Today we are going to talk about the dominance of AI companies like chat GPT meta Google are being challenged by Deepseek, yes deep seek is a new entry to the AI world so deep seek is a Chinese research lab which has come up with a new llm model named as R1 so it has amazed everyone with its performance and the low cost so let’s get into the details so deep seek is basically a very cost effective and quick fast we can say solution in the AI World which has come up in the recent two months so which was delivered by Chad GPT Googles and meta it took around years and billions of billions of investment dollars in terms of cost so it was very very cost effective and it has only took less than $6 million to deliver this llm model so competitive performance later on we will see that how it is very very good at performing the same task which chat gpt’s open AI or Google’s Gemini model is has performed so it is very good and very performance- wise very effective and it can run on the local systems so no need of GPUs the high performance processing chips by Nvidia so how it was started basically recently us has put a ban on Nvidia chips which were exported to China so this was just to maintain the dominance of AI companies but it has backfired how it has backfired in a way like they have managed to compute the same performance with the lowered version of Nvidia chips as we all say necessity is the mother of invention so once that ban was put on the latest Nvidia chips of importing it to China so they have managed to execute the same processing capabilities from the existing chips which was quite older but they managed to somehow extract the similar level of performance from those chips so H 800s is the chips they have used in deep 6 R1 model so h00 S series is the latest one in the market which was kind of not available to China due to this export ban but somehow they manage the older one h800 s chips to utilize the compute and performance wise the same area they have utilized so but how how they could use it so the basic thing which they reused the feature was reinforcement learning although reinforcement learning is a separate topic all together we we have a separate post available for that but basically at high level reinforcement learning is just a type of learning where we want agent or system to learn from the actions for each action it does we either give a award reward to it or a punishment to it it is just like a newborn child starts to learn walking how it works like how it learns the same way reinforcement learning is also based on like when a child starts to walk so it learns like these are the steps we I need to take otherwise I may get hurt so how this is how human starts to learn things based on each of the actions either that action would result in a reward or a punishment so that is how this reinforcement learning works it is the same thing they have utilized which allowed them to have the same kind of compute and the processing facilities available at a much lower cost because it does not need any supervised learning and how good deepcar is let’s let’s get into that so if we see aim these all are mathematical operation this blue bars are deep seeks model and the gray one are the open AI from Chad GPT if you see this graph all the mathematical operations coding wise code forces data is also at the similar or at the higher side from d and this is GP QA for QA Q and A answers so all these data if you see mathematics MML so all these bench marks if we see these are all at par or at higher from the existing open AI models so how good we can see this deep C car1 has performed in terms of performance with much lower cost with such an older GPU chips so this is amazing this is very amazing and what are the final thoughts so final thoughts are don’t always reinvent the wheel because what they they have done is they have came to the basics and they have reutilized the same basic concepts and delivered a much better performance.
Here are more details: https://www.youtube.com/watch?v=CwT3QfH66H8
Artificial Intelligence ML & Data Science Training Institute @ https://g.co/kgs/VFLQKy2