Speaker 1: This was just a side project. They didn't seek any external funding. They only have around 200 employees. For your comparison, OpenAI has over 4,000 employees. They don't even have the best GPUs out there, but this company just released an insane open source model. It's called DeepSeek R1, and this is completely free, open source, and uncensored. And you can use it right now. Some users are even able to use this on their iPhones or Android phones. And here's the crazy thing. This model even beats OpenAI's O1, which is their flagship model. This is a PhD level AI. DeepSeek R1 is able to beat it or match it on various benchmarks. So finally, open source has caught up to, or even surpassed, the best commercial models out there. So in this video, I'm going to go over what exactly is DeepSeek R1. I'm going to go over, its architecture, how it works, what are the different models you can use, and some really cool use cases that other users have shared. Let's jump right in. So I'm going to link to this main page in the description below, but they also have a technical paper, which you can read. If you really want to get into the nitty gritty details of how they trained this, how they designed this, I highly recommend you read this technical paper. In this video, I'm just going to really briefly go over their process on creating this. So this DeepSeek R1 model is very different from traditional large language models like ChatGPT or Cloud 3.5, which use supervised training, which I'll talk about in a second. In contrast, DeepSeek uses what is called reinforcement learning to train the model. So what exactly is reinforcement learning? In the simplest sense, here's what reinforcement learning looks like. So here we have the agent, or the AI basically, and it takes a certain action. And from that action, for example, if it answers the prompt, then it gets a reward. Or if it answers the prompt wrong, then it gets a punishment. So think of this as like training a dog, right? When the dog does something good, it gets a treat or the reward. Or if it does something bad, it might get scolded, which is basically the punishment. So over time, the dog learns which actions lead to treats and which actions lead to scolding. So similarly, when you're training an AI using reinforcement learning, through this feedback loop, it learns the best actions to take in order to, get the highest reward, which in our case would be like answering prompts correctly or achieving the highest accuracy or achieving the highest benchmark scores. So that's reinforcement learning. And so here they say that they trained this DeepSeek R1.0 using reinforcement learning without any supervised data. So what exactly does this mean? So traditionally, how like ChatGPT or Lama or other AI models were trained is that it's fed a ton of data, but you're also providing the answer for it to verify. So for example, let's say you're training this AI model to identify images of cats or dogs. Well, during training, you would feed it tons of images of cats or dogs, and it would spit out random answers at first, but then it would compare its response with the correct answer, which you give it. And if its response is wrong, then the AI model would basically reconfigure its weight in order to minimize the error and get it correct the next time. So because you're providing the answer for this AI to compare its response to, this is called supervised data or supervised learning. However, for this base model, DeepSeek R1.0, the authors say that it was only trained using reinforcement learning without any initial guidance or supervised data from humans. This means that it learned entirely from its own experiences. For example, if it's fed a math problem, well, we don't actually provide the correct answer, so we don't really have to do anything to help it. So not only does it have to try out different ways to solve the math problem, but it also has to verify its own responses. This is like a completely different model from what we've seen before, but as a result, it has developed advanced skills like self-checking its work and thinking through complex problems step by step. And this is a huge breakthrough in AI research. It shows that AI could just improve and learn by themselves without needing extra help from humans. Another really interesting insight from this paper, because it's using reinforcement learning and it's not guided by any humans or pre-existing answers, it has to figure out things on its own. Here's an example of DeepSeek actually discovering a new technique and calling it an aha moment. So here the model discovered a new way to solve this problem all by itself without any guidance from humans. So this again demonstrates the AI's ability to adapt and improve its problem-solving methods without any explicit instruction. So that sums up the creation of DeepSeek R1.0. Afterwards, they created an even better model called DeepSeek R1. So here it says, inspired by the promising results of DeepSeek R1.0, two questions arise. Can reasoning performance be further improved by incorporating a small amount of high-quality data as a cold start? So in other words, instead of just getting the AI model to learn everything from reasoning, it can be done by incorporating a small amount of reinforcement learning. Can we at the start feed it high-quality data and would this improve its performance? And then how can we train a user-friendly model that not only produces clear and coherent chain of thought, but also demonstrates strong general capabilities? So to address these two questions, they basically created a hybrid training approach to create this DeepSeek R1 model. So basically without going too deep into the technical details, they fed it with high-quality data and would this improve its performance? And then how can we train a user-friendly high-quality chain of thought data? And this data has detailed answers with reflection and verification, and it's in a very readable format. And this approach helped fix some issues like confusing language or unclear writing that was experienced by this original DeepSeek R1.0 model, which was only trained using reinforcement learning. And then after processing this high-quality data, then it continued with reinforcement learning, where there is no supervised data. It just learns to solve everything and double-check everything by itself. So again, note that this is a hybrid approach. It gets some high-quality supervised data initially, and then it continues with reinforcement learning or learning by itself. And by doing this hybrid training pipeline, it actually performs really damn well. So here is a table of various benchmarks from like English, coding, math, and Chinese, and the values in bold indicate the highest performance. So here is a table of various benchmarks from like English, coding, math, and Chinese, and the values in bold indicate the highest performance. So you can see that for most of these, DeepSeek is actually better than OpenAI's O1 model, which is like their flagship model. This is a PhD-level AI that can solve really hard math, coding, and science problems. But you can see DeepSeek, especially for math, just crushes O1. This is absolutely insane. Here's another chart, which I showed at the beginning of this video. DeepSeek R1 is the dark blue bar on the left, and you can see for R1, it's the dark blue bar on the left. And you can see for most of these benchmarks, it beats OpenAI's O1, which is in dark gray, or is at least tied with it. Again, keep in mind, DeepSeek is a completely open source and free model that you can run locally, whereas OpenAI is closed source. We have no idea how O1 was designed or what exactly it is. So this is incredible progress. Now, those are just their published metrics, but is it really that good? Let's look at some independent evaluations. New benchmark called Humanity's Last Exam. This is a pretty insane benchmark. This is designed to be the world's most difficult AI benchmark, and it really tests AI models on expert-level capabilities across various fields. And as you can see here, even for the top models like GPT-4.0 and Cloud 3.5 Sonnet and Gemini Thinking, or even O1, its performance is pretty bad. I mean, this is under 10%. GPT-4.0 is under 4%. But you can see out of all these models, DeepSeek R1 actually scored the highest. So this is pretty amazing considering out of all these models, this is like the only one that's free and open source and uncensored. So that's Humanity's Last Exam. Next, let's look at another independent evaluator called LiveBench by Abacus AI. And here you can see again, DeepSeek is ranked at number two, even above Google's flagship Gemini 2.0 Flash Thinking. Its average score is just a few percentage points below OpenAI's O1. So this DeepSeek model is legitimately good. Here's another independent evaluator called Artificial Analysis. And here's their leaderboard ranking the top AI models. And you can see again, DeepSeek is ranked at number two. It's only one point behind O1. But if you look at like the top five results, DeepSeek is the only one that's free and open source. Now, the models are already out. So you can actually download and run this locally or tweak these however you want. Or if you don't want to run it locally, there are also a few online platforms where you can use it for free. So for example, DeepSeek has released a native chat interface where you can chat with DeepSeek R1 for free. And there's actually two features of this chat. One is this DeepThink feature. So if you turn this on, you'll actually see its thinking process as it answers your question. So for example, if I want to answer a question, I'm going to press enter. And if I want to answer a question, I'm going to press enter. So here, because we've turned this DeepThink button on, you can see that it's actually thinking through how to answer our question, just like how a human would think. And then afterwards, it confirms that yes, 9.9 is bigger than 9.11. You can also click on this to search the web and retrieve the most relevant and latest information. So let's turn off DeepThink and let's search the web. Tell me about Trump's announcement of the $500 billion Stargate project. Now, this only happened two days ago, so there's no way that this AI would know about this in its training data. We'll need to turn this search feature on for it to search the web and retrieve the latest information. Let's click enter and see what that gives us. So you can see it's searching the web. It has found 50 results, and it's pulling information from all those results. And indeed, it gets all the details correct. It includes a partnership from OpenAI, Oracle, and SoftBank. And it's pulling information from all those results. And yes, it was unveiled on January 21st, 2025. So you can use this chat tool by DeepSeek just like Perplexity, a very powerful tool. All right, next, let's turn this off. You can also upload documents like PDFs for it to analyze. For example, I'm going to upload this technical paper about this new architecture called Titans onto here. And then I'm going to ask it to summarize this paper in one paragraph. In fact, if you haven't heard about this paper, this is basically a new AI model that incorporates memory so it can continually learn and remember new things. If you're interested in this, definitely check out my latest video where I do a deep dive on this paper. Anyways, back to DeepSeek. Here you can see it has indeed given me a one paragraph summary. And just from a quick scan of this, this does look accurate. So not only is this a great tool for chatting, but you can also upload documents for it to analyze. Now, this is just one of many platforms where you can use DeepSeek online for free. Here's another really cool place where you can use DeepSeek. This is a free Hugging Face space called AnyChat by Akalik. And you can select from many different AI models, including DeepSeek Coder. This basically uses DeepSeek to code up whatever you want. Plus, you can preview the output in this right panel here. So it's a really useful tool for prototyping. So let's go ahead and start with DeepSeek. So let's go ahead and start with DeepSeek. So let's go ahead and start with DeepSeek. Let's prompt it with create a Spotify home page clone, and then click generate and see what that gives us. So you can see in the left panel, it's generating the code for us. And then over here is a preview of what the page looks like. How insane is that? Very powerful tool. Let's try something else. Let's try create an interactive synth piano with adjustable settings. And let's click generate and see what that gives us. So again, note that it's first generating the code in the left panel here. And then after it's done, we're going to see a preview in this right panel. All right, so here's what we get. Let's click on a few of these keys and see if it works. Yes, it does. And let's select a different waveform. Let's select square. Awesome. Let's select sawtooth. Awesome. So you know, within like three seconds, notice I even spelled piano wrong, but it was able to generate this interactive synth piano that works. Now, those are just some basic examples. I'll link to this Hugging Face space in the description below. But here are some even cooler use cases that others have posted. So this person got DeepSeek R1 to create an entire animation explaining the Pythagorean theorem. And he says this was done in one shot. So with just one, prompt with no errors in less than 30 seconds. By the way, it's outputting this in a coding language called Manum, which is an open source tool for creating mathematical animations and explanatory videos. So pretty crazy already how DeepSeek R1 is able to create this. Now explaining the Pythagorean theorem is pretty easy. I mean, this is just like high school level stuff. So next he got it to create an animation about quantum electrodynamics, which is way, way beyond my knowledge, but he claims that it was also able to do this single shot without any mathematical error. So again, a really impressive and powerful AI model. Thanks to AI Portrait for sponsoring this video. A good professional photo on your LinkedIn or business profile makes a huge difference. Now, you could try to take them yourself or get a friend to do so, but most people aren't great at taking professional photos. Or you could hire a professional photoshopper to take a photo of you. But you could hire a professional photoshopper to take a photo of you. Or you could hire a professional photoshopper to take a photo of you. Or you could hire a professional photoshopper to take a photo of you. Or you could hire a professional photoshopper to take a photo of you. But this costs over $200 on average. Plus you need to schedule a session and spend hours awkwardly posing at a camera. And this brings us to AI Portrait. You can generate a portfolio of professional high-quality photos in just minutes. Just upload one photo, choose your gender, and it would generate a portfolio of 50 professional headshots in various settings. And it generates this within minutes. So if you're looking for high-quality professional photos without the hassle of a physical photoshoot, AI Portrait is your best bet. Check it out via the link in the description below. Here, this other user asked DeepSeek to implement a rotating triangle with a red ball bouncing in it. And they also prompted the same thing using OpenAI's O1 Pro on the left. So you can see even O1 Pro, which is one of their top models, is unable to generate this animation. But as you can see with DeepSeek, which is on the right, it's able to handle this perfectly. So those are just some cool and creative ideas you can use this for. Now, like I said, they've already released the models for you to download locally. And they've actually released a few models of various sizes. So here are the main models. Like I mentioned, DeepSeek R10. This is the base model that was created only using reinforcement learning. And then we have DeepSeek R1, which is a slightly bigger model. And then we have DeepSeek R2, which is a slightly bigger model. That was trained using both initial supervised data and reinforcement learning. Both these models have 671 billion parameters. And they have a context length of 120k, like most commercial AI models out there. Now, with 671 billion parameters, of course, this would be too large for most of us to run locally. Fortunately, not only did they release these two main models, they've also released several other smaller variants that are based on smaller models. So these are the two main models that are based on smaller models. So what that means is, for example, let's take this one, DeepSeek R1 Distill Llama 8b. This is based on the Llama 8 billion parameter model, which is way smaller than this, of course. And what this model means is they took the outputs of DeepSeek R1, and they used it to fine tune Llama 3.18b. So you could say, figuratively speaking, that it's like passing some of the intelligence from DeepSeek R1 into this Llama 8b model. And so there are a total of, I don't know, six smaller models that they've released. The smallest one is 1.5 billion parameters. Now, you might be wondering, these are smaller models. This one is like only 1.5b. These must suck compared to the base model and compared with, you know, the leading commercial models out there like 4.0 or Claude 3.5 Sonnet. Well, there's one key sentence that's hidden somewhere in this technical paper. Let me try to find it. Here we go. So it says, this DeepSeek R1 Distill Quen 1.5b model is based on 1.5b. This 1.5b model outperforms GPT 4.0 and Claude 3.5 Sonnet on math benchmarks. Are you kidding me? This model with 1.5 billion parameters actually beats the leading commercial models out there, GPT 4.0 and Claude 3.5 Sonnet on math benchmarks. I mean, both of these have like hundreds of billions of parameters. So the fact that this model, which is over 100 times smaller, is able to beat these two models in math is also just... mind-blowing. So that sums up all the models that are available. Now, if you do want to run this full model, but you don't have enough compute, you can also use the DeepSeek API. And here's the crazy thing. The cost to use DeepSeek is cheap as hell. So if you use their API, it only costs $2.19 per million output tokens. If you compare this with OpenAI's O1, that costs $60 per million output tokens. So I can't do public math. So if you compare this with OpenAI's O1, that costs $60 per million output tokens. So I can't do public math. I have to use Google. DeepSeek is like 27 times cheaper than OpenAI's O1. Plus it's free, it's open source, it's completely uncensored, and it matches the performance of O1. That is absolutely insane. We are at the age where intelligence is too cheap to meter. That being said, if you don't want to use their API and you want to run everything locally, well, there are a few options. Like I said, 1.5b is the smallest option, and some users have already successfully used it. So if you want to run everything on an iPhone, there are many options available. This comparison provides a clear example of how this can really be applied. This 깊이HAHAHAということで Canyon can run it on an iPhone. But if you don't have an iPhone and have an enjoyed, it also works. So here's an example of this user successfully installing and running this offline on his Android device. I have gotten an email from Verizon exchange.com telling me that, of course, they got the logs. But I want to taste human origin. And what do I do? Android phone. And again, note that this 1.5B version actually beats GPT-4.0 and Cloud 3.5 Sonnet in math, which is pretty ridiculous. And if you don't want to just run a tiny 1.5B model, some users have also successfully run the full 671 billion parameter model. So this person got it to run successfully on just two M.2 Ultras. And look at the speed of its response. This is blazing fast. And again, this is just incredible. We now have an open source model that's as good as OpenAI's O1, which you can run on just consumer grade hardware. You don't need to rent an entire GPU farm to run this. Here's another crazier setup. So this person apparently linked seven M4 Pro Mac minis together and one M4 MacBook Pro in order to run the full version of DeepSeek R1. But as you can see, it does work. And it answers very quickly. And I mean, compute is only going to get faster and more efficient. For example, Nvidia is set to roll out their AI supercomputer in spring of this year. And that enables us to run even larger models right in our homes. So this year is going to be really exciting. Finally, I want to end with this post from Dr. Jim Phan from Nvidia. So here's what he posted in regards to DeepSeek R1. We are living in a timeline where a non-US company is going to keep the original mission of OpenAI alive. Truly open frontier research that empowers all. It makes no sense. DeepSeek R1 not only open sources a barrage of models, but also spills all the training secrets. They are perhaps the first open source software project that shows major sustained growth for a reinforcement learning flywheel. What an ironic turn of events. OpenAI's original mission was to create open source. AI for the benefit of humanity. But right now, all their top models are closed. We have no idea how they were trained. We have no idea the architecture of these AI models. Ironically, DeepSeek, a Chinese company, is the one that has open sourced a state-of-the-art model that actually beats OpenAI's R1. So what a crazy plot twist. Anyways, that sums up my deep dive on DeepSeek R1. Hope that gives you a good sense of what it can do and how impressive this is. Let me know in the comments what you think of this, and if you've tried it out, let me know what other cool things you were able to achieve with this. As always, I will be on the lookout for the top AI news and tools to share with you. So if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content. Also, there's just so much happening in the world of AI every week, I can't possibly cover everything on my YouTube channel. So to really stay up to date with all that's going on in AI, be sure to subscribe to my YouTube channel, and I'll see you in the next video. Also, remember to subscribe to my free weekly newsletter. The link to that will be in the description below. Thanks for watching, and I'll see you in the next one.