95% Cheaper AI? How DeepSeek Is Taking on Big Tech

The release of DeepSeek AI from a Chinese company should be a wake-up call for our industries that we need to be laser-focused on competing to win because we have the greatest scientists in the world. Even Chinese leadership told me that.

Let me break down DeepSeek's latest AI innovation and why it's blowing everyone's mind and why you should care. Now, let me get you caught up to speed first. So a couple of weeks ago, this unknown Chinese company launched an LLM, or large language model. So it's equivalent to OpenAI's LLaMA or Anthropic's Claude. But after this model launched, it completely plummeted the Nasdaq, and Nvidia lost 20% of its market cap within days. That resulted in hundreds of billions of dollars being completely wiped out.

Wall Street rattled by fears that China caught up to America on artificial intelligence. Nvidia leading the sell-off, down about 15%, tanking on the news that DeepSeek has an AI model similar to ChatGPT but operating at a fraction of the cost.

What's the big fuss about DeepSeek? Now, let's first take a look at how traditional AI models are trained, and you'll be able to see why DeepSeek was such a big technological breakthrough. So when we're taking a look at how traditional AI models are trained, is trained using a neural network. Essentially, scientists was able to see how humans were able to learn these, and our brains are actually connecting neurons together so that we retain information. They then wanted to copy the same pattern into computers so that computers can learn things as well. And this was how the neural network was formed, and it's used to train large datasets. Now, there are 3 different components to a neural network. You have your input layer, you have all of your middle hidden layers, and finally you have your output layer. Let me give you an analogy so that this all makes sense. So let's suppose that you are going into a coffee shop to place an order. Your order the specific type of coffee, the amount of milk you want, that is the input layer. Then when the barista makes your coffee with the exact amount of coffee, milk, and sugar that you want, that is the hidden layer. And lastly, you have the output layer. This is the finished product. This is the prepared coffee that you will get. In the AI world, this is the prediction or the end result that the AI model was able to generate for you. So in layman's terms, how these LLMs are able to learn things is, let's say I show it a picture of a cat, then I show it another picture of a cat, then I show it another picture of a cat, and I ask it, what's this picture of? It's able to know that this is a cat. It's basically learning through repetition and pattern recognition, and it's very expensive to run these models. In fact, it costs about $100 million to train ChatGPT-4. So this Chinese company came in and said, what if we can rethink this? What if we can disrupt the market by asking different questions instead of trying to beat OpenAI at its own game? Why don't we reinvent and create a new paradigm? So the old paradigm is the more powerful the GPUs that you have, the more powerful the AI. That's why Nvidia was so powerful. But this Chinese company came in and said, what if we put in clever engineering so that we can get the exact same results without powerful semiconductors. This was also because President Biden put a ban on giving China the latest semiconductors because they wanted to be at the forefront of the AI revolution. So China was basically cut off from the latest semiconductors, which means Chinese AI companies can't have as much GPUs as companies like OpenAI or Anthropic, which are based in the US. So DeepSeek had to be innovative and be able to compete on a global scale with less resources. And here's the story of how they were able to do so. Also, if you're new to my channel, hi, my name is Helena. Please be sure to like and subscribe if you want to learn everything about AI and automation. So let's take a look at this model here. We already explained how a neural network works, and this is how OpenAI was able to train their O-1 model. So what DSeek did was they distilled the information from the teacher model by compressing it and simplifying it, and they were able to obtain this information by continuously querying the OpenAI API key. So essentially, they asked a question and asked another question and was able to collect all of the responses from OpenAI LLaMA 1, and it used this information to train their own model, which is a much smaller, more efficient neural network that's learned from the distilled knowledge. The student model requires fewer computational resources while maintaining performance close to the teacher's model. In fact, across all benchmarks, DeepSeek's model is very comparable to ChatGPT's L1 model. But here's the staggering thing: it only cost $5.5 million to train DeepSeek's R1 model, In comparison, it cost $100 million to train the O-1 model. This is basically a 95% decrease in the amount of cost, and this is reflected in DeepSeek's pricing as well. Currently, anyone can use the DeepSeek online version for completely free, which is allowing them to gain a lot of market share on top of OpenAI. And this is very scary for them because currently OpenAI is the market leader. But for how long, we don't know. The AI wars are really heating up. All right, the last thing that we can see from this model is the data. The data represents the dataset used to train both models. The teacher model first learns from this dataset and then helps the student model generalize the knowledge. A side note, if you think this is legal, it is actually not. OpenAI is actually pursuing legal actions against DeepSeek. But this also begs the question, where did OpenAI get all of their data from? It is actually well known that they trained their model on a lot of proprietary information across the web that they never had permission to use in the first place. So isn't this all a bit ironic? Let me know in the comments below. What I also want to point out is that DeepSeek rethought the entire game. They didn't try to beat the existing big tech at their own game. They created a whole new game to play in. So it's basically a smaller model that basically gives you the result of the bigger model because they went Well, let's say you have a number. Instead of writing 35 decimal places, DeepSeek came in and said, what if we just wrote this number up to 5 decimal places? It will be 95% accurate and it will save us 75% of the memory. And that's exactly what they did. On top of that, they invented a new multi-token system. So when we look at traditional AI, the AI system will read one word at a time. The dog ran, so forth and so forth. But Deepsea came in and said, why don't we read entire phrases at the same time? This will allow them to compute the answer 2 times faster, and it still retains 90% of the accuracy. And here's another bit of clever engineering. So traditional AI models have trillions of parameters. In fact, ChatGPT's L1 model has 1.8 trillion parameters. So Deepsea came in and first said, let's use the 80/20 rule. What's 20% of the parameters that can give us 80% of the result? So they created a model with just 671 billion parameters. And they also went, well, we don't need to use all of these parameters for every single inquiry, so why don't we just call the necessary parameters based on the input of the user? It's basically like having a expert system, as I would like to call it. So let's suppose you have a whole room of experts, but you only call on the expert that you need to use for that specific task. For example, if someone's asking DeepSeek to code, it's only calling on the coding expert. If they are asking a question about marketing, it's going to call on the marketing expert. Traditional AI models would activate all 1.8 trillion parameters all at once, which makes it very expensive to run. By reducing the number of parameters that's going to be used in every single inquiry, DeepSeek is able to save a lot on computational cost, and the results are absolutely staggering. Instead of having to spend $100 million to train an AI model, now DeepSeek was able to do it for just $5.5 million, representing a 95% decrease in cost. And instead of having to use 100,000 GPUs to run these AI models where not many of us have access to an entire data center. Now you can run your own AI model with just 2,000 GPUs. And DeepSeek also released a smaller model with just 6 to 7 billion parameters that any of us can now run on our local desktop. This has completely changed the game and the implications are profound. AI is no longer just a game for big tech. Anyone, including you and I, can make our own AI models now and participate in the AI race. This means that competition will be heating up. Hardware costs are going to be plummeting. This is super scary for NVIDIA because they have always been operating on the paradigm that the more powerful the GPU, the more powerful the AI. Now with DeepSeek, we know that's no longer the case. It's like any of us can now build a Ferrari in our garage. This would be really scary news for Ferrari, right? The best part is, instead of taking the OpenAI closed-system approach, DeepSeek decided to do an open-source approach, so anyone can access their code and cross-check their work and improve upon it. All of the existing big tech AI companies are now scrambling to incorporate these new technologies into their models as well, and maybe this is a start of a new Cold War between US and China. The biggest concern with using DeepSeek is privacy. Instead of having all of your information stored on American servers, whenever you do run DeepSeek, the online version, all of your data will be stored onto Chinese servers. Now, there is a way around this because all the code is open source. We can actually run the model on our very own desktop. So if you have a very powerful desktop with a lot of GPUs, you can actually run this 671 billion parameter model. But if you have an average desktop, you can probably run the 7 billion parameter model. So let me show you how you can do this. So first, to run your own AI model, you want to go to a website called lmstudio.ai and then download the appropriate version depending on what kind of computer you're using, whether it's a Mac, Windows, or Linux. Then you want to go up here across the top where it says Model Catalog to select the model that you want to run. So for example, if you want to run the DeepSeek Distilled model with 7 billion parameters, you would just open this up and then press on Use in LM Studio here, and it will start downloading this model inside of your LM Studio account. After you have this model fully downloaded, you just chat with it just as if you're chatting with ChatGPT. The interface, as you can see here, looks pretty much exactly the same. So it's as easy as that to now run your own AI model on your own desktop. And the implications of this is profound. Gone are the days of just throw more GPUs at it approach. AI development has now become more accessible than ever. Everyone can now participate in AI. It is no longer just limited to big tech, and hardware costs are going to plummet because now we can have more powerful AI models running on subpar semiconductors. In fact, we can all run it on our own desktops now. That is mind-blowing. This feels like one of those times that we're going to look back years from now as an inflection point. This is a point in history where we have leapfrogged in technology. It's like when we all got PCs in our homes and we no longer relied on mainframe computers that were reserved for large corporations. Now all of us have a second brain basically in our home at all times that we can access at any time. And DeepSeek, by open sourcing their model, pretty much reinvented the whole game. AI has just become more accessible and less expensive than ever before. DeepSeek charges 96% less than O1 model. In response, Sam Altman actually created a new O3 model mini that anyone can now use for free. So the question now is not if this will disrupt the current players, but how fast it's going to disrupt them. All right, thank you so much for joining me for today's video. If you learned anything new, Please like and subscribe because that will help out this channel so much, and I'll see you next time.