Speaker 1: OpenAI has been crushed, this is not the question, the question is whether the developers have also been crushed. In today's video, I will talk to you about the new model of DeepSeek, which has made AI world go crazy. How and what you should know about it, we will discuss in this video. So basically, OpenAI O1 was a very dangerous model. And I am saying it was because DeepSeek has released a model in today's date which has beaten OpenAI O1 in different benchmarks. Well, there is one thing to beat, beating OpenAI is a very big thing in itself. Then beating it through an open source model is a very big thing. The surprising thing is that you can use DeepSeek, we will use it in this video. I will tell you how you can do it. And you can use it commercially, that is, if you make money from it, then there is no problem, you can make DeepSeek. OpenAI O1's license was $200. And I am saying it was because maybe these people will have to revise their pricing. Because 26 to 28 times cheaper DeepSeek. Like today's DeepSeek. It is available on the 20th and anyone can run it on their computer. In fact, an ordinary person like me and you can run it on your computer. And I will show you how you can do it today. Well, these are good things. What impact will it have on jobs? What impact will it have on my and your jobs? What impact will it have on the jobs of normal developers? How is the future looking? We will also talk about this. Now the thing is that this model was taken out by China. And there are a lot of accusations against China. I will tell you what the accusations are. It's going to be a very interesting video. And I will tell you some things that you will get to learn a lot by listening to it. But you will also have a lot of fun. I will not go into politics in this video at all. I will show you a thread on Reddit. You guys read it yourself and decide what it is. But first, let's come to its research site. I read its paper. It is very intuitive. Let's bring its paper to the screen. Basically, what happened is that they did this. As we used to train LLM with fine-tuned data before. Now they have changed the approach a little. They said that we will use reinforcement learning instead of supervised fine-tuning. What does this mean? I will tell you. I will also tell you how to run it. But let's talk a little about Deepseek. Now Deepseek is not a company of today. Deepseek is a very old company. It is a Chinese artificial intelligence company. Those who are already involved in this kind of research. They have been doing it till today. But their models were good. It was not so much that they beat open AI. Deepseek was started as a side project. And it is used a lot in AI research in China. Now this is not the first model of Deepseek. Let me clear it for you. Many people have heard its name today. They are saying that Deepseek is not of today. It is a very old company. People have released their models. But they have brought a breakthrough in today's date. That's why there is so much talk about them. Hangzhou is their headquarters in China. And they are doing a lot of hard work from there. They are doing a good job. But there are some accusations. We will talk about it in a while. Now they also have money support. There is a very good support of a Chinese hedge fund called Hiflare. But if we compare with big companies like open AI. From Google and Facebook. So somewhere their name used to come down a little. And they don't have that much money to spend. I used to have a saying. In childhood. When I used to go to buy something. People used to say that it is very cheap. Buy it. If it works then till the moon. And if it doesn't work then till the evening. It seems that the model of Deepseek has worked till the moon. Now as I told you. It was used as a side project of Deepseek. They did not raise a very big, thick, long bunch of funding. But yes they had money. They have backing. But they don't have that much backing. They don't have that much open AI Google meta. Their employees are also not that much. There are only 200 employees. In comparison to open AI. Let me tell you. There are 4000 employees. Now how did they achieve success? By thinking differently. And brought innovation. Where everyone. Who was playing the same game. The same architecture. On LLMs. Data was being given. Data was being given. They were being trained. By doing different types of small tweaks. Things were being tried. And big models were being taken out. They used reinforcement learning a little. We will discuss about it later. Then once their reinforcement learning model was made. So they took a basic model. Through which they started cold. That yes we have a base model. Now we will improve it with reinforcement learning. And results were amazing. If I talk about architecture. Then I will not go into too much detail. But you guys read this paper. Let me tell you. One of the best written paper. When I read it. By looking at it. I thought how much I will understand. How much I will not understand. But when I read it. Then it was looking interesting. It was looking like a story book. It was actually fun to read. And yes. You guys will find it very complex. But if you guys will read. And step by step. You will understand the formula. Then you will understand. What you are actually doing. In short. I will tell you guys. Basically what they have done. That they have incorporated reinforcement learning. If you see in their abstract. Then they have explained this thing. We introduced our first generation reasoning models. DeepSeq R10. And DeepSeq R1. R10. A model trained. By large scale reinforcement learning. Without supervised fine tuning. In DeepSeq R10. They have used reinforcement learning only. Now what is reinforcement learning? Basically in reinforcement learning. There is an agent. Environment. Agent interacts with the environment. And. According to the actions of the agent. He is given a reward. And after that. To maximize his reward. He does good things. And in the long run. He only does the work. Which maximizes his rewards. Think of it as. There is a child. His father. He is a child. His dad gives him chocolates. When he works well. And if he misbehaves. He hits him with a slap. So. Over the long period of time. The child. He will try. Less lick. And he get more chocolates. And this is DeepSeq. R10. DeepSeq. R10. Here zero means zero. Supervised fine tuning. That is, we have not done supervised fine tuning on it. The model has learned that. By interacting with the environment itself. Now, what happened after that. They said. According to DeepSeq R10. There were some challenges. His readability was not good. Chinese was being mixed. There were too many problems of language mixing. So. What did they do? To address these issues, they came up with DeepSeek R1. In which they started a cold start with basic data. And after that, they incorporated the reinforcement learning spice in it. Now, the interesting thing here is that DeepSeek R1 has opened OpenAI O1 in different benchmarks. Which is considered a state-of-the-art model and is considered a very expensive model. And it comes through OpenAI, which has put the most dangerous players of this race behind them. I say that if you look at this graph, then come close to it. According to me, coming close to OpenAI O1 is a very big thing for any model. And DeepSeek R1 achieved this and showed it. Open source also did the model. And along with that, if you run it in today's date, it is 28 times cheaper. So if we don't call it a breakthrough, then what will we call it? This is a very big breakthrough in AI research. Now, here are some different benchmarks in which it has performed very well. But let me tell you an interesting thing. There is a benchmark called Humanities Last Exam. Humanities Last Exam. This is an AI exam. This is a benchmark. And if you look at it, then OpenAI O1 used to be at the top. These are the results. And you see here, in the Humanities Last Exam, the accuracy of DeepSeek R1 is 9.4%. The GPT-4 is 3.3%. You see, it has beaten with such a big margin. And being an open source model. Now, DeepSeek R1 is 28 times cheaper than O1. And it is going to be very accessible for the general public. The interesting thing is that the users have said that in iPhones and Android devices, they are able to run this model very easily. I saw some clips on YouTube and I enjoyed watching it. Now, people are running the open source model in the iPhone. That is, your AI has come into the iPhone. And such an AI is not going out of your phone. For organizations, this is huge. Because now they can use open source state-of-the-art AI without sending their data anywhere in their servers. Many organizations were forced to use O1 in today's date. Because now OpenAI has to use API because it is a state-of-the-art model. If they don't use it, they will be left behind. Will they use it? So, OpenAI will have to send data. But what can you do? If you have to send it, you have to send it. Now, what does OpenAI do? Once you send your data, your data goes away. After that, all these things are discussed. That data ethics will be followed. Your data is safe with us. All these things are discussed. What is being done with your data? Your data went out of your system. So, it went out. Going out is a very big thing. Now, that company is very ethical. Its founders are very good. Their server is secure. All these things are discussed. You know what I mean. I am not accusing anyone. Now, it is not that there were no good models in the market. There were very good models like LAMA. There were dangerous models. Google also did a very good job in releasing open source models. But that accuracy was not coming. OpenAI had set a different benchmark. DeepSeek did a great job. Now, what happened with this? See, good things have happened. A new model has come. We can use it commercially. Blah, blah, blah. Now, let me tell you some downsides of this. The first downside is in America. Global tech sell-off has come. Because now the people who were bullish on Nvidia, on OpenAI, the money that was being generated, the revenue of $200 was being generated by each person who was using OpenAI and OVEN. Now, what will happen to them? Obviously, their revenue is going to fall. And because of this, global tech sell-off has come. Let me show you the chart of NASDAQ. It looks something like this. If you see the graph of 5 days, you can see that OVEN has triggered somewhere. The whole thing is going on red. Anyway, I am not a stock market analyst. So, I will stop it for now. And the interesting thing is that I worked on a very small idea, converted it into reality, and produced an impact from it. What was the idea? What people were doing in a basic neural network? They were increasing the parameters. That is, the function of the neural network, they were making it complex. Now, if you fit any data on x square plus x plus 1, or if you fix it vis-a-vis, x to the power 4 plus x square plus 2x minus 1. So, the definite thing is that the big function whose degree is 4, the data will fit very well on it. Because its degree is 4. If you increase the parameters, if you make the function complex, then it can fit on big data. There are downsides to it, and there are also benefits to it. Yes, you have very powerful GPUs. Train it, train it, train it, and you will learn the AI model. But here, they used reinforcement learning, which did something different. This is a big example of producing an impact by doing something different. Now, what should we learn from this? The first thing we should learn is that if you do something different, do something different, do something new, then its result can either be positive, or your hard work will be wasted. You can either see it like this, that my hard work has been wasted, or you can see it like this, that you have learned something new. Now, here, DeepSeq must have learned a lot. This is not their first model, which beat O1. They must have tried a lot of other things. And finally, this thing worked for them. They published a paper, and they came out with a big heart. The second thing is that how can you make your applications? How can you make your SaaS, which is based on this particular model? And another very simple thing is that the more basic it is, the less people will use it. And that is, make something using this model. Make your own personal tools. In your own laptop, you can run this model and make a lot of things. So, make it and learn from it. Use this model. How to do it? I will tell you quickly. So, using the model is very easy. Download Ollama. I have already made a video. In fact, I have made multiple videos before. Download it for Windows. I am doing it for Windows here. If you are in Mac or Linux, download it for that. So, you can download and install Ollama. Download and install. It is very simple. I am installing Ollama here. Now, as soon as this Ollama is being installed, I will tell you a very interesting thing. This DeepSeek R1, it took only $5 million to train it. $5 million is not less, but it is very less than the money spent by OpenAI, which was tens of millions of dollars. Making this one of the cheapest models when it comes to training. Now, our Ollama has been installed. What I will do here is, I will run Ollama quickly. And after that, I will go to the models here and open it to run DeepSeek R1. And this is the smallest model. I will select it for this video. Which is of 1.5 billion parameters. I will copy it. After copying, I will paste the command Ollama run DeepSeek. Now, it will pull the manifest, download the model, and run it for me. Now, see, when you run it for the first time, it will run very slowly. Because it actually downloads the model. And after that, it will run it using the power of your GPU. Now, see, it is of 1.1 GB. It is showing the progress in this way. It will take a lot of time to download it. But let's wait for it to download. It is downloading very fast. It is showing 4 minutes here. So, we will download it. Let's wait for it to download. Now, as you can see, the model is running here. Now, again, let me tell you, that my laptop is using 30-50 GPUs. It is using 16 GB RAM. It is a very basic laptop. So, you can't expect much from it. But I will tell it, what is HTML? I mean, I will ask it. And see, it is running so fast. It is running so fast in my GPU. See, what it has written here. Now, I will tell it, do one thing, create a tiktok, tiktok, game in, HTML, CSS, JS. And it is running very amazing. Let me show you the specs of my computer. So, you can know, what you can do with it. So, see this, these are the specs of my computer. It is in front of you. It has 16 GB RAM. And NVIDIA 3050 Laptop GPU. It is very basic. I had bought this laptop for 55,000 rupees, one year ago. I had bought it for sale, for Flipkart. And see this, this is amazing. Sahab, this is amazing. This is amazing. State of the art model, open source model. You can also run it. Create its web interface. You can do a lot of things with it. So, do it. Now, let's talk about, the accusation on China. What is the accusation on China? Basically, the accusation on China is that, these people are putting censorship, they are putting bias, in the answers of this AI. Again, this is an accusation. I am not going to get into political things. You will see, but this is a very big accusation on them. It is being said that, they are hiding the facts. And, they are manipulating their answers. If any historical thing is being asked, they are manipulating it, and giving answers in their favour. And about this, there is a very big thread on Reddit, on which people have written a lot of things. You have to read it. I will put the links, both are below. And even if I don't put it, then you have to search it, by doing trending topics, you will put this title, it will come on Google. And it will also come on Reddit. You will write, censorship to questions about China, Deep Sea Garbhan, it will come. See for yourself, it will be fun to read. You should read all these things. Now, let's talk about, those things, about which I have promised you. We will talk about, what impact it will have on developers. So, the impact on developers is in front of you. This state of the art model, is doing everything. And when you will evaluate these responses, then you will know, your mind will blow. How accurately, these things are done. And on top of that, if you have taken it in context, then accuracy goes to a different level. I have faced this myself. I have seen it myself, how, accuracy reaches a different level. So, see, all these things, will impact the developers. Positive, negative, in both ways. In the negative way, they will say, yes, some experience will be given to the developers. Expectations will increase a lot from the developers. But again, all these models are also made by the developers. If you want to beat AI, you will have to learn AI. I always tell you this. Even today, my answer is the same. Whose job will go, whose will not go. We will discuss more on all this. Depression will increase. According to me, we should talk about, what to do. And what to do, is that, to use this kind of knowledge, to read this kind of papers. You have to get into AI ML. You have to learn about data science. Because, there is a future. If you are not in this domain, then I will tell you, just learn. But, even if you are a little bit, learn. Its future is very bright. Now, let's talk about how you can earn money from it. First of all, you can earn money by using it. Use it in your workflow. Make some SaaS. Making an app on today's date, is extremely easy. Make a Next.js app, 3 years before today. Or I will say, with the help of HTML, CSS, JavaScript, an e-commerce website, it was very, very, very difficult. By using such models on today's date, you can literally make a very good website in a week. Do it, man. Again, you don't do it. I mean, people don't do it. People come to me, ask a lot of questions. So many requests came for this video. But, you see, if something has come, which is beating the developers, try it. I have tried it. These things, are still very, very far behind, from a good experience developer. A good experience developer, by using these things, can increase his productivity. And, the expectation from him, will definitely increase. But, again, he can save his time. And, he can take humanity to a different level. You can see things like this. How can you make your workflow better? What tools can you make, which can solve your day-to-day life? What are those things, which you were doing today, you can do it with BI, so that your time is free, and you can do something bigger than that. Think like this. And if you think like this, then you will go very far. What do you have to say about this model? Try it. Try it. And tell me in the comment section, what do you think about this model? And if you want to see more videos like this, then tell me in the comment section, that you want to see more videos like this. I am not saying, just for the sake of getting comments. I literally want to know, that do you like these types of videos? Should I make more videos like this in the future? If yes, then tell me in the comment section. I hope you liked this video. Practically, we saw the model. And, there is a lot to talk about in this video. I enjoyed this video. I hope, you must have enjoyed it too. Thank you so much guys, for watching this video. And, I will see you next time.