Deepseek R2 Is About To Change That AI Industry (Deepseek R2 Leaks!)

So the AI industry is actually still continuing to heat up as deepseek are actually rushing to launch their new AI model and they are choosing to go all in on AI. If you aren't familiar with Deepseek's models, I'm pretty sure you haven't been in the AI space. But this is a company that completely put the AI industry on a hold because they did something that nobody expected. They managed to make really good frontier models for a fraction of the cost. And apparently now they are looking to launch their second iteration, R2. And, and I think this is even more concerning given the state of the AI industry. So you can see right here that it actually states that Deep Seek are actually planning to release this in early May, but they now want this early as possible. Two of them said, without providing specifics. So this is super interesting because it was only recently where deepseek managed to leapfrog other companies in terms of the price to output, in terms of the quality. But what we have here is an incredible situation that may turn the AI industry on its head, because if they can actually get Art 2, which will be the second iteration of their thinking models, to anywhere near the quality that we've recently seen from frontier labs like Anthropic and OpenAI for a fraction of the cost, then this could seriously put the Western AI nations, or should I say Western AI companies, in a very interesting predicament. Currently, I don't think developers and users have any loyalties to these brands. Of course, a lot of these brands do have, you know, brand names like ChatGPT, of course, Croc and Elon Musk. But many developers are always seeking cheaper prices and they're wondering how they can reduce costs for their apps or their software applications. And one of the key areas that they're actually going to be focusing on is going to be coding. So in this article it actually talks about how the company says that it hopes that the new model will produce better coding and be able to reason in languages beyond English. And details of the accelerated timeline have not been previously reported. Now, the reason I'm actually talking about coding here is because I do believe that coding is one of the hardest tasks to do in AI. It is certainly very interesting and I think the fact that a lot of people do use Claude 3.5 Sonnet, if we do get to a stage where Deepseek actually does better coding than Claude 3.5 or 3.6 on it, then that is going to be super interesting because that would basically remove Claude's entire market share and it wouldn't be the first time that this company has put the AI industry on a hold. For example, if we look at these recent coding benches, if we look at the agentic coding evaluation using Devon, Devon is the AI agent framework for software developers. And they are working with large enterprises to reduce the work that software developers have to do and automate a bunch of that stack. And you can see right here that the model that performs the best is of course, the Sonnet 3.7, which was newly released. Compared to, you know, DeepSeek R1, you can see that that one is at 51%. And you can see that other frontier models significantly outpace this one. But what happens if Deepseek R2 manages to surpass Sonnet 3.7 or even manages to get wedged in between here? I think this would mark a very pivotal moment for the AI industry as coding is something that many individuals do focus on and they are currently using to build out apps and many different pieces of software. So you can see right here, this is also another benchmark that's super interesting. This is the ADA Polyglot coding benchmark. And this is a tool designed to evaluate how effectively AI language models can translate natural coding requests into executable code that passes unit tests. Basically, this benchmark analyzes a variety of different code and it assesses not only the model's coding ability, but also the capacity to edit existing code and format those edits appropriately for integration into source files. Now, what we can see here as well is that, very interestingly, we can see that Deepseek R1 plus Claude Sonnet 3.5, that kind of agentic framework is actually one of the cheapest in this area. We can see that when we look at the total cost versus the performance metrics that we're getting here, we can see that this one is only $13 and is pretty high in terms of the score, whereas the other ones are $18, $100, and this one is $36. Now, I will say that it's quite likely that people will prefer accuracy over anything. So whil cost is going to matter. I do think for many different applications, this is something that they will have to focus on in terms of the fact that, look, accuracy will be important for Deepseek if they are to get that. But I think one of the things that makes Deepseek so unique is the fact that their accuracy doesn't matter that much. Of course, it's pretty accurate, but their price is so low that so many people want to use it because it is basically free. Now, if we also do look at the LMSYS arena, this is of course the qualitative benchmark which actually uses individuals such as yourself to look at how you can rank this. We can actually see that Deepseek R1 is only third in terms of the leaderboard space, or maybe fifth you can see right there. But this is still a relatively high spot for a non thinking model. And even Claude 3.7 Sonnet does actually rank lower on this benchmark. I do have to say, however, that this benchmark is quite likely going to be one where the vibes are really important. So for example, I do think models like GPT 4.5 would do well here, and I wouldn't really think that thinking models would do that well here. But it's going to be interesting to see where DeepSeek R2 does, because if that once again manages to leapfrog these companies, of course I don't think these benchmarks are that important, but I will say it will go to show the quality that they're going to be able to do. And also if they are able to get their second iteration out very, very quickly, as early as May, which is only around 60 to 90 days away from now, I think that is going to be a remarkable turnaround in terms of how they're managing to do this. And I think that would be rather worrying. Imagine you're a company that's able to churn out models really quickly for a fraction of the cost. The feedback loop in terms of improving each model cycle is going to get quicker each time. So it's going to be super, super interesting. And one of the things about R2 which is really worrying people is that recently Deepseek actually released some information on Twitter. They actually spoke on day six of their open source week that there was one more thing that they want to talk about. And they spoke about the statistics of their online service and the fact that their profit margin was 545%. That means that they're actually a profitable company. Now why is this something that should surprise you? Well, if you've been in AI, you'll know that they're, you know, on the back end of things. These companies aren't really making money and they have to continuously, you know, have these huge, huge funding rounds. For example, OpenAI had to have a $5 billion loss this year on a $3.7 billion revenue in this year. This was super, super interesting. Of course their revenue is growing year on year, which is good. But many people are speculating that these western tech companies are simply spending money on things that they won't need. I don't think that is the case in this specific scenario. I think AI is a specific scenario where you do need to spend a lot of money upfront and the profits will come later down the line once these gains are realized. But it is starting to beg the question, which stock market kind of took a massive dive is that, look, if these companies can actually be profitable and can actually have continued users, how on earth is OpenAI losing billions of dollars in revenue? And one of the things that you have to know is that models like GPT 4.5, they require significant investment in computational resources and infrastructure. Remember with GPT4, that apparently costs over $100 million to train, and those expenses actually contribute to the company's projected losses, such as this $5 billion loss that you're seeing right here. And of course, OpenAI, they've invested into Stargate initiative. They want to have massive data centers for future inference. But what is interesting is that even despite these losses, OpenAI actually continues to gain substantial investments due to its potential for long term profitability and the market leadership in AI. I remember that there were discussions of the recent fundraiser rounds, and apparently every single fundraiser round for OpenAI is oversubscribed. Basically saying that, look, every time they go for a round where people are allowed to invest, these investors essentially are so eager to invest in OpenAI that some of them don't even get the chance to. Which of course for OpenAI is most certainly a great problem to have. And you can see right here that Sam Altman actually talks about the fact that insane thing we are currently losing money on OpenAI Pro subscriptions. People use it much more than we expected. Now, something that I do think is interesting and something that a lot of people are overseeing is the fact that whilst, yes, other companies are making money. Well, you know, Deepseek R. I do think the point of this is, is that every single year the price of tokens does dramatically fall. So you can see right here, first GPT4 was $36 per million tokens. Then GPT4 Turbo was $14 per million tokens. Then it was $7, then $4. Then you can see GPT4 Mini was $0.25 per million tokens. And this is something that continues to get truer and truer and becomes clearer as the AI industry matures is the fact that every year the price of these models per token of intelligence will continue to go down. And I think this is somewhat of a Bad thing. Some people have said that, look, how are you guys going to make profit from this if other companies are going to continue to undercut you? But I think that this is just something that actually benefits OpenAI because essentially the service they're providing is going to be nearly free. And provided they have a good brand, then they are going to have one of the best products that people continue to want to use now. Now, even if Deepseek do manage to release R2, I do think that the rollout may be a bit shaky. R2, of course, is going to be the second iteration of the model, but you do have to remember previously there were situations where R2 is going to worry the US government because they've identified the leadership of AI as a national priority. And it says here that the release may further galvanize Chinese authorities and companies, dozens of which they have stated started integrating Deep Seq models into their products. And basically what they're worried about is the fact that Deepseek, they're probably going to ban this, okay? Because a few weeks ago, many government agencies in countries including South Korea and Australia have blocked access to Chinese AI startup Deepseek, mostly for government employees. So certain countries are not taking this well. They do not want Deepseek on their phones, of course. They are basically thinking that it may be Spyway or something like that. And for government safety, this is going to be banned. I do wonder if it is going to be another TikTok situation where they do basically ban the app due to this situation, where they think of it as a national security risk. But overall, I think that they should really be worried about China potentially surpassing them in the race to AGI. They've seemingly moved very quickly. And I think the reason that China have actually moved very, very quickly in terms of Deepseek is the fact that they have a very different management style. This is probably how they've been able to achieve so much in such a short space of time. You can see right here that it says that Liang gave us control and treated us as experts. He constantly asked us questions, learned alongside us, and this is someone who left the company. Deepseek allowed me to take control of ownership critical parts of the pipeline, which was very exciting. And essentially what we had here was something that was basically described as a flat management style. This was something that was very innovative, I believe, because this is something where they basically have everyone working on the same level. So it's not compartmentalized. You have everyone basically working together, and this allows the organization to move a lot more smoothly because individuals know how things are and they essentially work together more effectively. Whereas, you know, other companies are a lot more hierarchical, meaning that you have to speak to the manager and then this person and that person before it all goes up the chain and it all goes down the chain. So I think this kind of management style may actually be more commonplace in the future as we start to realize that, look, as a company, you can move a lot quicker if you have this flat management style. So what do you guys think about R2? Are you actually excited for the competition? Because I think that it's actually good for us because it actually forced OpenAI to release certain models and speed up certain releases. I think it's going to be super interesting to see what happens with R2 if it does manage to come out sometime next month, during April. I think that would be super, super interesting, considering that these companies typically take months and months and months to release new frontier models. But I cannot wait to see where things go next. But that being said, if you enjoyed the video, I'll see you in the next.