Akansha Yadav breaks down DeepSeek R1, a Chinese open-source AI model that matches ChatGPT-4's abilities for a fraction of the cost by using reinforcement learning instead of training from scratch. The model can reevaluate its own answers, and you can run it locally through Ollama.
We all know Deepseek is in hype. We all know Deepseek is trending, but what is Deepseek? So basically, Deepseek is the AI company which makes powerful LLM models, and they are based in China. They recently have created Deepseek R1 model. Deepseek R1 model is an open-source model which is accessible to the developers, researchers. Anyone can reuse it, anyone can modify the code, anyone can fine-tune the code as well. They have made it open source, still they are making money. Yes, We all know that because this is a company. They introduce this one moment in this which is aha moment. Aha moment says that your model is going to think like you think, like humans. So let me clarify this for you. So basically, just like we make decisions, we try to reevaluate our decisions, we try to make reevaluation on the basis of what we have been experienced or what we think that can be good for this situation, right? Same thing your model is going to do. Your model is going to rethink on the questions that you are going to provide to, and it will reevaluate its own decision and will give you the final one. ChatGPT-4 was launched few months back and it was in hype. But this Deepseek R1 has been launched recently and been compared with the ChatGPT 4. That is why Deepseek R1 is on hype, because ChatGPT has so many features and has come up with a lot more for us. Like we can do, uh, the problem solving stuff, we can ask questions related to coding, reasoning, and we have so many things, so many features on that platform. Same way, Deepseek R1 is there in the market with all those features at less price. Yes, while making the Deepseek R1 model, they have made sure that it should be cost efficient from the developer's end as well and the user's end as well. It should be cost efficient. How they have made that cost efficient? We know that ChatGPT-4 was built on $100 million, and that is why it was able to work on so many parameters. We have so many tokens for that as well. But R1 model is being made with the 1/20th value of that particular ChatGPT. That is why they have made it, you know, cost efficient. The main reason was they haven't trained that model from scratch. What they have exactly done is they have used the reinforcement learning in this. This reinforcement learning helps your model to learn from the experience or even from the feedback as well. This works on the feedback loop, and this is what they are doing in this particular model. They have used GRPO— they have used GRPO method instead of using the traditional Monte Carlo. Monte Carlo method was something else. It was training everything from scratch, taking the billions of data, and then train— making your train-test, uh, splits of the datasets, and then making your model work, right? But this DeepSeq R1 model is working on the reinforcement one which is GRPU, which stands for Generalized Reinforcement Policy Optimization technique. In this technique, they are using other LLM models to train this model, and this primary model is going to work on the feedback loop, and it will reevaluate it, it will use the aha moment, and then it's going to give you the results. I think you should better try it first. And how to try it? You can use it on Ollama website. Yes, you have to simply download Ollama on your system. Once you have downloaded Ollama, you just have to select the model that you want. This is not a single model which is an open-source model. There are so models in market which is an open source model. You can use any one of them which is just around the Ollama. Ollama is basically a platform that provides you a privilege to upload your models for free of cost and you can make it access to any developer. First they have to do is they have to download Ollama on their system, which is pretty easy. Within a single click you can download it. This is not something like we have— you have to do setup, you have to take care of certain things. No, you just have to download it on your system. Once you have downloaded, you have to choose the model that you wanted to run. You have to go to the models tab, and in the models tab you will be able see all the free open source models over there, and you have to choose one for yourself. And then you can copy that command and come to your CMD, Command Prompt Terminal, if you are using Windows. And if you are using Mac, you can go accordingly to the prompt. Then you have to enter that command. It will take few seconds to install that particular model on your system, to run it on your system, and make it compatible with all of the features. And then you can start chatting directly over there, which is pretty easy. So I hope this video was helpful for you to know the new technology that we have in market. So don't forget to subscribe to my channel and comment down what should I cover up in the next. Bye-bye.
