceived dominance within the AI house. China was capable of match OpenAIās o1 efficiency throughout benchmarks with worse {hardware}, much less information, whereas making it open-source.
I hope this helps make clear the information of the story and equips you with the best info.
Credit & Acknowledgments:
One of the best assets that I learn by far have been:
Share
If you realize somebody who would take pleasure in such a content material, inform them to subscribe to Yr 2049 at this hyperlink (year2049.substack.com) or share this submit with them.
