+- Communication Breakdown Open Source Community (https://communication-breakdown.com/mybb)
+-- Forum: Welcome To The Machine (https://communication-breakdown.com/mybb/Forum-welcome-to-the-machine)
+--- Forum: Here There And Everywhere (https://communication-breakdown.com/mybb/Forum-here-there-and-everywhere)
+--- Thread: Deep Seek starts bloodbath in the US stock market of over $1 trillion (/Thread-deep-seek-starts-bloodbath-in-the-us-stock-market-of-over-1-trillion)
Deep Seek starts bloodbath in the US stock market of over $1 trillion - chickensomething - 01-30-2025
Success from Deep Seek starts bloodbath in the US stock market of over $1 trillion. And the cause? The release of the DeepSeek R1 AI model from China. The Chinese model is as capable as the best US models, but it's free to use, open source, more efficient, and most shocking of all, it reportedly costs less than 3% of ChatGPT-4.
We were talking about an AI arms race between companies. Today that's evolved into an AI race between countries. In one corner, we have the United States. They have a long history of technological dominance. But then on the other side, we have China, a country with a very different ideology and motives. In this race to dominance, it's not about weapons, but it's about developing systems that are designed to think. Artificial intelligence. This race is reminiscent of the Cold War. The White House says that they're looking into "national security implications of China's DeepSeek AI platform." And to top it all off, OpenAI has accused DeepSeek of stealing its IP to train their own model. With the United States pouring in half a trillion dollars into the Stargate AI project, the global race is on, and this ongoing battle could be one of the biggest stories in tech this year. As artificial intelligence becomes a matter of national security, the technology would be forced to move faster than it is today. What a crazy time to be alive.
But before we get ahead of ourselves, what is really going on? How did a company from nowhere do all of this? Is this all just part of the AI hype cycle, or is this the real deal? It seems like the whole world is playing catch-up since the release, so let's try and make a guess. You are watching ColdFusion.
Historically, when technology meets a national security threat from an ideological opponent, we get inventions like the computer and jet aircraft from the competition of World War II, for example. But this time around, the United States was completely unchallenged in the field of AI... for a time. But that all changed on January 20th, 2025, with the release of R1.
DeepSeek R1, which is free, has performance reportedly on par with OpenAI's $200 a month model. And this is performance in the context of tasks such as language reasoning, mathematics, and coding. The free model also beats out Anthropic's Claude Sennet and Google's Gemini. But what many people may not know is that DeepSeek does things a little bit differently to the current state-of-the-art models. It's in part because there's no competition for that level of AI performance for free, users have been flocking to it, with DeepSeek becoming number one in Apple's App Store. But here are the stats of why people's jaws are dropping. The AI was built in 2 months and reportedly cost less than $5.6 million to build. The AI company Anthropic says that $100 million to $1 billion is the general amount needed to develop an AI. And to that end, Meta plans to spend $65 billion on AI, so creating something that performs this well with just $5.6 million is groundbreaking. But all may not be as it seems. More on that later.
"I think there are two very important things that people need to know about what's happening with DeepSeek AI and the way it's being interpreted on Wall Street. The first is, it doesn't matter if it's a Chinese government PsyOP or not. The technological innovation of having an LLM train itself through reinforcement learning is impressive. DeepSeek R1 being open source means that its code is freely available for whoever wants to use it and for whatever they want to use it for. Users can modify and modify the code to their liking. Please, all for free. This is totally the opposite approach of OpenAI, which is pretty ironic. This is all horrific news for USA AI companies because it means that suddenly they are all out of balance. DeepSeek, with its 671 billion parameters, can run locally, pros. In contrast, investors and companies have poured billions of dollars into the market. Service. Services that they've been providing. Maybe in the future, it's not going to be so much the models that would make the most money, but the applications that run on top of them. Has this all been a massive mistake from US investors? No one knows for sure, and that's why the markets are selling off. One bright spot for US companies though, is that users of AI systems may not feel comfortable in giving their data directly to China, especially in corporate settings."
In order to compete, Sam Altman, CEO of the ChatGPT maker OpenAI, has announced that their GPT 3.0 mini model will be given away for free. As for Mark Zuckerberg... Meta, they're internally panicking. But it's not just the Americans. Over in China, the effect is the same. Other Chinese tech giants such as ByteDance, the maker of TikTok, Alibaba, and Tencent, have freaked out and had to cut the prices of their AI models to compete. And despite the low price charged by DeepSeek, it remains profitable, while its rivals lose money. Interestingly, OpenAI told the Financial Times that they have evidence that DeepSeek was using the output from ChatGPT to train their own models, suggesting that IP theft has occurred. It should also be noted that it seems like Chinese AI developers are still managing to get their hands on top-of-the-line Nvidia GPUs, despite US sanctions.
"This is one of the 200 stolen cards that we just opened this morning. I'll show you how to open it later. In the last video, a friend, a lot of friends, said in the comments that this is illegal. We are illegal in this area, for the law of the United States. Trump will be very angry about this, right?"
But that begs the question, who are DeepSeek and how did DeepSeek get here?
But the story is interesting so far. DeepSeek founder Liang Wenfang isn't from the typical tech world. He actually has a background in finance and co-founded a hedge fund called High Flyer. His company used AI to predict market trends and help make investment decisions. And he was very successful at that, and his fund now manages $8 billion. But after his initial success, he wanted more. He wanted to build "human-level AI". In 2021, he started buying thousands of Nvidia GPUs as part of his quote "AI side project". This was right before the Biden administration began limiting US exports of AI hardware to China. Liang eventually spun off his AI side project into another company, DeepSeek, and R1 is their latest model.
But honestly, the more I've been reading up on the Liang story, the more interesting it gets. So let me know in the comment section if you want to see a dedicated episode on the DeepSeek founder.
So DeepSeek R1 was trained with Reinforcement Learning with Human Feedback (RLHF). That means there weren't any humans who helped it learn. And the method that DeepSeek uses for their model architecture is different to most of the other players. It's a technique called... Sky News explains it well.
"...and it sends it to a particular part of the digital brain to be dealt with. This lets the other parts remain switched off, saving time, energy, and most importantly, the need for more computer science."
The YouTube channel Computerphile explains further.
"...the right one, as opposed to, but one model can do it. So that's much more efficient."
To add to the efficiency is a process called distillation. Basically, using larger models to train smaller models in targeted domains. The result is equivalent performance with significantly less computing power. And this was the big shock for AI developers and financial markets.
"...reasoning completely open and visible. Choice. OpenAI basically does the opposite. It's essentially write down a step-by-step process for solving the problem and slowly solve it and then write down the answer. You tend to get much better at solving problems that require multiple steps. If you want to just, 'why is the sky blue?' it will just regurgitate that pretty easily from text that's learned on the internet. But if you're asking like problem-solving skills, it's hard to do in one shot, so you kind of take a little bit of time to just, their actual internal monologue, which is essentially a trade secret. What R1 is doing is it's doing a chain of thought which is similar to ChatGPT, but it's fully public. They've released all the models, they've released all the code, you can talk to it, you can see the entire monologue. They've also trained it with massively more limited data."
As mentioned earlier, things may not be as they seem. The claim that it only cost 5.6 million dollars to create the model may not be complete. In fact, in a paper released by DeepSeek themselves, they mentioned that that 5.6 million dollar figure includes only the official training of DeepSeek v3 and does not include costs of prior research, experiments on architectures, algorithms, or data. That does put a question mark on all the headlines we've been seeing that this thing was built for under 6 million dollars, but what does this mean? Well, it's likely to be much less than what US companies have been spending.
In the latest news, DeepSeek has also dropped an open image model, and at this rate, a video model will probably soon follow, to rival OpenAI's Sora or Google's anticipated Gemini Video. In terms of search interest right now, DeepSeek is trending, and it became one of the most downloaded apps on the App Store. And then towards the end of January, things absolutely blew up and went wild.
China, during Chinese New Year, went crazy. First, Alibaba comes out with Tongyi Qianwen 2.5, a very capable AI that could one-shot this code animation. Just asking a computer to code an animation and then it goes out and does it. It's so intuitive that I think kids of the future will believe that this is how coding always worked. Alibaba's Tongyi Qianwen 2.5 Max outperforms DeepSeek and even GPT-4.0 in some tasks. And then there's Baidu's ERNIE Bot 3.5 released around the same day.
Ok, before you all rush out to sign up to DeepSeek, please be aware of something. It collects data such as chat history, any text or audio inputs, uploaded files, keystroke patterns - basically anything you input into the model. Now OpenAI does similar things, but the difference is that with DeepSeek, your data goes straight to servers in the People's Republic of China. So I guess the question is, do you want to be spied on by the US? Or do you want to be spied on by China? There is an option for complete privacy. Here's the YouTube channel SomeOrdinaryGamers, running it locally.
"...write me code for a simple login webpage. So at this moment in time, it'll think, it'll be like, all right, the user is asking for code to create a simple login page. So first, it's going to structure HTML, then it's gonna style, then it's gonna validate. And then here it is, it's actually writing me the HTML code. So we're sitting in a world where like, I feel so scared for like junior coders these days, because God damn AI is really coming for some of the jobs that people least expected to lose first. So again, it writes this actual login page. And of course, once it's done, it'll actually also provide you a preview in this chat box software. So you can see it for yourself before you actually throw it into, you know, production or testing or whatever. So right here, I'm just gonna hit that preview button and boom, there it is."
This was also a warning to many, as it seems like the program may not be as ready as it seems. So what does Sam Altman think? He's only directly referenced the company once, saying, "DeepSeek's R1 is an impressive model, particularly around what they're able to deliver for the price. We will obviously deliver much better models, and it's also legit invigorating to have a new competitor. We will pull up some releases." We'll see what's around the corner for OpenAI. But the joke is, AI took ChatGPT's job.
What we're seeing here is the technological version of Thucydides Trap. Basically, it states "When a rising power challenges an existing power, conflict arises." In an interview with Caijing, republished in the China Academy of Information and Communications Technology back in mid-2024, DeepSeek's founder Liang made his ambitions clear. He said, "For years, Chinese companies have been accustomed to leveraging technological advantages. Our goal isn't quick profit, but adventure, enhancing the technological frontier to drive ecosystem growth. Why is Silicon Valley so innovative? Because they dare to try. When ChatGPT debuted, China lacked confidence in frontier research. From investors to major tech firms, many felt the gap was too wide and focused instead on innovations, but innovation requires confidence, and old people tend to have more of it."
With such a mindset, DeepSeek may force AI innovation forward, and China could be at the forefront of the global AI race. Competitors around the world will be forced to reduce their costs and rethink how they're creating AI models. Efficiency will be the aim of the game. We don't know how it will play out, but we do know that we'll be having some rapid advancements in the coming years. If we do remain positive, we could see breakthroughs to diseases like cancer and Alzheimer's. And I have done a video on that topic years ago before AI blew up, so you can check that out. But as usual, in all of this, let's just keep a close eye and see where this goes.