Speaker 1: Hey friends, hope you're well. In case you missed it, DeepSeek R1 has been making waves in the AI and tech scene. It's an open-source AI model that was apparently developed for less than $6 million, a fraction of the billions of dollars spent by OpenAI and Google, for example, to create their AI models. The good news for all of us is that DeepSeek's free to use and it's shot up to the most downloaded app on App Store, surpassing ChatGPT within days. It's now one of the most advanced free and open-source AI models we can use. I've been playing around with DeepSeek R1 for a couple of days and I have to say it's game changing, but it's not without its flaws. So let's run through what DeepSeek R1 is capable of and see what the fuss is all about. So let's get up to speed. One of the main reasons why DeepSeek R1 is so hyped is because it doesn't rely on expensive human-labeled datasets or supervised fine-tuning, which is how most AI models are trained and it costs millions, if not billions. Instead, DeepSeek R1 uses a self-reinforced learning method without the need for human supervision and effort. You can think of supervised fine-tuning like teaching a child to cook by writing up a long and precise recipe and then showing them step by step while reinforcement learning is allowing the child to sort of like experiment in the kitchen and gently guiding them when dishes don't turn out well. So they're learning through trial and error and that's exactly how DeepSeek was trained. And the benchmark results are incredible. On the AIME 2024 Mathematics Benchmark, DeepSeek achieves 71% accuracy while GPT-01 Mini achieves 63.6% accuracy. And on the Math500 Benchmark, it beats both O1 Mini and O10912, but it performs worse on coding tasks in CodeForce and LiveCode benchmarks. But of course, there's much more to benchmarks. So let's jump onto the laptop and I'll show you what I found while playing with DeepSeek over the past couple of days. So jumping onto deepseek.com, here's where you can create an account or you can go ahead and download the app on your phone. But currently their servers are super slow because of the crazy demand. So I recommend avoiding signing up with an email. You'll be probably waiting forever for an email verification code. So I suggest logging directly through a Google account. So once you're in here, toggle on the DeepThink R1 model here. It's an advanced reasoning model similar to GPT-01 model, but without GPT-01's 50 message per week restriction. And also R1 is able to work alongside Internet search toggle, this toggle right here, simultaneously, something I believe O1 still can't do yet. OK, so R1 model uses the chain of thought prompting approach, which basically encourages the A1 model to break down the reasoning into simple to understand steps. This isn't new, but DeepSeek R1 does this really well. So let's use this simple maths problem as an example. The first part here is the problem to solve. And the second is the prompt that I've added to show its chain of thoughts. So that's specifically let's solve this step by step for each step, explain your thinking and show your calculations. So hitting enter, you can see DeepSeek thinking and reasoning with itself. And this is what makes R1 different. It transparently reasons through each step individually and figures it out in the same response in real time. Whereas GPT can often be sort of clinical and political. I found DeepSeek R1 to be direct, but also great at showing you the reasoning. And you can also extract the reasoning and send it to other AI models, to something that's unique to DeepSeek R1. The other cool thing is how DeepSeek R1 solves hallucinations. So hallucinations is a term to describe when AI gives you an incorrect answer. And it's a big challenge with current AI models. But I've noticed that R1 is particularly good at understanding why it hallucinates almost as if it's truly self-aware. And then it also corrects itself. So I started recording this specific clip here when I noticed that it gave me an incorrect answer to the vague question of what happened to Hershey's in 1998. It says Hershey's launched Almond Kisses in 1988, when in reality they were actually launched in 1990. So I pointed out the mistake and asked why it made the mistake. Because of its chain of thought approach, it's fascinating to see it run a search on this mistake, confirming why it made a mistake, and then it corrects itself here. Compared to other AI models, DeepSeek R1 thinks way more naturally, almost human-like, and elaborates on its mistake clearly. So I highly recommend challenging R1 when it hallucinates and give this a go yourself. It does seem to be slower though than ChatGPT 4.0, especially when it comes to coding tasks. I've been playing around with creating games on DeepSeek. Like if we ask it to create a Tetris game and then take the Python code and run it in HTML, it takes longer than it would in 4.0 before you can preview the game right from the chat. So if you have coding tasks, 0, 1, and particularly Claude 3.5 Sonnet still does a better job overall and will help remove the need to debug as a coder. But if you're looking for a free option or an open source option, R1 here is definitely the way to go currently and worth checking out. So based on my short time with R1, I feel like DeepSeek was probably trained on GPT 4.0-generated data. The responses on both models are eerily similar. And if you're concerned about privacy but still want to leverage DeepSeek R1, you can actually run it locally. Because it's open source, you can download and use the Olama app to run this R1 model on a local server. So all your questions and interactions remain completely private rather than on the cloud. But it is a very large model, so you'll need a beast of a setup to run its full R1 model locally. It's roughly like 1,300 gigabytes of VRAM that you'll need to run it fully. But there are distilled LLM versions of R1 that run on a single GPU. Version 1.5b in particular works fine on my Mac Studio M2 Ultra, for example. So that's my first look into DeepSeek R1. Clearly, some really incredible things happening in the AI space. When I began using DeepSeek, I was skeptical but very quickly realized it really is something special considering it's low cost to build and it's free for users to use. It's a very exciting time in the AI space and I'm keen to see how others like OpenAI respond to DeepSeek. If you made it to the end of this video, comment the code word R1 and I'll give it a like for making it to the end of this video. Make sure you subscribe for the latest in tech and AI content. And as always, thanks for watching and I'll see you in the next video.